Emergency blood guarantee inventory dynamic optimization system based on edge computing

Abstract

The application belongs to the technical field of computer and edge computing, and discloses an emergency blood guarantee inventory dynamic optimization system based on edge computing. The system adopts a cloud-edge-end collaborative architecture, and comprises a global collaborative optimization module, a regional dynamic decision module and a local real-time sensing and execution module. The regional dynamic decision module integrates multi-source real-time data through a data fusion engine, performs fine demand prediction by using a space-time graph neural network, and generates inventory replenishment and allocation instructions based on model predictive control rolling optimization. The application realizes minute-level rapid response and accurate prediction optimization of emergency blood demand, improves blood guarantee rate, and reduces expiration loss and transportation cost.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computer and edge computing technology, specifically relating to a dynamic optimization system for emergency blood supply inventory based on edge computing. Background Technology

[0002] In the field of medical resource management and emergency response, improving the efficiency and capacity of resource allocation through information and intelligent technologies is a crucial aspect of enhancing public health emergency response capabilities. Blood, as an irreplaceable and time-sensitive special medical resource, has its inventory's dynamic management and optimization directly impacting the success rate of emergency treatment and the operational efficiency of medical institutions.

[0003] Dynamic optimization of emergency blood supply is a complex technical area in medical resource management. This approach aims to establish predictive and decision-making models to accurately predict and intelligently schedule blood inventory levels at different medical institutions, blood banks, and emergency stations, thereby ensuring supply while minimizing the risks of blood expiration and shortages.

[0004] Current technologies primarily rely on traditional time-series forecasting models based on historical data or simple heuristic rules for inventory decisions. These methods are slow to react and have low prediction accuracy when dealing with sudden, multi-source emergency events, and they fail to capture the spatiotemporal correlations between demands at different locations. Due to the high uncertainty and geographical specificity of emergency events, models under traditional centralized data processing architectures struggle to quickly analyze and respond to real-time, high-frequency demand data generated at the edge, leading to delayed inventory strategy adjustments. Furthermore, existing systems lack the ability to deeply integrate and perform real-time computing on multi-source heterogeneous data, resulting in inventory optimization decisions often based on isolated, outdated information, unable to adapt to the rapidly changing supply and demand situation in emergency scenarios. Therefore, how to construct a dynamic support system capable of real-time perception, rapid response, and precise optimization of emergency blood inventory has become a key technical challenge that needs to be addressed. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic optimization system for emergency blood supply inventory based on edge computing, in order to solve the problems of response lag, insufficient prediction accuracy, inability to integrate multi-source real-time data and capture spatiotemporal correlation caused by the reliance on centralized data processing and traditional prediction models in the existing emergency blood supply inventory management.

[0006] This invention provides a dynamic optimization system for emergency blood supply inventory based on edge computing, comprising:

[0007] The global collaborative optimization module, deployed in a cloud data center, is used to execute long-term strategic planning and macro-level resource allocation across regions.

[0008] The regional dynamic decision-making module is deployed in the edge computing center of each city or geographic region to be responsible for real-time tactical decision-making within the region, including a data fusion engine, a spatiotemporal graph neural network prediction unit, and a real-time inventory optimization decision-making unit.

[0009] The data fusion engine is used to access and preprocess multi-source heterogeneous data streams in real time, and to perform time alignment, spatial gridding mapping and feature vectorization on the data to generate a multi-dimensional feature tensor with unified spatiotemporal coding.

[0010] The spatiotemporal graph neural network prediction unit is used to construct a spatiotemporal graph structure based on the multidimensional feature tensor, with each medical institution and emergency station in the region as nodes and the geographical distance and traffic accessibility between nodes as edge weights. It also uses a gated cyclic graph convolutional network architecture to perform forward propagation calculations and outputs the probability distribution of predicted demand for each node in each prediction time window for different blood types and blood components.

[0011] The real-time inventory optimization decision unit is used to receive the predicted demand probability distribution and, in conjunction with the real-time inventory status of each node, run a rolling optimization algorithm based on model predictive control. The optimization objective is to minimize the expected shortage rate, blood expiration amount, and internal transfer cost in the region. The unit constructs and solves a mixed integer linear programming model to generate specific inventory replenishment and allocation instructions.

[0012] The local real-time sensing and execution module is deployed in each hospital blood bank or emergency station to execute sensing and operation commands, including the intelligent warehouse sensing submodule and the command execution submodule;

[0013] The global collaborative optimization module, the regional dynamic decision-making module, and the local real-time perception and execution module interact with each other and transmit instructions through an encrypted communication network.

[0014] Preferably, the global collaborative optimization module incorporates a multi-agent collaborative decision-making model based on deep reinforcement learning. This multi-agent collaborative decision-making model aims to maximize the overall blood supply rate and minimize blood expiration losses and emergency transportation costs nationwide. It receives aggregated data and status reports uploaded from dynamic decision-making modules in various regions and outputs blood reserve baseline adjustment instructions, cross-regional blood allocation plans, and emergency resource mobilization plans for sudden large-scale events for each region.

[0015] Preferably, the intelligent warehouse sensing submodule connects to the radio frequency identification reader, temperature and humidity sensor, and intelligent shelf weight sensor in the blood bank through Internet of Things technology, and collects the unique identification code, blood type, composition, collection date, current storage temperature, storage location and inventory quantity of each bag of blood in real time, and uploads this data to the corresponding regional dynamic decision-making module in real time.

[0016] The instruction execution submodule is used to receive and parse instructions from the regional dynamic decision-making module. The instruction types include inbound instructions, outbound instructions, and internal transfer preparation instructions. It also controls the corresponding equipment to complete the blood receiving, verification, shelving, picking, outbound handover, packaging, and transfer preparation operations.

[0017] Preferably, the multi-source heterogeneous data streams accessed by the data fusion engine specifically include:

[0018] The system includes real-time inventory data and emergency medical request data reported by the local real-time perception and execution module, real-time emergency medical event flow data from the city emergency medical command platform, real-time traffic condition data from the traffic management department, population heat and flow trend data from mobile communication operators, and short-term weather forecast data from the meteorological department.

[0019] Preferably, the spatial gridding mapping process of the data fusion engine specifically involves dividing the geographical area of ​​this region into regular square grids and mapping all data with spatial location attributes to the grid in which they reside.

[0020] Each grid generates a feature vector within each time slice, which includes the number of emergency events in that grid, population density index, average road speed, and weather condition code.

[0021] Preferably, the gated cyclic graph convolutional network of the spatiotemporal graph neural network prediction unit employs a first-order local spectral filtering method with Chebyshev polynomial approximation for its graph convolution operation.

[0022] At each time step, the network first normalizes the adjacency matrix of the graph, then performs graph convolution to aggregate the feature information of adjacent nodes, and then inputs the aggregated features and the hidden state of the node itself in the previous time step into the gated recurrent unit to update the current hidden state and output of the node.

[0023] Preferably, the model predictive control rolling optimization algorithm of the real-time inventory optimization decision unit has a decision cycle of 15 minutes;

[0024] In each decision cycle, the model predictive control rolling optimization algorithm takes the predicted demand for the next two prediction cycles as input and obtains the latest inventory snapshot.

[0025] The decision variables of the mixed integer linear programming model include the delivery volume from the regional central blood bank to each node in the next two cycles, the amount of mutual transfer between nodes, and the amount of blood that each node plans to scrap.

[0026] In the objective function of the mixed-integer linear programming model, the stockout penalty coefficient is set to 3 times the expiration penalty coefficient.

[0027] Preferably, when the instruction execution submodule receives an outbound instruction, its picking priority logic is as follows:

[0028] First, blood with identical blood type and composition is matched; when multiple bags of blood meet the requirements, the blood with the closer expiration date is selected first; if the expiration dates are the same, the blood in the storage location with the closer physical distance to the outbound handover point is selected first; the picking path planning adopts the nearest neighbor search algorithm based on dynamic programming.

[0029] Preferably, the deep reinforcement learning multi-agent collaborative decision-making model in the global collaborative optimization module adopts a paradigm of centralized training and distributed execution during the training process;

[0030] During the training phase, the model's global state space consists of the inventory status of all regions, the amount of blood in transit, recent demand history, and the weather forecast for the next 72 hours.

[0031] The action space of each agent is defined as the amount of resources requested from the global center and the adjustment range of the reserve baseline of its region;

[0032] The global reward function consists of the weighted negative values ​​of the national total coverage rate, total expiring loss rate, and total transportation cost; the model is trained offline using a near-end policy optimization algorithm.

[0033] Preferably, the system also includes a security and auditing blockchain sub-module that spans the cloud, edge, and terminal layers; the security and auditing blockchain sub-module is used to record operational events and the entire lifecycle information of blood circulation;

[0034] Every state change event of each bag of blood from collection, testing, warehousing, storage, allocation, and release to final use is generated by the local module at the trigger point. The transaction includes the event type, timestamp, location, operator, and blood identification code. After being digitally signed, it is uploaded to a permissioned blockchain network jointly maintained by multiple regional edge nodes for evidence storage.

[0035] The security and auditing blockchain sub-module adopts a permissioned blockchain network. The consensus nodes use a practical Byzantine fault-tolerant algorithm to verify and sort transactions. After reaching a consensus, new blocks are added to the chain to form an immutable audit trail chain.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] 1. This invention constructs a distributed computing architecture that integrates cloud, edge, and endpoint, decentralizing prediction and optimization computing capabilities to regional edge nodes. This enables localized real-time processing and rapid response to emergency medical data. The regional dynamic decision-making module can update predictions and make optimization decisions on a minute-level timescale based on multi-source fusion real-time data. This overcomes the decision-making lag problem caused by data transmission delays and processing bottlenecks in traditional centralized systems, improving the system's agility in responding to sudden and regional emergency medical needs.

[0038] 2. This invention employs a spatiotemporal graph neural network. By explicitly modeling the spatial topological relationships and time-series patterns of various medical nodes within a region, it can accurately capture the complex correlations and diffusion effects of emergency medical needs in the spatiotemporal dimension. Compared to traditional isolated time-series forecasting methods, the forecasting unit of this invention can more accurately predict the chain demand of surrounding hospitals that may be triggered by a single emergency medical event, thereby providing a more reliable forward-looking input for inventory optimization and reducing the risk of stockouts or redundancies due to inaccurate forecasts.

[0039] 3. This invention constructs the real-time inventory optimization problem as a rolling optimization problem based on model predictive control, and solves a specific mixed-integer linear programming model in each decision cycle. This method transforms dynamic optimization into a series of continuous static optimizations, ensuring both the real-time nature of decision-making and considering future uncertainties through a rolling time-domain approach. The optimization model simultaneously balances stockouts, expiration, and transportation costs, and assigns a higher penalty weight to stockouts in the objective function, ensuring that the system always prioritizes ensuring the safety of emergency supply under complex constraints, achieving an optimal balance between inventory costs and service levels.

[0040] 4. This invention, by introducing a security and auditing blockchain sub-module, establishes a complete, transparent, and tamper-proof electronic traceability file for each bag of blood. This greatly enhances the security and credibility of blood management, provides a reliable technical basis for quality tracking and responsibility definition, and the stored circulation information also provides high-quality real-world data feedback for global and regional optimization models, forming a complete quality closed loop from decision-making to execution to audit feedback, continuously driving system optimization. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0042] Figure 2 This is a schematic diagram of the multi-level interaction relationship and data flow of the distributed architecture and data interaction relationship of "cloud-edge-device" collaboration in this invention;

[0043] Figure 3 This is a schematic diagram of the core principle framework of the regional dynamic decision-making module in this invention;

[0044] Figure 4 This is a logical flowchart of the data fusion engine and the spatiotemporal graph neural network prediction unit in this invention;

[0045] Figure 5 This is a schematic diagram of the rolling optimization principle framework of the real-time inventory optimization decision unit based on model predictive control in this invention. Detailed Implementation

[0046] refer to Figures 1 to 5 This invention proposes a dynamic optimization system for emergency blood supply inventory based on edge computing. The overall technical architecture adopts a three-tiered collaborative distributed topology of "cloud-edge-device." By deploying complementary computing and sensing units at different levels, it constructs a closed-loop blood supply system with high responsiveness, strong predictive capabilities, and adaptive optimization capabilities. Please refer to the appendix. Figure 1 The diagram illustrates the overall technical architecture of the system. The global collaborative optimization module is located in the cloud data center, the regional dynamic decision-making module is deployed in edge computing centers in various cities or geographical regions, and the local real-time perception and execution module is distributed in blood banks and emergency stations of various medical institutions. The three modules form a vertically integrated and horizontally collaborative data interaction and command transmission mechanism through an encrypted communication network.

[0047] During system operation, the local real-time perception and execution module, as the front-end perception and execution unit, continuously collects and uploads blood inventory status and emergency event information; the regional dynamic decision-making module fuses multi-source heterogeneous data and completes ultra-short-term to short-term refined blood demand prediction based on spatiotemporal graph neural networks, thereby generating rolling optimized inventory replenishment and allocation instructions; the global collaborative optimization module, at a higher strategic level, coordinates the allocation of resources nationwide, formulates cross-regional allocation plans and emergency mobilization schemes, thereby achieving full-chain collaborative optimization from micro-operation to macro-control.

[0048] The local real-time sensing and execution module includes an intelligent warehouse sensing submodule and an instruction execution submodule. The intelligent warehouse sensing submodule connects to RFID readers, temperature and humidity sensors, and intelligent shelf weight sensors deployed within the blood bank via IoT technology to monitor the status of each bag of blood products throughout their entire lifecycle. RFID readers read the electronic tags embedded in the blood bags to obtain information such as their unique identifier, blood type, component type, collection date, and expiration date. Temperature and humidity sensors collect temperature and humidity data of the blood storage environment every 5 seconds to ensure it remains within the prescribed cold chain range. Intelligent shelf weight sensors monitor the number of blood bags in each storage location in real time using a high-precision weighing device, automatically triggering inventory updates when inbound or outbound operations occur. All the above sensing data is packaged into structured messages and uploaded in real time to the regional dynamic decision-making module of the relevant area via a secure transmission protocol.

[0049] The instruction execution submodule is responsible for receiving and parsing various operation instructions issued by the regional dynamic decision-making module, including inbound instructions, outbound instructions, and internal transfer preparation instructions. Upon receiving an inbound instruction, the submodule first verifies whether the electronic tag information of the blood bag matches the blood type, composition, and quantity specified in the instruction. If they match, it activates the automated guided vehicle (AGV) or robotic arm, guiding it to the designated unloading area to complete the receiving, verification, and shelving of the blood bags, and simultaneously updates the local inventory database. Upon receiving an outbound instruction, the submodule performs blood picking according to a preset picking priority logic: first, it filters blood bags whose blood type and composition perfectly match; if multiple candidates exist, it prioritizes blood bags with a closer expiration date; if the expiration dates are the same, it further compares the physical distance between the storage location of each candidate blood bag and the outbound handover point, prioritizing the blood bag with the closer distance. The picking path planning uses a nearest neighbor search algorithm based on dynamic programming to minimize the total travel distance of the AGV or robotic arm throughout the picking process, improving outbound efficiency. When an internal transfer preparation instruction is received, the instruction execution submodule moves the designated blood bag from the regular storage area to the transfer area and automatically generates an electronic transfer order containing information such as blood identification code, target node, and estimated shipping time for subsequent logistics scheduling.

[0050] Please refer to the appendix. Figure 2This diagram details the data interaction relationships between the "cloud-edge-device" three layers. The data stream uploaded from the local real-time perception and execution module to the regional dynamic decision-making module mainly includes real-time inventory snapshots, blood status change events, and emergency response request records; the regional dynamic decision-making module then issues specific inventory operation instructions to the local module. Simultaneously, the regional dynamic decision-making module periodically uploads aggregated data to the global collaborative optimization module, including the predicted total blood demand for the next 24 to 72 hours, current inventory health indicators, and cross-regional transfer requests; the global collaborative optimization module then issues blood reserve baseline adjustment instructions, cross-regional transfer plans, and emergency resource mobilization plans to each regional dynamic decision-making module. All data interactions employ end-to-end encryption and two-way authentication mechanisms to ensure the confidentiality and integrity of data during transmission.

[0051] The regional dynamic decision-making module is deployed in the edge computing center of each city or geographic region, and its internal structure is shown in the attached figure. Figure 3 As shown, it mainly includes three functional sub-modules: a data fusion engine, a spatiotemporal graph neural network prediction unit, and a real-time inventory optimization decision-making unit. The data fusion engine is responsible for real-time access and preprocessing of multi-source heterogeneous data streams from multiple external systems. These data sources include: real-time inventory data and emergency request data reported by the local real-time perception and execution module; real-time emergency event stream data provided by the city's emergency command platform, including the time, location, type of event, number of injured, and preliminary diagnosis information; real-time traffic data released by traffic management departments, covering the traffic speed, congestion index, and accident information of major roads; population heat and flow trend data provided by mobile communication operators, reflecting the spatial distribution and temporal evolution of population density in the region; and short-term weather forecast data released by meteorological departments, including the probability of precipitation, temperature changes, and extreme weather warnings for the next 6 hours. The data fusion engine performs unified time alignment processing on the above data, synchronizing all data streams to time slices with 15-minute intervals. Then, it performs spatial gridding mapping, dividing the geographical area into regular square grids with sides of 500 meters, mapping all data points with spatial location attributes to their respective grid cells. Finally, for each grid, a feature vector is generated within each time slice. This feature vector contains a classification of the number of emergency events within that grid: 0 for no events, 1 for minor events, 2 for moderate events, and 3 for severe events; population density index; average road speed; and weather status encoding: sunny = 0, cloudy = 1, rain = 2, snow = 3, thunderstorm = 4. All grid feature vectors are organized into a three-dimensional tensor with dimensions [time steps, total number of grids, feature dimension], and output as a unified spatiotemporal encoding multidimensional feature tensor to the spatiotemporal graph neural network prediction unit.

[0052] The spatiotemporal graph neural network prediction unit, based on the aforementioned multidimensional feature tensor, performs ultra-short-term predictions for the next 1 to 6 hours and short-term predictions for the next 6 to 24 hours. This unit first constructs a spatiotemporal graph structure with each medical institution and emergency medical station within the region as nodes. The number of nodes in the graph equals the total number of hospital blood banks and emergency medical stations involved in blood supply within the region, and each node corresponds to a specific geographical location. The edge weights of the graph are determined by the geographical distance and transportation accessibility between nodes, specifically calculated as follows: , For edge weights, Geographical distance, This is the average travel time. and This is an adjustable weighting coefficient used to balance the effects of distance and time factors. In this embodiment, The value is 0.6. The value is set to 0.4. The spatiotemporal graph neural network adopts a gated recurrent graph convolutional network architecture, combining graph convolution operations with gated recurrent units. At each time step... The network first normalizes the adjacency matrix of the graph to obtain a symmetric normalized adjacency matrix; then, it performs graph convolution operations to aggregate the feature information of adjacent nodes. To accommodate the limited computing power of edge computing nodes, the graph convolution operation adopts a first-order local spectral filtering method with Chebyshev polynomial approximation, mathematically expressed as follows:

[0053] ;

[0054] The first image obtained after graph convolution operation Layer node feature matrix, Given an adjacency matrix with self-loops, It is the original graph adjacency matrix. It is the identity matrix. Let it be its corresponding degree matrix. Let l be the node feature matrix of the l-th layer. The weight matrix is ​​a learnable matrix. The activation function is non-linear (ReLU is used in this embodiment). The output of the graph convolution operation is fed into the gated recurrent unit (GRU), participating in the state update along with the node's hidden state from the previous time step. The GRU, through update and reset gate mechanisms, alleviates the vanishing gradient problem in traditional recurrent neural networks, enabling it to better capture long-term temporal dependencies. After multiple layers are stacked, the network finally outputs the probability distribution of predicted demand for different blood types and blood components for each node within future prediction time windows. This probability distribution is given in the form of a Poisson or negative binomial distribution, and its parameters are obtained from the output of the last layer of the network through a Softmax or exponential transformation.

[0055] The real-time inventory optimization decision unit receives the prediction results output by the spatiotemporal graph neural network prediction unit and, combined with the current real-time inventory status of each node, generates specific inventory replenishment and allocation instructions. This real-time inventory optimization decision unit runs a rolling optimization algorithm based on model predictive control, with a decision cycle set to 15 minutes. At the beginning of each decision cycle, the algorithm first calls the latest round of prediction data to obtain the predicted demand values ​​for various types of blood at each node for the next two prediction cycles; it then obtains the latest inventory snapshot from the local real-time sensing and execution module, including the current inventory level of each node, the quantity of blood nearing its expiration date, and the upper limit of inventory capacity. The algorithm constructs a mixed-integer linear programming model, with decision variables including the delivery volume from the regional central blood bank to each node. i represents the central blood bank, k represents the blood type; the amount of inter-node transfers. j and m are two nodes, n is the time step, and n is the planned amount of blood to be scrapped at each node. The optimization objective of this mixed-integer linear programming model is to minimize the sum of the expected shortage rate, the amount of expired blood, and the internal transportation costs in the region. Its objective function can be expressed as:

[0056] ;

[0057] For nodes In time Regarding the shortage of blood type k, Expired quantity For nodes and The transportation distance between them , , These are the stockout penalty coefficient, the expiration penalty coefficient, and the unit transportation cost, respectively. In this embodiment, the stockout penalty coefficient is set to three times the expiration penalty coefficient to reflect the principle of "prioritizing supply security." The model's constraints include: inventory capacity limits for each node, blood shelf-life constraints, transport vehicle capacity constraints, and flow conservation constraints. This mixed-integer linear programming model is solved in real-time on edge computing nodes using a commercial solver, with the solution time strictly controlled within 10 minutes to ensure that instructions can be issued before the start of the next decision cycle. After the solution results are generated, the real-time inventory optimization decision unit packages the specific replenishment quantity suggestions, blood component allocation suggestions, and internal transfer route planning into an instruction set and sends it to the corresponding local real-time sensing and execution module through an encrypted channel.

[0058] The global collaborative optimization module is deployed in a cloud data center and is responsible for executing long-term strategic planning and macro-level resource allocation across regions. This module periodically (every 6 hours) receives aggregated data uploaded from all regional dynamic decision-making modules, including the total blood demand forecast for each region for the next 24 to 72 hours, inventory health indicators, and cross-regional transfer requests. The module incorporates a deep reinforcement learning-based multi-agent collaborative decision-making model, which treats each region as an independent agent. During the training phase, the model employs a paradigm of centralized training and distributed execution. The global state space consists of the inventory status of all regions, the amount of blood in transit, recent demand history (past 7 days), and the weather forecast for the next 72 hours. The action space for each agent is defined as the amount of resource transfer requested from the global center and the adjustment range of the reserve baseline for their region. The global reward function consists of the weighted negative values ​​of the national total guarantee rate, total expired loss rate, and total transportation cost, and its mathematical form is: , As a reward, Total out-of-stock amount, Total expired amount, For total transportation costs, , , These are the weighting coefficients, and The model is trained offline on a training dataset constructed from historical real events and a large-scale simulation environment using a proximal policy optimization algorithm until the policy converges. In actual operation, this multi-agent collaborative decision-making model outputs instructions for adjusting blood reserve baselines for each region, cross-regional blood allocation plans, and emergency resource mobilization plans for sudden large-scale events, based on the current global state.

[0059] This system also includes a security and auditing blockchain sub-module spanning the cloud, edge, and device layers. This sub-module employs a permissioned blockchain architecture, with a distributed ledger maintained by multiple regional edge nodes. For each bag of blood, every status change event from collection, testing, warehousing, storage, allocation, and dispensing to final use is recorded by a local module at its trigger point. This transaction record includes the event type, a timestamp accurate to the second, geographical coordinates, operator identification, and a unique blood identifier. The transaction record is digitally signed immediately upon generation and broadcast to consensus nodes in the blockchain network. Consensus nodes use a practical Byzantine fault-tolerant algorithm to verify and sort the transactions, and after reaching consensus, append a new block to the chain. This resulting immutable audit trail chain not only provides a complete electronic traceability record for blood management but also provides high-quality real-world feedback data for global and regional optimization models. For example, the regional dynamic decision-making module can periodically extract historical dispensing records from the blockchain to calibrate the accuracy of its predictive models; the global collaborative optimization module can analyze blood circulation efficiency nationwide and optimize cross-regional allocation strategies.

[0060] In summary, this embodiment constructs a dynamic optimization system for emergency blood supply inventory that is fast-responding, accurate in prediction, intelligent in decision-making, and fully traceable by organically integrating a "cloud-edge-device" collaborative architecture, multi-source data fusion, spatiotemporal graph neural network prediction, model prediction control optimization, and blockchain auditing. It solves the problems of slow response, inaccurate prediction, insufficient collaboration, and difficulty in traceability in existing technologies.

[0061] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0062] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic optimization system for emergency blood supply inventory based on edge computing, characterized in that, include: The global collaborative optimization module, deployed in a cloud data center, is used to execute long-term strategic planning and macro-level resource allocation across regions. The regional dynamic decision-making module is deployed in the edge computing center of each city or geographic region to be responsible for real-time tactical decision-making within the region, including a data fusion engine, a spatiotemporal graph neural network prediction unit, and a real-time inventory optimization decision-making unit. The data fusion engine is used to access and preprocess multi-source heterogeneous data streams in real time, and to perform time alignment, spatial gridding mapping and feature vectorization on the data to generate a multi-dimensional feature tensor with unified spatiotemporal coding. The spatiotemporal graph neural network prediction unit is used to construct a spatiotemporal graph structure based on the multidimensional feature tensor, with each medical institution and emergency station in the region as nodes and the geographical distance and traffic accessibility between nodes as edge weights. It also uses a gated cyclic graph convolutional network architecture to perform forward propagation calculations and outputs the probability distribution of predicted demand for each node in each prediction time window for different blood types and blood components. The real-time inventory optimization decision unit is used to receive the predicted demand probability distribution and, in conjunction with the real-time inventory status of each node, run a rolling optimization algorithm based on model predictive control. The optimization objective is to minimize the expected shortage rate, blood expiration amount, and internal transfer cost in the region. The unit constructs and solves a mixed integer linear programming model to generate specific inventory replenishment and allocation instructions. The local real-time sensing and execution module is deployed in each hospital blood bank or emergency station to execute sensing and operation commands, including the intelligent warehouse sensing submodule and the command execution submodule; The global collaborative optimization module, the regional dynamic decision-making module, and the local real-time perception and execution module interact with each other and transmit instructions through an encrypted communication network.

2. The edge computing-based emergency blood supply dynamic optimization system according to claim 1, characterized in that, The global collaborative optimization module incorporates a multi-agent collaborative decision-making model based on deep reinforcement learning. This model aims to maximize the overall blood supply rate while minimizing blood expiration losses and emergency transportation costs nationwide. It receives aggregated data and status reports uploaded from dynamic decision-making modules in various regions and outputs blood reserve baseline adjustment instructions, cross-regional blood allocation plans, and emergency resource mobilization plans for sudden large-scale events for each region.

3. The edge computing-based emergency blood supply inventory dynamic optimization system according to claim 2, characterized in that, The intelligent warehouse sensing submodule connects to the radio frequency identification reader, temperature and humidity sensor, and intelligent shelf weight sensor in the blood bank through Internet of Things technology. It collects the unique identification code, blood type, composition, collection date, current storage temperature, storage location and inventory quantity of each bag of blood in real time, and uploads this data to the corresponding regional dynamic decision-making module in real time. The instruction execution submodule is used to receive and parse instructions from the regional dynamic decision-making module. The instruction types include inbound instructions, outbound instructions, and internal transfer preparation instructions. It also controls the corresponding equipment to complete the blood receiving, verification, shelving, picking, outbound handover, packaging, and transfer preparation operations.

4. The edge computing-based emergency blood supply dynamic optimization system according to claim 3, characterized in that, The multi-source heterogeneous data streams accessed by the data fusion engine specifically include: The system includes real-time inventory data and emergency medical request data reported by the local real-time perception and execution module, real-time emergency medical event flow data from the city emergency medical command platform, real-time traffic condition data from the traffic management department, population heat and flow trend data from mobile communication operators, and short-term weather forecast data from the meteorological department.

5. The edge computing-based emergency blood supply inventory dynamic optimization system according to claim 4, characterized in that, The spatial gridding mapping process of the data fusion engine specifically involves dividing the geographical area of ​​this region into regular square grids and mapping all data with spatial location attributes to the grid in which they reside. Each grid generates a feature vector within each time slice, which includes the number of emergency events in that grid, population density index, average road speed, and weather condition code.

6. The edge computing-based emergency blood supply inventory dynamic optimization system according to claim 5, characterized in that, The gated cyclic graph convolutional network of the spatiotemporal graph neural network prediction unit employs a first-order local spectral filtering method with Chebyshev polynomial approximation for its graph convolution operation. At each time step, the network first normalizes the adjacency matrix of the graph, then performs graph convolution to aggregate the feature information of adjacent nodes, and then inputs the aggregated features and the hidden state of the node itself in the previous time step into the gated recurrent unit to update the current hidden state and output of the node.

7. The edge computing-based emergency blood supply inventory dynamic optimization system according to claim 6, characterized in that, The model prediction control rolling optimization algorithm of the real-time inventory optimization decision unit has a decision cycle of 15 minutes. In each decision cycle, the model predictive control rolling optimization algorithm takes the predicted demand for the next two prediction cycles as input and obtains the latest inventory snapshot. The decision variables of the mixed integer linear programming model include the delivery volume from the regional central blood bank to each node in the next two cycles, the amount of mutual transfer between nodes, and the amount of blood that each node plans to scrap. In the objective function of the mixed-integer linear programming model, the stockout penalty coefficient is set to 3 times the expiration penalty coefficient.

8. The edge computing-based emergency blood supply inventory dynamic optimization system according to claim 7, characterized in that, When the instruction execution submodule receives an outbound instruction, its picking priority logic is as follows: First, match blood with identical blood type and composition; when multiple bags of blood meet the requirements, prioritize the blood with the closest expiration date; if the expiration dates are the same, prioritize the blood stored in the location with the closest physical distance to the outbound handover point. The picking path planning uses a nearest neighbor search algorithm based on dynamic programming.

9. The edge computing-based emergency blood supply inventory dynamic optimization system according to claim 8, characterized in that, The deep reinforcement learning multi-agent collaborative decision-making model in the global collaborative optimization module adopts a paradigm of centralized training and distributed execution during the training process. During the training phase, the model's global state space consists of the inventory status of all regions, the amount of blood in transit, recent demand history, and the weather forecast for the next 72 hours. The action space of each agent is defined as the amount of resources requested from the global center and the adjustment range of the reserve baseline of its region; The global reward function consists of the weighted negative values ​​of the national total coverage rate, total expiring loss rate, and total transportation cost; the model is trained offline using a near-end policy optimization algorithm.

10. The edge computing-based emergency blood supply inventory dynamic optimization system according to claim 1, characterized in that, The system also includes a security and audit blockchain sub-module that spans the cloud, edge, and device layers; the security and audit blockchain sub-module is used to record the entire lifecycle information of operational events and blood circulation. Every state change event of each bag of blood from collection, testing, warehousing, storage, allocation, and release to final use is generated by the local module at the trigger point. The transaction includes the event type, timestamp, location, operator, and blood identification code. After being digitally signed, it is uploaded to a permissioned blockchain network jointly maintained by multiple regional edge nodes for evidence storage. The security and auditing blockchain sub-module adopts a permissioned blockchain network. The consensus nodes use a practical Byzantine fault-tolerant algorithm to verify and sort transactions. After reaching a consensus, new blocks are added to the chain to form an immutable audit trail chain.

Images