Intelligent analysis system and method for medicine supply chain based on cloud platform

Abstract

The invention discloses an intelligent analysis system and method for a medicine supply chain based on a cloud platform, and relates to the technical field of computer internet, and the system comprises the steps: building a multi-source heterogeneous data federated sensing network, carrying out the differential privacy processing, building a space-time coupled public opinion diffusion dynamic graph model to capture a propagation rule, and carrying out the analysis of a medicine supply chain based on a prediction result and an inventory state. A regional demand disturbance index of multi-modal fusion is calculated, and a collaborative replenishment strategy is generated by using a distributed constraint optimization algorithm; and when the public opinion and demand abnormity continuously exceeds a threshold value, automatically triggering a cross-node emergency allocation process, and executing collaborative inventory optimization and emergency collaborative scheduling under a privacy protection framework, so as to solve the problems of cross-mechanism privacy leakage, single public opinion modeling, demand perception lag and the like of the medicine supply chain. High-safety and high-precision medicine supply chain dynamic management is realized, and the predictability of demand prediction and the resource allocation efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of computer and internet technology, specifically to a cloud-based intelligent analysis system and method for the pharmaceutical supply chain. Background Technology

[0002] With the accelerated digitalization of the pharmaceutical supply chain, cloud-based intelligent analysis methods have demonstrated significant potential in improving drug distribution efficiency and emergency response capabilities. However, current mainstream sentiment-driven demand forecasting methods still have significant limitations. For example, patent application CN120875749A discloses a method and system for analyzing drug inventory demand based on sales data. This method assesses the impact of public opinion by analyzing the co-occurrence frequency of pharmacy sales data and drug names in text. However, pharmacy sales records involve corporate trade secrets and personal privacy, making real-time cross-institutional sharing difficult in practice, leading to distorted estimates of drug consumption. Furthermore, relying solely on the co-occurrence relationship between diseases and drugs to measure semantic associations fails to characterize the spatiotemporal diffusion dynamics of public opinion, resulting in a delayed and one-sided perception of regional demand disturbances.

[0003] Therefore, there is an urgent need for a cloud-based intelligent analysis method for the pharmaceutical supply chain that can build a multi-entity collaborative federated perception mechanism while strictly protecting the sensitive data of each drug supplier, integrate multimodal heterogeneous data to accurately depict regional demand mutations driven by public opinion, and thereby achieve privacy-preserving dynamic inventory optimization and cross-agency emergency dispatch. Summary of the Invention

[0004] The purpose of this invention is to provide a cloud-based intelligent analysis system and method for the pharmaceutical supply chain, in order to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a cloud-based intelligent analysis method for the pharmaceutical supply chain, comprising the following specific steps; Step 1: Construct a multi-source heterogeneous data federated perception network: Deploy local edge computing nodes for hospitals, chain pharmacies and regional distribution centers under the cloud platform architecture. Each node only uploads the drug consumption feature vector after differential privacy perturbation to the cloud aggregation server, without transmitting the original sales records. Step 2: Construct a spatiotemporally coupled dynamic graph model for public opinion diffusion: Based on social media text, news reports, geolocation check-in data, and drug image recognition results, extract disease keywords, drug entities, geographical locations, and timestamp quadruplets, construct a dynamic graph structure with regional divisions as nodes and cross-regional population flow intensity as edge weights, and learn the propagation path and intensity decay law of public opinion in the spatiotemporal dimension through a gated graph neural network. Step 3: Calculate the regional demand disturbance index of multimodal fusion: Align the public opinion dissemination intensity output in Step 2 with the regional drug consumption trend obtained in Step 1, and use a weighted fusion formula to calculate the regional demand disturbance index for each region. Step 4: Perform collaborative inventory optimization under privacy protection: Based on the demand disturbance index, each drug supplier runs a secure multi-party computation protocol locally to jointly solve a constrained linear programming model and generate a local replenishment strategy that meets service level requirements and minimizes total holding costs. During the process, only encrypted gradient information is exchanged, and each supplier’s inventory level and cost parameters are not disclosed. Step 5: Perform cross-agency emergency collaborative scheduling simulation: When the demand disturbance index of any region exceeds the preset threshold for multiple consecutive time units, and the safety stock of its subordinate nodes is lower than the warning line, the emergency response mechanism is activated. The minimum cost flow problem is solved based on the operations optimization engine embedded in the cloud platform. The objective function is to minimize the total allocation cost. The constraints include supply capacity, demand gap and non-negative allocation amount. After obtaining the optimal allocation plan, it is pushed to the management terminal.

[0006] Furthermore, the process of constructing the drug consumption feature vector in step 1 includes: obtaining the identity information of the drugs, including the generic name or code of the drugs, and collecting all the drug identity information to form a drug set; obtaining the average daily sales gp of the p-th drug within a unit period; adding Laplace noise no to the average daily sales gp at the local edge computing node to obtain the sales feature value vp of the p-th drug; generating a sales feature vector by compressing the sales feature value vp; obtaining the sales feature vectors uploaded by N participating nodes, and calculating the regional drug consumption feature vector Vap of the p-th drug, where Vap is the arithmetic mean of the N sales feature vectors.

[0007] Furthermore, the construction of the dynamic graph model in step 2 includes: extracting disease keywords d and drug names m from social media texts and news reports; collecting user geographic location information and converting it into national standard geographic division codes l; and recording the event timestamp t to form a quadruple (d, m, l, t); aggregating all the geographic division codes to form a node set V, with each unique division code corresponding to a graph node; establishing a directed edge between any two nodes, with the edge weight wuv calculated by the number of people moving between the two locations per unit time, the flow data originating from operator signaling or traffic checkpoint records; aggregating all edges to form an edge set E, and all edge weights to form an edge weight set W, thereby establishing a dynamic graph G = (V, E, W); for any timestamp t, mapping all quadruples at that moment to the corresponding graph node to form the node's initial feature vector.

[0008] Furthermore, the state update rule of the gated graph neural network in step 2 is as follows: For the v-th node in the node set V, let its hidden state at time t0 be h. v,t0 If its neighbor set is Q(v), then its hidden state update at time t0+1 is: , where w uv The edge weights between nodes u and v are defined by GRU, a gated recurrent unit used to fuse the current node state with neighbor weighted information, thereby modeling the propagation and decay of public opinion in the spatiotemporal dimension.

[0009] Furthermore, the calculation of the Regional Demand Disturbance Index (RDI) in step 3 includes: standardizing the regional drug consumption feature vector Vap to obtain the trend score Trend=(Vap-μbase) / σbase, where μbase and σbase are the historical mean and standard deviation of the p-th drug in the event-free baseline period, respectively. The hidden state of the nodes output by the gated graph neural network at the prediction time is input into a single-layer fully connected neural network, which outputs the regional flow intensity score Sent; the weight coefficients α and β are set to satisfy α+β=1, and the regional disturbance index RDI=α×Trend+β×Sent is calculated.

[0010] Furthermore, step 4, collaborative inventory optimization, includes: predicting the total regional demand D = D0 × (1 + κ × RDI) based on the RDI value, where D0 is the historical average demand for the same period, and κ is the disturbance amplification factor, with a value within a predetermined range; calculating the mean μ and standard deviation σ based on the predicted demand sequence for several future days; and calculating the dynamic safety stock using the service level factor z. , where L is the prediction cycle length; SS is used as the trigger threshold for the automatic replenishment system, and the replenishment process is started when the local inventory is lower than SS.

[0011] Preferably, the activation condition for emergency collaborative scheduling in step 5 is: the demand disturbance index (RDI) of any region exceeds a preset threshold for multiple consecutive time units, and the safety stock of a predetermined proportion or more of the hospital or pharmacy nodes in that region is lower than the warning line, i.e., the current inventory is less than the predetermined trigger threshold (SS); at this time, the cloud platform activates operations optimization, constructs a minimum cost flow model, and the objective function is minΣ(cij×xij), i.e., minimizing the product of cij×xij, where cij is the unit transfer cost from the i-th drug supplier to the j-th demander, and xij is the transfer quantity from the i-th drug supplier to the j-th demander; the constraints include: for the i-th drug supplier, Where Si is the available inventory of the i-th drug supplier, and J0 represents the total number of demanders connected to the i-th drug supplier; for the j-th demander, Where I0 represents the total number of drug suppliers, Fj is the predicted shortage of the j-th demander, Hj is the local inventory of the j-th demander, and xij≥0.

[0012] The predicted shortage Fj represents the predicted total demand of the j-th demander over the next L days, which is obtained by allocating the predicted total regional demand D from step 4 according to the historical proportion of each institution. , where Bj is the average daily consumption of the p-th drug by the j-th demander during the event-free baseline period, and B0 represents the sum of the average daily consumption of the p-th drug by all demanders during the event-free baseline period.

[0013] Preferably, an example of the method's implementation in a real-world scenario is as follows: Taking the demand forecast for oseltamivir during a peak influenza season in a certain city as an example, the average daily sales μbase during the event-free baseline period is the historical baseline value, and σbase is the corresponding standard deviation; after the perturbation uploaded by the edge node, Vap is higher than the baseline, so Trend is a positive value; the Sent output of the public opinion graph model is a positive number; assuming α and β are predetermined weight coefficients, RDI is the weighted result; D0 is the historical average weekly demand for the same period, κ is the perturbation amplification coefficient, so D is the amplified total demand; the mean demand μ and standard deviation σ for the next few days can be predicted, and SS is calculated in conjunction with the service level factor z; if a predetermined proportion of inventory in multiple hospitals in a certain district is lower than the warning line and RDI exceeds the preset threshold for multiple consecutive time units, emergency dispatch is triggered, and the system generates an allocation plan after solving the minimum cost flow.

[0014] Furthermore, the scale parameter ε of the Laplace noise in the differential privacy perturbation is set within a predetermined range to achieve a balance between data utility and privacy protection; the single-layer fully connected neural network contains a predetermined number of hidden units, uses the ReLU activation function, and the output layer uses the Sigmoid function to normalize Sent to a predetermined interval; the secure multi-party computation protocol adopts a gradient exchange mechanism based on homomorphic encryption to ensure that each drug supplier cannot infer the original inventory or cost parameters of other parties during the joint optimization process.

[0015] A cloud-based intelligent analysis system for the pharmaceutical supply chain includes: a federated learning coordination module, a spatiotemporal graph construction module, a joint reasoning module, an operations research optimization module, and an instruction execution module. The federated learning coordination module is used to manage the model training cycle, encrypted parameter upload, and global model distribution for each participating client. The spatiotemporal graph construction module is used to integrate geographic information, population flow, logistics trajectory and public opinion data to build and maintain a dynamically updated multi-relationship graph structure; The joint inference module is used to fuse the demand output from federated learning with the perturbation signals propagated in the spatiotemporal graph to generate regional demand forecasts; The operations research and optimization module is used to solve the optimal drug allocation and distribution plan based on demand forecasting results and resource constraints. The instruction execution module is used to perform rule verification on the optimization scheme and transmit it to the delivery scheme execution system.

[0016] The federated learning coordination module includes a local feature extraction unit and a secure aggregation unit; The local feature extraction unit is used to perform data desensitization and vector encoding on each participating party's side; The secure aggregation unit is used to perform noise aggregation on model parameters at the central server using a differential privacy mechanism.

[0017] The spatiotemporal graph construction module includes a node attribute management unit and a weight management unit; The node attribute management unit is used to maintain the basic static attributes and dynamic consumption indicators of each region; The weight management unit is used to calculate the edge weights between regional nodes based on population migration records; The joint inference module includes a disturbance detection unit and a graph propagation computation unit; The disturbance detection unit is used to monitor the public opinion heat and abnormal medical visit indicators at each node; The graph propagation computation unit is used to perform information transmission and state updates based on graph neural networks.

[0018] The operations research and optimization module includes a constraint modeling unit and a solution scheduling unit; The constraint modeling unit is used to transform inventory, transportation capacity, and policy requirements into mathematical constraints. The solution scheduling unit is used to call the optimization solver to generate allocation plans that meet timeliness requirements.

[0019] Compared with the prior art, the beneficial effects of the present invention are: 1. By integrating privacy-preserving drug consumption data under a federated learning framework with a multimodal dynamic graph model of public opinion, this invention overcomes the limitations of traditional methods that rely on single text co-occurrence or raw sales data. By introducing differential privacy Laplace noise at local edge nodes and combining it with a secure multi-party computation protocol, it ensures that raw sales data, patient privacy, and enterprise inventory cost parameters do not leave the local system throughout their entire lifecycle, thus resolving the information sharing dilemma faced by various stakeholders in the pharmaceutical supply chain due to commercial competition and legal compliance requirements.

[0020] 2. Compared to traditional lagging forecasting methods that rely solely on sales history, this system constructs a spatiotemporally coupled dynamic graph model of public opinion diffusion, enabling early perception of the propagation momentum of social public opinion in geographic space. Through deep modeling of population flow weights using GGNN, the system can proactively assess demand disturbances in the physical space based on public opinion signals in the digital space, before actual consumption surges occur.

[0021] 3. It can automatically trigger emergency dispatch simulation based on the minimum cost flow model and generate the optimal allocation plan covering the shortage institutions within a predetermined time period. Attached Figure Description

[0022] Fig. 1 This is a flowchart illustrating a cloud-based intelligent analysis method for the pharmaceutical supply chain according to the present invention. Fig. 2 This is a schematic diagram illustrating a deployment example of a cloud-based intelligent analysis system for the pharmaceutical supply chain according to the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example: Figs. 1-2 As shown, the present invention provides a technical solution: a cloud-based intelligent analysis system and method for the pharmaceutical supply chain.

[0025] The methods include: In the aforementioned cloud-based intelligent analysis method for the pharmaceutical supply chain, there is deep collaboration between the cloud cluster and edge hardware distributed across various medical institutions, retail terminals, and warehousing centers. The system's physical architecture is divided into a perception layer, a transmission layer, a computing layer, and a decision layer. The perception layer consists of local edge computing nodes deployed in various hospitals, chain pharmacies, and regional distribution centers. These nodes are connected to the cloud via industrial-grade gateways or dedicated lines. Furthermore, each local edge computing node is equipped with a processing unit featuring a hardware-level trusted execution environment for storing highly sensitive encryption keys and executing differential privacy algorithms.

[0026] In its implementation, this invention's cloud-based intelligent analysis method for the pharmaceutical supply chain first achieves data anonymization of the underlying raw data by constructing a multi-source heterogeneous data federated sensing network. Specifically, each local edge computing node extracts raw drug sales records, prescription information, and inventory balances in real time from the hospital's Hospital Information System (HIS), the pharmacy's Enterprise Resource Planning System (ERP), and the distribution center's Warehouse Management System (WMS) through a structured query interface. To ensure data compliance, this raw information is strictly confined to the memory of the local computing node and is prohibited from direct external transmission. Each node first standardizes the collected drug information, mapping it uniformly to the national standard drug coding system, thereby compiling a drug collection covering all dosage forms, including granules, capsules, and injections.

[0027] For drug p, the system first matches its unique identifier through the national drug coding database or generic name dictionary to form a drug set, and then calculates its average daily sales gp in the current unit period for drug p.

[0028] Subsequently, within each preset unit period, such as 24 hours, the local edge computing node will calculate the average daily actual sales of the p-th drug. To eliminate the extreme sensitivity of individual samples to global data, the local node activates the differential privacy module to inject Laplace noise into the average daily sales, thereby generating the sales feature value vp = gp + no, where no represents the Laplace noise, and the Laplace noise follows a Laplace distribution with parameter n. o ~Lap(Δf / ε), where Δf is the global sensitivity, dynamically adjusted based on the maximum fluctuation threshold of the node's historical sales, while the privacy budget parameter ε is set in the range of 0.1 to 1 in practical applications to balance data availability and privacy protection. This process is completed entirely at the edge; the original sales records do not leave the local device.

[0029] Next, the sales feature values ​​are compressed using a combination of hash mapping and sparse coding. For example, the code for the p-th drug is hashed using the SHA-256 function to generate a 256-bit digest, and the lower 64 bits are used as the integer index idx. A sparse vector is constructed, with only the idx position assigned a value of 1 and the rest set to 0, forming a high-dimensional sparse vector. This high-dimensional sparse vector is then compressed using a compression algorithm, such as the Zstandard algorithm, to generate a compact sales feature vector vp. At the end of each time unit, each edge node packages and uploads the sales feature vectors corresponding to all drugs to the cloud aggregation server. After receiving the sales feature vectors from N participating nodes, the cloud aggregation server performs an arithmetic average of the N sales feature vector values ​​for the p-th drug to calculate the regional drug consumption feature vector Vap. , where v p,kThis represents the sales characteristic value of the p-th drug uploaded by the k-th node. This aggregation process relies solely on the perturbed values ​​and does not involve any original transaction details, ensuring the security of the commercial secrets of each drug supplier.

[0030] In the above method, step 2 constructs a spatiotemporally coupled dynamic graph model of public opinion diffusion, which is implemented as follows: The system collects multimodal information in real time from public data sources such as social media and mainstream news websites, as well as user geographic location data, such as electronic map heat points, telecommunications operator base station signaling data and drug packaging image recognition results. A pre-trained named entity recognition model is used to jointly extract disease keywords d and drug names m from the text. The geographic information parsing module converts GPS coordinates or IP addresses into six-digit regional division codes as specified by relevant national standards as geographic division codes. The timestamp t is automatically recorded by the data collection time. The four constitute a structured quadruple (d, m, l, t).

[0031] After all historical and real-time quadruples are collected, the system extracts all unique geographic division codes appearing in them to form a node set V, where each element corresponds to a region. The construction of the edge set E is based on the intensity of cross-regional population flow: the system accesses operator signaling data or highway ETC checkpoint records to calculate the number of people flowing between any two regions vᵤ and vᵥ in a unit of time, such as 24 hours, fᵤᵥ. The edge weight wᵤᵥ is defined as wᵤᵥ=tanh(fᵤᵥ) to suppress numerical overflow caused by extremely high flow.

[0032] All directed edges (vᵤ, vᵥ) and their weights wᵤᵥ constitute the edge set E and the edge weight set W, respectively, thus establishing a dynamic graph G=(V,E,W). This graph structure is updated in a rolling time window, with a time granularity synchronized with step 1, which is 24 hours.

[0033] For any timestamp t, the system maps all quadruples at that time to corresponding graph nodes: if a quadruple contains a zoning code l, its disease keyword d and drug m are transformed into vectors through an embedding layer, and then fused with the frequency of similar historical events to form the initial feature vector of node v. This initial feature is input into a Gated Graph Neural Network (GGNN) for spatiotemporal propagation modeling. Each layer of the GGNN performs a state update: for the v-th node in the node set V, let its hidden state at time t0 be hᵥ, t0 If its neighbor node set Q(v)={u|(u,v)∈E}, then its hidden state update rule at time t0+1 is: ; Here, GRU stands for Gated Recurrent Unit, which contains update and reset gates to control the fusion ratio of historical state and neighbor aggregation information; wᵤᵥ is the edge weight, reflecting the propagation strength of information from u to v. After several iterations, the final hidden state hᵥ is output. t This state encodes the propagation path, intensity attenuation, and inter-regional coupling effect of public opinion in the spatiotemporal dimensions.

[0034] In the above method, step 3 calculates the regional demand disturbance index of multimodal fusion. The specific implementation details are as follows: The system first processes the regional drug consumption feature vector V output in step 1. ap Standardization is performed. The standardized parameters μbase and σbase represent the baseline period for the p-th drug in the absence of a significant surge in cases. In this example, a 30-day sliding window of no major public health events in the past 12 months can be added as the baseline period. The historical daily average sales mean and standard deviation during the baseline period are collected, stored long-term in the cloud, and updated periodically. The trend score (Trend) is calculated as: Trend = (Vap - μbase) / σbase. The Trend value reflects the degree of deviation of current drug consumption from the historical baseline; a positive value indicates rising demand, while a negative value indicates falling demand.

[0035] Simultaneously, the node hidden state hᵥ output by the gated graph neural network in step 2 at prediction time t, t Input a single-layer fully connected neural network. This network uses the ReLU activation function, and the output layer uses the Sigmoid function to normalize the result to the [0,1] interval, obtaining the regional flow intensity score Sent. The higher the Sent value, the stronger the influence of public opinion on the region, and the greater the risk of a surge in potential demand.

[0036] Subsequently, the system sets weight coefficients α and β, satisfying α + β = 1. The values ​​of α and β are dynamically adjusted according to the type of drug: in the example, for antiviral drugs, such as oseltamivir, α = 0.4 and β = 0.6, emphasizing public opinion-driven factors; for chronic disease drugs, such as antihypertensive drugs, α = 0.7 and β = 0.3, focusing on historical consumption trends.

[0037] Ultimately, the Regional Disturbance Index (RDI) is calculated as α × Trend + β × Sent. RDI is a dimensionless indicator used to quantify the overall strength of regional demand affected by external disturbances.

[0038] In the above method, step 4 performs collaborative inventory optimization under privacy protection. The specific implementation process is as follows: Each drug supplier, such as hospitals, pharmacies and drug distribution centers, runs a secure multi-party computation protocol based on homomorphic encryption locally.

[0039] First, the total regional demand D is predicted based on the RDI value. The calculation formula is D = D0 × (1 + κ · RDI), where D0 is the total regional demand for the same historical period, such as the average of the same week over the same year, and κ is the disturbance amplification factor, with a preferred value range of [0.1, 0.3], dynamically allocated according to the degree of drug shortage. For example, for oseltamivir during the flu season, κ = 0.25.

[0040] Subsequently, the system calculates the mean μ and standard deviation σ based on the predicted demand sequence for the next 7 days, or randomly generated according to the fluctuations of D. The dynamic safety stock SS is calculated using a service level factor z, with z corresponding to a 95% service level and a value of 1.96. The prediction period length L = 7 days, substituting... This SS value serves as the trigger threshold for the automatic replenishment system.

[0041] Throughout the optimization process, drug suppliers do not share original inventory, cost, or demand data. The SMPC protocol employs a Paillier homomorphic encryption scheme. Each party calculates the gradient of the objective function locally and uploads the encrypted gradient to the coordination server. The server aggregates the gradient in the ciphertext domain and distributes it. Each party decrypts the gradient locally and updates its local inventory strategy parameters. This process iterates until convergence, ensuring a globally optimal solution while preventing drug suppliers from inferring sensitive parameters from other parties.

[0042] In the above method, step 5 executes a cross-agency emergency collaborative scheduling simulation, specifically implemented as follows: The system continuously monitors the RDI value of each region and the inventory status of its subordinate nodes. The activation condition is: the RDI of any region exceeds a preset threshold θ for three consecutive time units, where θ=2, and the current inventory of more than 30% of the hospital or chain pharmacy nodes in that region is lower than the trigger threshold of the automatic replenishment system. When the conditions are met, the cloud platform automatically activates the embedded operations optimization.

[0043] Minimum Cost Flow Model: Let I be the set of organizations participating in the scheduling, where the supplier set A1⊆I and the demand set A2⊆I. The objective function is min{∑ i ∑ j Let (cᵢⱼ×xᵢⱼ)}, where i∈A1, j∈A2, and cᵢⱼ is the unit transfer cost from node i to node j, including transportation costs, labor costs, and loss rate adjustments, and xᵢⱼ is the transfer quantity. Constraints include: Where Si is the available inventory of the i-th drug supplier, i.e., the current inventory minus the safety stock SS; J0 represents the total number of demanders connected with the i-th drug supplier; for the j-th demander, Where I0 represents the total number of drug suppliers, Fj is the predicted shortage of the j-th demander, Hj is the local inventory of the j-th demander, and xij≥0.

[0044] After determining the optimal allocation plan, the system generates structured scheduling instructions, which are pushed to the management terminals of each institution via API. The instructions include the sending party, receiving party, drug code, quantity, estimated delivery time, and logistics carrier information. Simultaneously, the system records scheduling logs for post-event auditing and model feedback optimization.

[0045] To enable those skilled in the art to more clearly understand the technical solution of the present invention, the following detailed calculation examples of the method of the present invention are provided in conjunction with specific implementation scenarios and data: Suppose a municipality is monitoring the supply chain of its core drug "oseltamivir" during the flu season. In step 1, the city has 300 hospitals and 1500 pharmacies as local edge nodes. Taking a large chain pharmacy node k as an example, its average daily actual sales volume gp of oseltamivir within a unit period is 120 boxes. The generated Laplace noise no is 2.5, then the sales volume feature value vp uploaded by this node = 120 + 2.5 = 122.5.

[0046] After acquiring data from N nodes across the city via the cloud, the regional drug consumption feature vector Vap was calculated to be 135 boxes / node / day.

[0047] In step 2, the system detects a large-scale influenza outbreak in a neighboring province and calculates the flow intensity weight w from neighboring province A to the city using signaling data. uv =0.75. The gated graph neural network extracts quadruples containing keywords such as "fever," "oseltamivir," and "influenza." After three rounds of GRU iteration, the hidden state h of the nodes in the target region is... v,t It exhibits strong characteristics of high risk.

[0048] In step 3, given the baseline mean μbase = 80 and standard deviation σbase = 15, the trend score is calculated as Trend = (135 - 80) / 15 = 3.67.

[0049] h v,t The input is a fully connected neural network, and the output flow intensity score is Sent=0.88. With weights set α=0.7 and β=0.3, the regional demand disturbance index RDI is obtained as RDI=0.7×3.67+0.3×0.88=2.83.

[0050] In step 4, the historical weekly demand for the same period is D0 = 100,000 boxes, and the disturbance amplification factor κ = 0.25 is set. The predicted total weekly demand for the future is D = 100,000 × (1 + 0.25 × 2.83) = 170,750 boxes. Based on the demand sequence, the standard deviation σ = 450, the service level factor z = 1.96, corresponding to a 95% service level, and the forecast period L = 7 days, the dynamic safety stock is calculated. Approximately 2333 boxes.

[0051] In step 5, the system detected that the city's RDI value exceeded the threshold of 2 for three consecutive days, and 70% of the city's pharmacy inventory was below 2333 boxes, triggering emergency dispatch. A distribution center has a surplus inventory Si = 50,000 boxes. The predicted total demand Fj of a hospital, based on its historical proportion of 2%, is calculated as 170750 × 0.02 = 3415 boxes, while its current inventory Hj = 1200 boxes, resulting in a shortage of 2215 boxes. The system, by solving the minimum cost flow model, determines to transfer 2215 boxes from the distribution center to the hospital, with a unit transfer cost c. ij =5 yuan / box, the lowest total cost.

[0052] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A cloud-based intelligent analysis method for the pharmaceutical supply chain, characterized in that: The method includes the following steps: Based on a cloud platform architecture, local edge computing nodes are deployed in hospitals, chain pharmacies, and regional distribution centers. Each node generates a perturbed sales feature value by adding noise to the average daily sales of medicines at its own node, and uploads the compressed sales feature vector to the cloud aggregation server to obtain the regional drug consumption feature vector. Based on social media text, news reports and geolocation data, disease keywords, drug names, geographical locations and timestamps are extracted. A dynamic graph structure is constructed with regional divisions as nodes and cross-regional population flow intensity as edge weights. The propagation path and intensity decay law of public opinion in the spatiotemporal dimension are learned through a gated graph neural network. Aligning the intensity of public opinion dissemination with drug consumption trends, a weighted fusion method was used to calculate the regional demand disturbance index for each region; Each drug supplier generates a local replenishment strategy that meets service level requirements and minimizes costs by solving a constrained linear programming model based on the regional demand disturbance index. When the demand disturbance index of any region exceeds the preset threshold for several consecutive time units, and the safety stock of its subordinate nodes is lower than the warning line, the emergency response mechanism is activated. The optimal allocation plan is obtained by minimizing the total allocation cost as the objective function and then pushed to the management terminal.

2. The intelligent analysis method for the pharmaceutical supply chain based on a cloud platform according to claim 1, characterized in that: The method for generating the sales characteristic value includes: The process involves: acquiring the identity information of the drugs and forming a drug set; calculating the average daily sales of the p-th drug within a unit period; adding Laplace noise to the average daily sales at the local edge computing node to obtain the sales feature value of the p-th drug; compressing the sales feature value to generate a sales feature vector; uploading the sales feature vector to the cloud platform; and then averaging the sales feature values ​​of the p-th drug from N nodes to obtain the drug consumption feature vector of the p-th drug in the region.

3. The intelligent analysis method for the pharmaceutical supply chain based on a cloud platform according to claim 1, characterized in that: The construction of the dynamic graph structure includes: Obtain a quadruple consisting of disease keywords, drug names, geographic region codes, and timestamps; collect all unique geographic region codes to form a node set. Calculate the edge weight between any two nodes based on the number of people moving between two locations per unit time, and establish a set of directed edges; map the quadruplets under each timestamp to the corresponding graph nodes to form the initial feature vector of the nodes; The state update rule of the gated graph neural network is as follows: for any node's hidden state at time t0, its hidden state at time t0+1 is generated by fusing the current node's state and the weighted states of its neighboring nodes through a gated recurrent unit, where the contribution of the neighboring nodes is weighted by the corresponding edge weights.

4. The intelligent analysis method for the pharmaceutical supply chain based on a cloud platform according to claim 1, characterized in that: The calculation of the regional demand disturbance index includes: The regional drug consumption feature vector is standardized to obtain a trend score; the hidden state of the nodes output by the gated graph neural network is input into a single-layer fully connected neural network to output a regional flow intensity score normalized to a predetermined interval; the trend score and the regional flow intensity score are weighted to obtain the regional demand disturbance index.

5. The intelligent analysis method for the pharmaceutical supply chain based on a cloud platform according to claim 1, characterized in that: The local replenishment strategy includes: The regional demand disturbance index is amplified proportionally by a disturbance amplification factor to obtain a disturbance reference value. The predicted value of the total regional demand is obtained by multiplying the historical average demand for the same period by the disturbance reference value. The disturbance amplification factor is set within a predetermined range based on the degree of drug shortage. The mean and standard deviation are calculated based on the predicted demand sequence for several days in the future. The dynamic safety stock level is calculated by combining the service level factor and the forecast period length, which serves as the local replenishment trigger threshold. When the local inventory falls below this threshold, the replenishment process is initiated.

6. The intelligent analysis method for the pharmaceutical supply chain based on a cloud platform according to claim 1, characterized in that: The response methods of the emergency response mechanism include: If the demand disturbance index in any region exceeds the response threshold for three consecutive time units, and the current inventory of the hospital or pharmacy nodes in that region that exceed the percentage threshold is less than half of the safety stock; The constraints of the objective function include: the total allocation amount of each supplier does not exceed its available inventory, the total receiving amount of each demander is not less than the predicted shortage, and all allocation amounts are non-negative; the optimal allocation plan obtained by solving includes the transferor, transferee, drug code, quantity, and expected delivery time, and is pushed to the management terminal through the application programming interface.

7. A cloud-based intelligent analysis system for the pharmaceutical supply chain, used to execute the cloud-based intelligent analysis method for the pharmaceutical supply chain as described in claim 1, characterized in that: The system includes: The modules include a federated learning coordination module, a spatiotemporal graph construction module, a joint reasoning module, an operations research optimization module, and an instruction execution module. The federated learning coordination module is used to manage the model training cycle, encrypted parameter upload, and global model distribution of each participating client. The spatiotemporal graph construction module is used to integrate geographic information, population flow, logistics trajectory and public opinion data to construct and maintain a dynamically updated multi-relationship graph structure. The joint inference module is used to fuse the demand output from federated learning with the perturbation signal propagated in the spatiotemporal graph to generate regional demand forecasts. The operations research and optimization module is used to solve the optimal drug allocation and distribution plan based on demand forecasting results and resource constraints. The instruction execution module is used to perform rule verification on the optimization scheme and transmit it to the delivery scheme execution system.

8. The intelligent analysis system for the pharmaceutical supply chain based on a cloud platform according to claim 7, characterized in that: The federated learning coordination module includes a local feature extraction unit and a secure aggregation unit; The local feature extraction unit is used to perform data desensitization and vector encoding on each participating party's side; the secure aggregation unit is used to perform noise aggregation on the model parameters on the central server side using a differential privacy mechanism.

9. The intelligent analysis system for the pharmaceutical supply chain based on a cloud platform according to claim 7, characterized in that: The spatiotemporal graph construction module includes a node attribute management unit and a weight management unit, and the joint inference module includes a disturbance detection unit and a graph propagation calculation unit. The node attribute management unit is used to maintain the basic static attributes and dynamic consumption indicators of each region. The weight management unit is used to calculate the edge weights between regional nodes based on population migration records; The disturbance detection unit is used to monitor the public opinion heat and abnormal medical treatment indicators of each node. The graph propagation computation unit is used to perform information transmission and state updates based on graph neural networks.

10. The intelligent analysis system for the pharmaceutical supply chain based on a cloud platform according to claim 7, characterized in that: The operations research and optimization module includes a constraint modeling unit and a solution scheduling unit; The constraint modeling unit is used to transform inventory, transportation capacity and policy requirements into mathematical constraints; the solution scheduling unit is used to call the optimization solver to generate allocation schemes that meet timeliness requirements.

Images