Intelligent management method for medicine supply chain based on big data processing

Abstract

The invention discloses a medicine supply chain intelligent management method based on big data processing, and the method comprises the steps: collecting multi-source data, and carrying out the preprocessing of the multi-source data to generate a standardized data set; constructing a unified entity set, and generating full-link tracing relation data; constructing a multivariable time sequence, and forming a variable Token; an improved iTransform model is constructed, and a structured result is obtained; outputting a scene set and a disposal record based on approximate nearest neighbor search; generating an executable instruction set according to the structured result, the scene set and the disposal record; and outputting to a service execution end, and generating an execution result and a tracing audit record. According to the invention, through combination of prediction and early warning of the improved iTransform model and similar scene retrieval of approximate nearest neighbor search, perceptible, traceable and executable closed-loop intelligent management of the medicine supply chain is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent supply chain management technology, and in particular to an intelligent management method for the pharmaceutical supply chain based on big data processing. Background Technology

[0002] The pharmaceutical supply chain covers procurement, warehousing, transportation, distribution, traceability, and quality management. Business data is typically stored in different data carriers from various sources, including orders, inventory, transportation, traceability, and temperature control. These data are generated and maintained by different departments and systems at different time granularities, easily leading to inconsistencies in field meanings, inconsistent coding standards, missing or offset timestamps, and a coexistence of duplicate and missing data. Due to the lack of a unified entity set and relational mapping model for pharmaceutical business, it is difficult to establish a stable and continuously updated mapping relationship between drug codes, batch numbers, traceability codes, expiration dates, temperature zones, and logistics carriers such as boxes, waybills, nodes, and storage locations. Inbound, picking, outbound, loading, in-transit, arrival, receipt, temperature control data collection, anomaly handling, and recall isolation events are difficult to link together along a single chain. This results in difficulty in real-time perception of the entire supply chain status, incomplete traceability chains, time-consuming cross-system location of batch risks and responsibility attribution, and fragmented audit evidence with high costs for aggregation and verification.

[0003] Current pharmaceutical supply chain management largely relies on rule configuration, experience thresholds, or single-point forecasting analysis. Common practices separate demand forecasting and risk warning from business decisions regarding replenishment, allocation, picking and warehousing, cold chain disposal, and distribution adjustments. Model outputs are mostly reports or risk alerts, making it difficult to directly generate a set of executable instructions and create verifiable traceability audit records after execution, given the compliance of batch numbers, expiration dates, temperature zones, and traceability codes. In multivariate time series forecasting scenarios, the relationships between variables are complex. Traditional modeling methods are insufficient in characterizing variable relationships and business levels, leading to significant impacts on forecast and warning stability due to data fluctuations. When historical similarities are needed to assist in handling, existing similarity searches often use static indexes or simple candidate access sequences, which are ill-suited to the degradation of index structures caused by changes in data distribution due to seasonality, policies, channels, and supply and demand fluctuations. Unstable search results further affect the timeliness and consistency of coordinated decisions regarding replenishment, allocation, and cold chain disposal.

[0004] Therefore, how to provide a smart management method for the pharmaceutical supply chain based on big data processing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent management method for the pharmaceutical supply chain based on big data processing. This invention comprehensively utilizes multi-source data preprocessing and standardization, unified entity and association mapping modeling, improved iTransformer multi-channel embedding, variable hierarchical grouping self-attention and multi-task structured decoding, graph index dynamic edge update based on near nearest neighbor search, and candidate access order structure internal sorting to realize the process from full-link traceability relationship construction, prediction and early warning and decision suggestion output, similar scenario retrieval to executable instruction generation and traceability audit record generation. Compared with the prior art, this invention achieves perceptible and traceable connection of batch number, expiration date, temperature zone and traceability code, and links replenishment, allocation, distribution and recall disposal under the premise of compliance, with the advantages of strong executableness and engineering implementation.

[0006] A pharmaceutical supply chain intelligent management method based on big data processing according to an embodiment of the present invention includes: Collect multi-source data from the pharmaceutical supply chain, preprocess the multi-source data, and generate a standardized dataset; A unified entity set is constructed based on a standardized dataset, and the relationship mapping between entities is established to generate full-link traceability relationship data. By tracing relationship data across the entire chain, a multivariate time series is constructed, and the historical sequence of each variable is used to construct a variable token. An improved iTransformer model is constructed. The variable Token is embedded and encoded separately through multi-channel embedding and then fused. Grouping self-attention with variable hierarchy awareness is introduced to group the variables and perform self-attention processing in sequence. A multi-task output head is introduced for parallel decoding to obtain structured results. The scene vectors are constructed from the structured results and written into the vector library. A graph index for vector library retrieval is established. Similar scene retrieval is performed on the scene vectors based on approximate nearest neighbor search. Local edge reconstruction is performed by updating the dynamic edges in the graph index. Sorting is optimized by using in-structure sorting. The output is a set of similar scenes and historical processing records. Based on the structured results, similar scenario sets, and historical disposal records, an executable instruction set is generated, and each instruction is associated with batch number, expiration date, temperature zone, and traceability code information; The set of executable instructions is output to the business execution end, and execution result records and traceability audit records are generated through execution result data and traceability evidence data.

[0007] Optionally, the multi-source data includes order data, inventory data, transportation data, traceability data, and temperature control data.

[0008] Optionally, generating the standardized dataset includes: Collect order data, inventory data, transportation data, traceability data, and temperature control data, and perform field extraction on the multi-source data, extracting the drug code field, batch number field, traceability code field, node identifier field, document identifier field, time field, quantity field, and status field; The field extraction results undergo standardization processing, which includes field mapping, data type conversion, unit unification, and time unification. Time fields are converted to timestamp fields, quantity fields are converted to units of measurement, and temperature control fields are converted to temperature units. Data that meet the criteria of consistent drug code, node identifier, and batch number and fall within the same time granularity are aggregated according to time granularity to generate aggregated records. Data cleaning is performed on the aggregated records. The data cleaning process includes constructing a deduplication key based on the drug code, node identifier, batch number, timestamp and record type and deleting duplicate records with the same deduplication key, filling missing fields, setting the quantity field to zero for records with negative values, writing anomaly markers to temperature control records that exceed the preset temperature range, and outputting a standardized dataset.

[0009] Optionally, generating end-to-end traceability data includes: Entity tables are created based on standardized datasets. The entity tables include drug entities, batch entities, traceability code entities, node entities, storage location entities, container entities, waybill entities, and order entities. An association mapping table is generated based on the entity table. The association mapping table includes order-drug code mapping, order-batch number mapping, order-traceability code mapping, batch number-expiration date mapping, batch number-temperature zone mapping, traceability code-box pallet mapping, box pallet-waybill mapping, waybill-node mapping, node-warehouse location mapping, and waybill-temperature control record mapping. An effective timestamp and an expiration timestamp are written for each mapping record. Full-chain traceability relationship data is generated based on the association mapping table and timestamps. The full-chain traceability relationship data uses the traceability code or batch number as the primary key, and the events are linked in sequence according to the timestamp order to output the full-chain traceability relationship data.

[0010] Optionally, constructing the historical sequence of each variable into a variable token includes: The management objects are identified by tracing the relationship data of the entire chain. The management objects consist of node identifiers and drug codes. A timeline and historical window length are determined for each management object. The timeline consists of a continuous timestamp sequence with time granularity. Aggregate variable sequences for managed objects on the timeline by timestamp. The variable sequences include demand variable sequences, inventory variable sequences, in-transit variable sequences, transportation timeliness variable sequences, temperature control variable sequences, traceability event variable sequences, and compliance status variable sequences. A multivariate time series matrix is ​​generated based on the variable sequence. The rows of the multivariate time series matrix correspond to the timestamp sequence, and the columns correspond to the variable set. The historical window length is truncated for each column in the variable set to obtain the historical sequence of the variable. The historical sequence of the variable is used to construct a variable token, and the index of the variable token corresponds to the variable in the variable set.

[0011] Optionally, obtaining the structured result includes: An improved iTransformer model is constructed, including a multi-channel embedding module, a grouped self-attention encoding module, and a multi-task output head module; The multi-channel embedding module performs three-channel encoding processing on each variable Token, dividing each variable Token into a numerical channel sequence, an event channel sequence, and a compliance status channel sequence. It then performs in-channel embedding encoding to generate numerical embedding vectors, event embedding vectors, and compliance embedding vectors, and performs fusion mapping on the three types of embedding vectors to generate variable Token representation vectors. The grouped self-attention encoding module performs grouping and two-level self-attention processing on the variable Token representation vector. It groups the variables according to demand, inventory and in-transit, transportation and cold chain, and quality and compliance. Within each group, it performs multi-head self-attention operation on the variable Token representation vector to obtain the group aggregate representation. Then, it uses the group aggregate representation as input to perform multi-head self-attention operation between groups to obtain the global aggregate representation. The multi-task output header module performs parallel decoding on the global aggregate representation and outputs demand forecast results, risk warning results, and decision suggestion results respectively. The demand forecast results are the demand time series within the forecast step, the risk warning results are the risk probability time series within the forecast step, and the decision suggestion results are the replenishment suggestion time series and allocation suggestion time series within the forecast step, thus obtaining structured results.

[0012] Optionally, the output set of similar scenarios and historical processing records includes: Scenario vectors are constructed based on the structured results and written into the vector library. The scenario vectors are generated by splicing together the demand forecast results, risk warning results, decision suggestion results and management object identifiers, and a timestamp is written for each scenario vector. A graph index for approximate nearest neighbor search is constructed based on a vector library. The graph index consists of a node set and an edge set. The node set stores the scene vector identifier and scene vector value, and the edge set stores the list of adjacent nodes for each node. The local edge reconstruction is performed by updating the dynamic edge in the graph index. When a new scene vector is written to the vector library, a new node is generated in the graph index, and the new node identifier and the new scene vector value are registered in the node set. The neighborhood candidate node set is retrieved for the new node in the edge set and the edge between the new node and the neighborhood candidate node set is established. At the same time, the adjacent node list of each node in the neighborhood candidate node set is replaced and updated to complete the local edge reconstruction. Similar scene retrieval is performed by sorting the candidate nodes according to their access order. The query scene vector is received as the query vector. The entry node in the graph index is added to the candidate queue. The distance between the scene vector corresponding to each node in the candidate queue and the query vector is calculated and sorted in ascending order of distance. Candidate nodes are taken out in order of sorting results and the list of adjacent nodes of the candidate nodes is expanded to generate an expanded node set. The expanded node set is added to the candidate queue and the candidate queue sorting is updated. The process of recursively retrieving candidate nodes, expanding adjacency, updating candidate queues, and sorting candidate queues continues until a preset threshold for the number of accessed nodes is reached or the candidate queue no longer generates nodes with smaller distances. The output is a set of similar scenarios and the corresponding historical processing records.

[0013] Optionally, generating the set of executable instructions includes: Receive structured results, similar scenario sets, and historical handling records, and extract demand forecast results, risk warning results, decision suggestion quantity results, and historical handling action parameters; Based on the demand forecast results, risk warning results, decision suggestion results, and historical action parameters, an executable instruction set is generated, and each type of instruction is written with an object type field, an object identifier field, an operation type field, a quantity field, a node field, and a time window field. Based on the full-chain traceability data, batch number, expiration date, temperature zone and traceability code fields are written to the set of executable instructions, and document identifier and generation timestamp are written to each instruction.

[0014] Optionally, the generation of execution result records and traceability audit records includes: Output a set of executable instructions to the business execution end to trigger replenishment, transfer, picking and outbound, cold chain disposal and delivery adjustment, as well as recall or isolation disposal operations, and collect execution result data, and generate execution result records based on document identifiers and object identifiers; Collect traceability evidence data and generate traceability audit records, and store the traceability audit records and execution result records in association according to document identifier, batch number and traceability code.

[0015] The beneficial effects of this invention are: This invention extracts fields, aligns time, unifies units, aggregates, and cleans multi-source data from orders, inventory, transportation, traceability, and temperature control to form a standardized dataset with consistent definitions. Furthermore, it establishes unified entities and associated mappings for drug codes, batch numbers, traceability codes, expiration dates, temperature zones, boxes / pallets, waybills, and nodes, generating end-to-end traceability data. Compared to methods relying on scattered ledgers, manual verification, or single-system queries, this invention can sequentially link key events such as warehousing, picking, outbound, loading, in transit, signing, and temperature control data collection in chronological order, improving the consistency and traceability of end-to-end status awareness and shortening the time required for batch risk identification, responsibility link verification, and recall scope confirmation.

[0016] This invention employs an improved iTransformer model to perform multi-channel embedding and fusion of the variable Token, variable hierarchical grouping self-attention encoding, and multi-task structured decoding. It outputs demand forecasts, risk warnings, and decision suggestions, and combines near nearest neighbor search to construct a graph index to complete similar scenario retrieval. This generates executable instructions for replenishment, allocation, picking and outbound, cold chain disposal, delivery adjustment, and recall isolation, and forms execution results and traceability audit records. Compared with single-point prediction, static rules, or post-event report-based management, this invention achieves prediction and early warning linkage with business actions under the conditions of batch number, expiration date, temperature zone, and traceability code, enhancing the executability of instructions and the consistency of decisions. At the same time, it improves the stability and real-time performance of similar retrieval under data distribution changes through dynamic edge updates of the graph index and optimization of candidate access order within the structure. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0018] Figure 1 This is a flowchart of a pharmaceutical supply chain intelligent management method based on big data processing proposed in this invention; Figure 2 This is a structural block diagram of the improved iTransformer model for a pharmaceutical supply chain intelligent management method based on big data processing proposed in this invention. Figure 3 This is a functional diagram illustrating the approximate nearest neighbor search of a pharmaceutical supply chain intelligent management method based on big data processing proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1, Figure 2 and Figure 3 A smart management method for the pharmaceutical supply chain based on big data processing includes: Collect multi-source data from the pharmaceutical supply chain, preprocess the multi-source data, and generate a standardized dataset; A unified entity set is constructed based on a standardized dataset, and the relationship mapping between entities is established to generate full-link traceability relationship data. By tracing relationship data across the entire chain, a multivariate time series is constructed, and the historical sequence of each variable is used to construct a variable token. An improved iTransformer model is constructed. The variable Token is embedded and encoded separately through multi-channel embedding and then fused. Grouping self-attention with variable hierarchy awareness is introduced to group the variables and perform self-attention processing in sequence. A multi-task output head is introduced for parallel decoding to obtain structured results. The scene vectors are constructed from the structured results and written into the vector library. A graph index for vector library retrieval is established. Similar scene retrieval is performed on the scene vectors based on approximate nearest neighbor search. Local edge reconstruction is performed by updating the dynamic edges in the graph index. Sorting is optimized by using in-structure sorting. The output is a set of similar scenes and historical processing records. Based on the structured results, similar scenario sets, and historical disposal records, an executable instruction set is generated, and each instruction is associated with batch number, expiration date, temperature zone, and traceability code information; The set of executable instructions is output to the business execution end, and execution result records and traceability audit records are generated through execution result data and traceability evidence data.

[0021] In this embodiment, the multi-source data includes order data, inventory data, transportation data, traceability data, and temperature control data.

[0022] In this embodiment, generating the standardized dataset includes: Collect order data, inventory data, transportation data, traceability data, and temperature control data, and perform field extraction on the multi-source data, extracting the drug code field, batch number field, traceability code field, node identifier field, document identifier field, time field, quantity field, and status field; The field extraction results undergo standardization processing, which includes field mapping, data type conversion, unit unification, and time unification. Time fields are converted to timestamp fields, quantity fields are converted to units of measurement, and temperature control fields are converted to temperature units. Data that meet the criteria of consistent drug code, node identifier, and batch number and fall within the same time granularity are aggregated according to time granularity to generate aggregated records. Data cleaning is performed on the aggregated records. This includes constructing a deduplication key based on drug code, node identifier, batch number, timestamp, and record type, deleting duplicate records with the same deduplication key, filling in missing fields, setting the quantity field to zero for records with negative values, writing anomaly markers to temperature control records exceeding a preset temperature range, and outputting a standardized dataset. The preset temperature range is specifically defined as follows: The preset temperature range is determined by the temperature zone identifier and the drug code. A temperature upper and lower limit parameter table is established for each temperature zone identifier. The parameter table is then associated with the temperature zone identifier corresponding to the batch number. The temperature upper and lower limit parameter table is set into four categories according to the commonly used zoning of cold chain transportation and storage: frozen zone, refrigerated zone, cool zone, and room temperature zone. The frozen zone is set to -15 degrees Celsius to -30 degrees Celsius, the refrigerated zone is set to 8 degrees Celsius to 2 degrees Celsius, the cool zone is set to 20 degrees Celsius to 0 degrees Celsius, and the room temperature zone is set to 30 degrees Celsius to 15 degrees Celsius.

[0023] In this embodiment, generating end-to-end traceability data includes: Entity tables are created based on standardized datasets. These entity tables include entities for medicines, batches, traceability codes, nodes, storage locations, pallets, waybills, and orders. Specifically, creating these entity tables based on standardized datasets involves: The entity table is generated by deduplicating and aggregating the standardized dataset according to the entity primary key fields. The drug entity uses the drug code as the primary key and summarizes the fields of drug name, specifications and unit of measurement. The batch entity uses the drug code and batch number as the composite primary key and summarizes the fields of expiration date and temperature zone. The traceability code entity uses the traceability code as the primary key and associates the fields of drug code and batch number. The node entity uses the node identifier as the primary key and summarizes the fields of node type and address. The storage location entity uses the node identifier and storage location identifier as the composite primary key and summarizes the field of storage location type. The box and pallet entity uses the box code or pallet code as the primary key and associates the fields of drug code, batch number and traceability code. The waybill entity uses the waybill identifier as the primary key and summarizes the fields of carrier, dispatch node, arrival node and timestamp. The order entity uses the document identifier as the primary key and summarizes the fields of order type, node identifier, drug code, quantity and timestamp. A relational mapping table is generated based on the entity table. This table includes mappings for order-drug code, order-batch number, order-traceability code, batch number-expiration date, batch number-temperature zone, traceability code-box / pallet, box / pallet-waybill, waybill-node, node-warehouse location, and waybill-temperature control record. Each mapping record is written with an effective timestamp and an expiration timestamp. Specifically, generating the relational mapping table based on the entity table involves: The association mapping table is generated from the key-value pairs corresponding to the primary keys of each entity in the standardized dataset. The order mapping is generated by pairing the document identifier with the drug code, batch number, and traceability code respectively. The batch number mapping is generated by pairing the drug code and batch number with the expiration date and temperature zone respectively. The traceability code mapping is generated by pairing the traceability code with the box code or pallet code. The box and pallet mapping is generated by pairing the box code or pallet code with the waybill identifier. The waybill mapping is generated by pairing the waybill identifier with the dispatch node identifier, arrival node identifier, and temperature control collection record identifier respectively. The node mapping is generated by pairing the node identifier with the storage location identifier. The effective timestamp is the earliest timestamp of the paired record, and the expiration timestamp is the latest timestamp of the paired record. Full-chain traceability relationship data is generated based on the association mapping table and timestamps. The full-chain traceability relationship data uses the traceability code or batch number as the primary key, and the events are linked in sequence according to the timestamp order to output the full-chain traceability relationship data.

[0024] In this embodiment, constructing the historical sequence of each variable into a variable Token includes: The management objects are identified through end-to-end traceability data. Each management object consists of a node identifier and a drug code. A timeline and historical window length are determined for each management object. The timeline is composed of a continuous sequence of timestamps with time granularity. Specifically, the identification of management objects through end-to-end traceability data involves: The managed objects are obtained by grouping event records in the full-chain traceability data according to node identifiers and drug codes. Each group corresponds to a managed object. The node identifier is taken from the warehouse, distribution center, store or hospital node identifier in the event occurrence node field. The drug code is taken from the drug code field associated with the batch number or traceability code. Only the group with the number of events or transactions reaching the threshold of 20 within a statistical period of 30 consecutive days is retained as the managed object. Aggregating variable sequences for managed objects along a timeline by timestamp. These variable sequences include demand, inventory, in-transit, transportation timeliness, temperature control, traceability event, and compliance status sequences. Specifically, the aggregation of these variable sequences along the timeline by timestamp involves: Using the node identifier and drug code of the managed object as the filtering key, matching records are filtered in the standardized dataset and the full-link traceability relationship data. The filtered records are mapped to each time bucket on the time axis according to the timestamp. For each time bucket, the demand variable is calculated as the sum of the order quantities in the time bucket, the inventory variable is the inventory balance at the end of the time bucket, and if the inventory balance is missing, the balance after the inventory changes in the time bucket is used. The in-transit variable is the sum of the quantities that have been shipped but not signed for in the time bucket. The transportation timeliness variable is the average time difference between the signing timestamp and the shipping timestamp corresponding to the signing records in the time bucket. The temperature control variable is the maximum, minimum and average temperature in the time bucket. The traceability event variable is the number of times the events of warehousing, warehousing, in-transit, signing, abnormal handling and recall isolation occur in the time bucket. The compliance status variable is the remaining days of expiration, temperature zone category code and isolation status code in the time bucket. The calculation results of each time bucket are arranged in the order of the time axis to form a variable sequence. A multivariate time series matrix is ​​generated based on the variable sequences. The rows of the multivariate time series matrix correspond to timestamp sequences, and the columns correspond to variable sets. The historical window length is truncated from each column in the variable set to obtain the historical sequence of the variable. The historical sequences of the variables are then used to construct variable tokens, with the index of each token corresponding to a variable in the variable set. Specifically, constructing the historical sequences of the variables into variable tokens involves: For each column in the variable set, take consecutive historical window lengths of sampled values ​​in timestamp order to form a one-dimensional vector. Then, bind the one-dimensional vector with the column variable identifier to generate variable tokens. Numerical variable tokens directly use the one-dimensional vector as the input sequence, event variable tokens use the event count one-dimensional vector as the input sequence, and compliance status variable tokens use the one-dimensional vector obtained by concatenating the remaining days of validity, temperature zone code, and isolation status code as the input sequence. Assign a token index that is consistent with the column number of the variable set to each variable token, and then assemble all variable tokens into a variable token sequence according to the token index order.

[0025] In this embodiment, obtaining the structured result includes: An improved iTransformer model is constructed, comprising a multi-channel embedding module, a grouped self-attention encoding module, and a multi-task output head module, wherein: A multi-channel embedding module is embedded into the embedding layer of the variable token input. The single linear embedding method is replaced by independent embedding encoding and fusion mapping of numerical channel sequence, event channel sequence and compliance status channel sequence. A grouped self-attention encoding module is connected at the encoder self-attention calculation position. The unified attention calculation of all variable tokens is replaced by grouping rules of demand class, inventory and in-transit class, transportation and cold chain class, and quality and compliance class. After the self-attention output within the group, the self-attention calculation between the groups is concatenated to generate a global aggregate representation. Finally, at the output end, a multi-task output head module is used to replace the single-task prediction head. Parallel decoding is performed on the global aggregate representation to output the demand prediction result, risk warning result and decision suggestion result respectively. The multi-channel embedding module performs three-channel encoding processing on each variable Token, dividing each variable Token into a numerical channel sequence, an event channel sequence, and a compliance status channel sequence. It then performs intra-channel embedding encoding to generate numerical embedding vectors, event embedding vectors, and compliance embedding vectors, respectively. Finally, it performs a fusion mapping on the three types of embedding vectors to generate a variable Token representation vector, where: The three-channel encoding process for each variable Token is as follows: Based on the variable identifier, the variable Token is split into a numerical channel sequence, an event channel sequence, and a compliance status channel sequence. The numerical channel sequence consists of continuous sampled values ​​of demand, inventory, in transit, timeliness, and temperature statistics. The event channel sequence consists of counts or event markers of inbound, outbound, receipt, temperature control violation, damage reporting, and recall isolation events. The compliance status channel sequence consists of the remaining days of validity period, temperature zone code, and isolation status code aligned with timestamps. The process of generating numerical embedding vectors, event embedding vectors, and compliance embedding vectors through in-channel embedding encoding is as follows: For the numerical channel sequence, linear projection is performed to map the historical window length dimension to the embedding dimension to obtain the numerical embedding vector. For the event channel sequence, the event type is first mapped to the event index and an embedding lookup table is performed to obtain the event embedding sequence. Then, pooling or linear projection is performed on the event embedding sequence to obtain the event embedding vector. For the compliance status channel sequence, the temperature zone code and the isolation status code are respectively embedded and concatenated with the remaining days of validity period. Then, linear projection is performed to obtain the compliance embedding vector. Embedding lookup refers to using a trainable embedding matrix to convert discrete codes into continuous vectors. First, the temperature zone code, event type code, or isolation state code is represented as an integer index. Then, the row corresponding to the index is taken from the embedding matrix as the embedding vector of the code. The number of rows in the embedding matrix is ​​equal to the number of values ​​of the discrete code, and the number of columns is equal to the embedding dimension. During training, the matrix parameters are updated through backpropagation so that the embedding vectors corresponding to codes with similar semantics are closer in the vector space. The execution of the fusion mapping specifically involves: concatenating the numerical embedding vector, the event embedding vector, and the compliance embedding vector into a fusion vector, and performing a linear mapping on the fusion vector to a unified embedding dimension to obtain a variable Token representation vector; The grouped self-attention encoding module performs grouping and two-level self-attention processing on the variable Token representation vector. It groups the variables according to demand, inventory and in-transit, transportation and cold chain, and quality and compliance. Within each group, it performs multi-head self-attention operations on the variable Token representation vector to obtain the intra-group aggregate representation. Then, using the intra-group aggregate representation as input, it performs multi-head self-attention operations between groups to obtain the global aggregate representation. Where: The process of performing multi-head self-attention operation on the variable token representation vector within each group is as follows: An input matrix is ​​formed from the variable token representation vectors within the same group; a linear transformation is performed on the input matrix to generate a query matrix, a key matrix, and a value matrix; the query matrix and the key matrix are multiplied by their transposes to obtain a similarity matrix; the similarity matrix is ​​normalized row-wise to obtain an attention weight matrix; the attention weight matrix is ​​multiplied by the value matrix to obtain an attention output matrix; the operation is repeated for multiple attention heads, and the outputs of each attention head are concatenated along the feature dimension and linearly mapped to obtain the in-group output matrix; the in-group output matrix is ​​then used to generate an in-group aggregated representation using average pooling. The process of performing multi-head self-attention operations between groups using intra-group aggregation representations as input specifically involves: forming an inter-group input matrix by arranging the intra-group aggregation representations of the demand group, inventory and in-transit group, transportation and cold chain group, and quality and compliance group in group order; performing linear transformations on the inter-group input matrix to generate an inter-group query matrix, an inter-group key matrix, and an inter-group value matrix; multiplying the inter-group query matrix and the inter-group key matrix by transpose to obtain an inter-group similarity matrix; performing row-wise normalization on the inter-group similarity matrix to obtain an inter-group attention weight matrix; multiplying the inter-group attention weight matrix and the inter-group value matrix to obtain an inter-group attention output matrix; calculating and concatenating multiple attention heads and then performing linear mapping to obtain the inter-group output matrix; and generating a global aggregation representation from the inter-group output matrix using average pooling. The multi-task output header module performs parallel decoding on the global aggregate representation, outputting demand forecast results, risk warning results, and decision suggestion results respectively. The demand forecast result is the demand time series within the forecast step, the risk warning result is the risk probability time series within the forecast step, and the decision suggestion results are the replenishment suggestion time series and allocation suggestion time series within the forecast step, yielding structured results. The parallel decoding of the global aggregate representation specifically involves: The global aggregation representation is simultaneously input into the demand forecasting branch, the risk warning branch, and the decision suggestion branch. The three branches adopt different decoding structures and processing flows. The demand forecasting branch copies the global aggregation representation into a time step vector of prediction step size, and concatenates the time step index for each time step before inputting it into two fully connected layers to output the demand values ​​for each time step. The values ​​are arranged in the order of time steps to form a demand time series. The risk warning branch maps the global aggregation representation into a risk score of prediction step size through a fully connected layer, and then inputs it into a one-dimensional convolutional network to perform temporal shaping on the risk score sequence. The risk score sequence is then converted into a risk probability between zero and one by the Sigmoid function and arranged in the order of time steps to form a risk probability time series. The decision recommendation branch concatenates the global aggregated representation with the demand time series output by the demand forecast branch by time steps and inputs them into a fully connected layer to output the replenishment recommendation quantity series and the transfer recommendation quantity series respectively. ReLU activation is then performed on the output results to obtain the replenishment recommendation quantity time series and the transfer recommendation quantity time series.

[0026] In this embodiment, the output of the set of similar scenarios and historical handling records includes: Scenario vectors are constructed based on the structured results and written into the vector library. The scenario vectors are generated by splicing together the demand forecast results, risk warning results, decision suggestion results and management object identifiers, and a timestamp is written for each scenario vector. A graph index for approximate nearest neighbor search is constructed based on a vector library. The graph index consists of a node set and an edge set. The node set stores the scene vector identifier and scene vector value, and the edge set stores the list of adjacent nodes for each node. Local edge reconstruction is performed by dynamically updating the edges in the graph index. When a new scene vector is written to the vector library, a new node is generated in the graph index, and the new node identifier and the new scene vector value are registered in the node set. A set of neighboring candidate nodes is retrieved for the new node in the edge set, and edges are established between the new node and the neighboring candidate node set. Simultaneously, the adjacent node list of each node in the neighboring candidate node set is replaced and updated to complete the local edge reconstruction. Specifically, retrieving the set of neighboring candidate nodes for the new node in the edge set and establishing edges between the new node and the neighboring candidate node set involves: Using the newly added scene vector as the query vector, perform an approximate nearest neighbor search in the graph index. Calculate the distance between the newly added scene vector and the scene vector corresponding to the visited node and maintain a candidate queue. Select the top 30 nodes with the smallest distance from the candidate queue as the neighborhood candidate node set. Create an adjacent node list for the newly added node in the edge set and write the node identifier of the neighborhood candidate node set into the adjacent node list to form the outgoing edge set of the newly added node. At the same time, append the newly added node identifier to the adjacent node list corresponding to each node in the neighborhood candidate node set to form a reverse edge. Similar scene retrieval is performed by sorting candidate nodes according to their access order. The query scene vector is received as the query vector. Entry nodes from the graph index are added to the candidate queue. The distance between the scene vector corresponding to each node in the candidate queue and the query vector is calculated and sorted in ascending order of distance. Candidate nodes are then retrieved sequentially according to the sorting result, and the adjacent node list of each candidate node is expanded to generate an expanded node set. The expanded node set is added to the candidate queue, and the candidate queue sorting is updated. Specifically, the query scene vector is: Demand forecasting results, risk warning results, and decision suggestion results are extracted from the structured results. Each type of result is expanded into a one-dimensional vector according to the forecast step size, and then concatenated with the embedding vector of the management object identifier to obtain the query scenario vector. The embedding vector of the management object identifier is generated by concatenating the node identifier and the drug code after embedding and encoding them respectively. The process of recursively retrieving candidate nodes, expanding adjacencies, updating candidate queues, and sorting candidate queues continues until a preset threshold for the number of accessed nodes is reached or the candidate queue no longer generates nodes with smaller distances. The process outputs a set of similar scenarios and corresponding historical processing records. The preset threshold for the number of accessed nodes is specifically defined as follows: The preset threshold for the number of accessed nodes is set to four hundred nodes. During the retrieval process, each time a candidate node is taken from the candidate queue and expanded, an access is counted once. When the access count reaches four hundred, the loop stops and the twenty nodes with the smallest distance from the candidate queue are taken as a set of similar scenarios. When the minimum distance value in the candidate queue is not less than the currently recorded minimum distance value after fifty consecutive accesses, it is determined that the candidate queue will no longer generate nodes with smaller distances and the loop stops.

[0027] In this embodiment, generating the executable instruction set includes: Receive structured results, similar scenario sets, and historical handling records, and extract demand forecast results, risk warning results, decision suggestion quantity results, and historical handling action parameters; Based on demand forecasting results, risk warning results, decision suggestion results, and historical action parameters, an executable instruction set is generated. For each type of instruction, an object type field, object identifier field, operation type field, quantity field, node field, and time window field are written. Specifically, generating the executable instruction set involves: The structured results of parsing the management object identifier are used to obtain the demand forecast sequence, risk probability sequence, replenishment suggestion sequence and transfer suggestion sequence. The record with the most recent timestamp and consistent management object identifier is selected from the historical disposal action parameters as the parameter template. Replenishment instructions are generated based on the replenishment suggestion sequence, transfer instructions are generated based on the transfer suggestion sequence, and cold chain disposal and distribution adjustment instructions are generated based on the temperature control abnormality risk probability of not less than 0.7 in the risk probability sequence. Based on the recall risk probability being greater than 0.7 in the risk probability sequence, recall or isolation disposal instructions are generated. Based on the stockout risk probability being greater than 0.7 and the inventory variable being zero, picking and outbound instructions or alternative supply instructions are generated. Each instruction completes the source node, target node, waybill identifier, box code or pallet code fields according to the parameter template, and writes the object type field, object identifier field, operation type field, quantity field, node field and time window field, and summarizes them to form an executable instruction set. Based on the full-chain traceability data, batch number, expiration date, temperature zone and traceability code fields are written to the set of executable instructions, and document identifier and generation timestamp are written to each instruction.

[0028] In this embodiment, the generation of execution result records and traceability audit records includes: Output a set of executable instructions to the business execution end to trigger replenishment, transfer, picking and outbound, cold chain disposal and delivery adjustment, as well as recall or isolation disposal operations, and collect execution result data, and generate execution result records based on document identifiers and object identifiers; Collect traceability evidence data and generate traceability audit records, and store the traceability audit records and execution result records in association according to document identifier, batch number and traceability code.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to the pharmaceutical supply chain between a chain pharmacy system and a cold chain distribution center in Taoyuan. During the summer, the demand for insulin and multivalent vaccines fluctuates significantly among the stores. Data is scattered across five sources: orders, inventory, transportation, traceability, and temperature control. Mapping gaps frequently occur between batch numbers, traceability codes, waybills, and pallets, leading to problems such as stores receiving goods but being unable to quickly verify batch number and temperature control compliance, and stores experiencing stock shortages but struggling to promptly restock from neighboring stores. The platform integrates data from the distribution center and 12 stores, accumulating 186,000 order records, 920,000 inventory change records, 54,000 transportation node events, 3.1 million traceability code scans, and 8.6 million temperature control data collection points, involving 1,240 drug codes and 3,980 batch numbers, with refrigerated areas accounting for 46%.

[0030] In the application process, the five types of data are first preprocessed into a standardized dataset. Drug codes, batch numbers, traceability codes, expiration dates, temperature zones, pallets, waybills, nodes, and storage locations are then mapped as unified entities, enabling each traceability code to be linked to inbound, picking, outbound, loading, in-transit, arrival, signing, and temperature control data collection events on a timeline. Subsequently, multivariate time series are generated using store or distribution center node identifiers and drug codes as management objects. Variable tokens are constructed and fed into the improved iTransformer model. Multi-channel embedding simultaneously encodes numerical values, events, and compliance status. Grouped self-attention aggregation is performed hierarchically based on demand, in-transit inventory, cold chain transportation, and quality compliance. Multi-task output heads simultaneously provide demand forecasts, risk probabilities, and replenishment or transfer recommendations. The structured results are concatenated into scene vectors and written into the vector library. Approximate nearest neighbor search uses graph index to dynamically update edges to maintain index quality, and the intra-structure sorting of candidate access order is used to stably retrieve similar scenes. Finally, executable instructions for replenishment, allocation, picking and outbound, cold chain disposal, distribution adjustment and isolation recall are generated, and the instructions are bound with batch number, expiration date, temperature zone and traceability code to form an audit chain.

[0031] On the same day that two stores received mixed shipments of the same product from different batches, the platform quickly located the anomaly to the corresponding transportation task by linking the traceability code with the pallet, waybill, and node events. It also identified a short-term boundary violation signal in the temperature control data collection event and discovered an abnormal trajectory of the same pallet being split during the signing and scanning events. The platform automatically generates isolation and verification instructions, requiring the affected batch numbers and traceability code ranges to be isolated and then verified. Simultaneously, it generates cross-store transfer and replenishment instructions to cover the day's prescription needs. After the instructions are executed, a complete execution result record and traceability audit record are generated. The audit record includes the batch number, expiration date, temperature zone, traceability code, temperature control evidence, and signing evidence, supporting one-click verification and traceability queries for the destination, temperature control compliance, and closed-loop handling of the same batch between stores and distribution centers.

[0032] Table 1 Comparison of Intelligent Management Methods for the Pharmaceutical Supply Chain

[0033] As shown in Table 1, the method of this invention demonstrates the most significant advantages in terms of end-to-end perceptibility and traceability, particularly in both traceability and location time and temperature control anomaly detection rate. The traceability and location time is 4.8 minutes, significantly lower than the rule-driven method's 27.5 minutes, ARIMA's 24.1 minutes, traditional Transformer's 18.6 minutes, LSTM+rules' 20.9 minutes, and static approximate nearest neighbor search's 15.7 minutes. This indicates that unified entity and association mapping, along with structured accumulation of audit links, effectively reduces cross-system verification and traceability query time. The temperature control anomaly detection rate reaches 96.2%, a significant improvement compared to rule-driven (71.4%), ARIMA (68.9%), traditional Transformer (82.6%), LSTM+rules' 79.1%, and static approximate nearest neighbor search (84.3%). This suggests that multi-task risk output and similar scenario retrieval linkage are more conducive to detecting temperature control out-of-bounds anomalies.

[0034] From the perspective of forecasting and operational results indicators, this invention achieves optimal or near-optimal results in demand forecasting MAPE, stockout rate, and expired loss rate. The demand forecasting MAPE is 8.9%, superior to the traditional Transformer's 11.8%, static approximate nearest neighbor search's 12.9%, LSTM+rules' 13.6%, and ARIMA's 15.4%, and significantly lower than the rule-driven method's 21.7%, indicating that the improved iTransformer's multi-channel embedding and grouped self-attention can more fully characterize the coupling relationship between demand, inventory, transportation cold chain, and compliance status. The stockout rate is reduced to 1.7%, lower than rule-driven (4.9%), ARIMA (3.8%), traditional Transformer (3.1%), LSTM+rules' 3.6%, and static approximate nearest neighbor search's 3.3%. The expired loss rate is 0.42%, also significantly better than the compared methods, demonstrating the improvement in inventory structure and expiration date risk control through the linkage of forecasting and early warning with replenishment and allocation instructions.

[0035] In terms of decision executability and retrieval performance, this invention also leads in allocation hit rate and instruction execution rate, while maintaining low retrieval latency. The allocation hit rate reaches 88.5%, significantly higher than the traditional Transformer's 74.2%, static approximate nearest neighbor search 71.1%, LSTM+rules' 69.8%, ARIMA's 61.7%, and rule-driven 52.3%, indicating that with experience in handling similar scenarios, the allocation suggestions are closer to a realistic and feasible supply-demand balance path. The instruction execution rate is 93.4%, higher than the comparative methods' 68.6% to 81.5%, indicating that instructions with complete output fields and bound to traceable elements are easier to implement. Regarding retrieval latency, this invention is 12ms, lower than the static ANN's 35ms, demonstrating that the structural sorting optimization of the candidate access order improves retrieval efficiency while ensuring retrieval effectiveness.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent management of the pharmaceutical supply chain based on big data processing, characterized in that, include: Collect multi-source data from the pharmaceutical supply chain, preprocess the multi-source data, and generate a standardized dataset; A unified entity set is constructed based on a standardized dataset, and the relationship mapping between entities is established to generate full-link traceability relationship data. By tracing relationship data across the entire chain, a multivariate time series is constructed, and the historical sequence of each variable is used to construct a variable token. An improved iTransformer model is constructed. The variable Token is embedded and encoded separately through multi-channel embedding and then fused. Grouping self-attention with variable hierarchy awareness is introduced to group the variables and perform self-attention processing in sequence. A multi-task output head is introduced for parallel decoding to obtain structured results. The scene vectors are constructed from the structured results and written into the vector library. A graph index for vector library retrieval is established. Similar scene retrieval is performed on the scene vectors based on approximate nearest neighbor search. Local edge reconstruction is performed by updating the dynamic edges in the graph index. Sorting is optimized by using in-structure sorting. The output is a set of similar scenes and historical processing records. Based on the structured results, similar scenario sets, and historical disposal records, an executable instruction set is generated, and each instruction is associated with batch number, expiration date, temperature zone, and traceability code information; The set of executable instructions is output to the business execution end, and execution result records and traceability audit records are generated through execution result data and traceability evidence data.

2. The intelligent management method for the pharmaceutical supply chain based on big data processing according to claim 1, characterized in that, The multi-source data includes order data, inventory data, transportation data, traceability data, and temperature control data.

3. The intelligent management method for the pharmaceutical supply chain based on big data processing according to claim 1, characterized in that, The generation of the standardized dataset includes: Collect order data, inventory data, transportation data, traceability data, and temperature control data, and perform field extraction on the multi-source data, extracting the drug code field, batch number field, traceability code field, node identifier field, document identifier field, time field, quantity field, and status field; The field extraction results undergo standardization processing, which includes field mapping, data type conversion, unit unification, and time unification. Time fields are converted to timestamp fields, quantity fields are converted to units of measurement, and temperature control fields are converted to temperature units. Data that meet the criteria of consistent drug code, node identifier, and batch number and fall within the same time granularity are aggregated according to time granularity to generate aggregated records. Data cleaning is performed on the aggregated records. The data cleaning process includes constructing a deduplication key based on the drug code, node identifier, batch number, timestamp and record type and deleting duplicate records with the same deduplication key, filling missing fields, setting the quantity field to zero for records with negative values, writing anomaly markers to temperature control records that exceed the preset temperature range, and outputting a standardized dataset.

4. The intelligent management method for the pharmaceutical supply chain based on big data processing according to claim 1, characterized in that, The generation of end-to-end traceability relationship data includes: Entity tables are created based on standardized datasets. The entity tables include drug entities, batch entities, traceability code entities, node entities, storage location entities, container entities, waybill entities, and order entities. An association mapping table is generated based on the entity table. The association mapping table includes order-drug code mapping, order-batch number mapping, order-traceability code mapping, batch number-expiration date mapping, batch number-temperature zone mapping, traceability code-box pallet mapping, box pallet-waybill mapping, waybill-node mapping, node-warehouse location mapping, and waybill-temperature control record mapping. An effective timestamp and an expiration timestamp are written for each mapping record. Full-chain traceability relationship data is generated based on the association mapping table and timestamps. The full-chain traceability relationship data uses the traceability code or batch number as the primary key, and the events are linked in sequence according to the timestamp order to output the full-chain traceability relationship data.

5. The intelligent management method for the pharmaceutical supply chain based on big data processing according to claim 1, characterized in that, The process of constructing a variable token from the historical sequence of each variable includes: The management objects are identified by tracing the relationship data of the entire chain. The management objects consist of node identifiers and drug codes. A timeline and historical window length are determined for each management object. The timeline consists of a continuous timestamp sequence with time granularity. Aggregate variable sequences for managed objects on the timeline by timestamp. The variable sequences include demand variable sequences, inventory variable sequences, in-transit variable sequences, transportation timeliness variable sequences, temperature control variable sequences, traceability event variable sequences, and compliance status variable sequences. A multivariate time series matrix is ​​generated based on the variable sequence. The rows of the multivariate time series matrix correspond to the timestamp sequence, and the columns correspond to the variable set. The historical window length is truncated for each column in the variable set to obtain the historical sequence of the variable. The historical sequence of the variable is used to construct a variable token, and the index of the variable token corresponds to the variable in the variable set.

6. The intelligent management method for the pharmaceutical supply chain based on big data processing according to claim 1, characterized in that, The obtained structured results include: An improved iTransformer model is constructed, including a multi-channel embedding module, a grouped self-attention encoding module, and a multi-task output head module; The multi-channel embedding module performs three-channel encoding processing on each variable Token, dividing each variable Token into a numerical channel sequence, an event channel sequence, and a compliance status channel sequence. It then performs in-channel embedding encoding to generate numerical embedding vectors, event embedding vectors, and compliance embedding vectors, and performs fusion mapping on the three types of embedding vectors to generate variable Token representation vectors. The grouped self-attention encoding module performs grouping and two-level self-attention processing on the variable Token representation vector. It groups the variables according to demand, inventory and in-transit, transportation and cold chain, and quality and compliance. Within each group, it performs multi-head self-attention operation on the variable Token representation vector to obtain the group aggregate representation. Then, it uses the group aggregate representation as input to perform multi-head self-attention operation between groups to obtain the global aggregate representation. The multi-task output header module performs parallel decoding on the global aggregate representation and outputs demand forecast results, risk warning results, and decision suggestion results respectively. The demand forecast results are the demand time series within the forecast step, the risk warning results are the risk probability time series within the forecast step, and the decision suggestion results are the replenishment suggestion time series and allocation suggestion time series within the forecast step, thus obtaining structured results.

7. The intelligent management method for the pharmaceutical supply chain based on big data processing according to claim 1, characterized in that, The output set of similar scenarios and historical processing records include: Scenario vectors are constructed based on the structured results and written into the vector library. The scenario vectors are generated by splicing together the demand forecast results, risk warning results, decision suggestion results and management object identifiers, and a timestamp is written for each scenario vector. A graph index for approximate nearest neighbor search is constructed based on a vector library. The graph index consists of a node set and an edge set. The node set stores the scene vector identifier and scene vector value, and the edge set stores the list of adjacent nodes for each node. The local edge reconstruction is performed by updating the dynamic edge in the graph index. When a new scene vector is written to the vector library, a new node is generated in the graph index, and the new node identifier and the new scene vector value are registered in the node set. The neighborhood candidate node set is retrieved for the new node in the edge set and the edge between the new node and the neighborhood candidate node set is established. At the same time, the adjacent node list of each node in the neighborhood candidate node set is replaced and updated to complete the local edge reconstruction. Similar scene retrieval is performed by sorting the candidate nodes according to their access order. The query scene vector is received as the query vector. The entry node in the graph index is added to the candidate queue. The distance between the scene vector corresponding to each node in the candidate queue and the query vector is calculated and sorted in ascending order of distance. Candidate nodes are taken out in order of sorting results and the list of adjacent nodes of the candidate nodes is expanded to generate an expanded node set. The expanded node set is added to the candidate queue and the candidate queue sorting is updated. The process of recursively retrieving candidate nodes, expanding adjacency, updating candidate queues, and sorting candidate queues continues until a preset threshold for the number of accessed nodes is reached or the candidate queue no longer generates nodes with smaller distances. The output is a set of similar scenarios and the corresponding historical processing records.

8. The intelligent management method for the pharmaceutical supply chain based on big data processing according to claim 1, characterized in that, The generation of the executable instruction set includes: Receive structured results, similar scenario sets, and historical handling records, and extract demand forecast results, risk warning results, decision suggestion quantity results, and historical handling action parameters; Based on the demand forecast results, risk warning results, decision suggestion results, and historical action parameters, an executable instruction set is generated, and each type of instruction is written with an object type field, an object identifier field, an operation type field, a quantity field, a node field, and a time window field. Based on the full-chain traceability data, batch number, expiration date, temperature zone and traceability code fields are written to the set of executable instructions, and document identifier and generation timestamp are written to each instruction.

9. The intelligent management method for the pharmaceutical supply chain based on big data processing according to claim 1, characterized in that, The generated execution result records and traceability audit records include: Output a set of executable instructions to the business execution end to trigger replenishment, transfer, picking and outbound, cold chain disposal and delivery adjustment, as well as recall or isolation disposal operations, and collect execution result data, and generate execution result records based on document identifiers and object identifiers; Collect traceability evidence data and generate traceability audit records, and store the traceability audit records and execution result records in association according to document identifier, batch number and traceability code.

Images