A material in-out warehouse intelligent prediction method based on big data analysis

Abstract

The application discloses a kind of based on big data analysis's material warehousing intelligent prediction method, it is related to big data analysis technical field, including: collecting material warehousing business data, obtains standardization material warehousing data;Generation material business object set, constructs material-business object two department correlation diagram;Execute SimRank algorithm, obtain business hierarchical structure similarity matrix;Build target material itself flow chain, build business hierarchical similarity lattice chain set;Get implicit structure lattice chain sequence;Build improved Lattice LSTM model, obtain warehousing prediction state representation;Generation target material's warehousing prediction result, inventory risk early warning and warehousing scheduling suggestion.The application introduces improved Lattice LSTM model and SimRank algorithm, realizes the intelligent prediction of target material warehousing quantity, warehouse quantity, inventory balance, replenishment demand and inventory risk state.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and in particular to an intelligent prediction method for the inbound and outbound of materials based on big data analytics. Background Technology

[0002] As enterprise warehousing scale expands and production pace accelerates, material inbound and outbound management increasingly involves multiple business processes, including procurement, warehousing, inventory, requisition, replenishment, stockouts, and production tasks. Different materials often exhibit relationships such as substitution, shared requisition, periodic replenishment, and concentrated consumption. Material inventory status is also affected by order arrival times, changes in production plans, fluctuations in requisition frequency, and supply response cycles. Existing warehouse management systems can typically record changes in material quantities and the flow of business documents. However, in scenarios with numerous material types, long business chains, and complex data sources, inventory changes exhibit significant temporality, correlation, and uncertainty. Relying solely on manual experience or fixed inventory thresholds is insufficient to reflect the true material demand in a timely manner.

[0003] Existing material inbound / outbound forecasting technologies are mostly based on historical inbound and outbound quantities and current inventory levels, combined with statistical analysis, time series forecasting, or conventional recurrent neural network models to predict future demand changes for materials. While these basic methods can handle general quantity trends, they do not adequately consider the business transmission relationships between procurement, requisition, replenishment, and stockouts, and struggle to differentiate the similarity impact of different materials across different business layers. When materials experience continuous outbound shipments, critical inventory levels, delayed replenishment, or recurring stockouts, existing methods are prone to problems such as forecast lag, misjudgment of similar materials, inaccurate replenishment recommendations, and untimely inventory risk warnings.

[0004] Therefore, how to provide an intelligent prediction method for the inbound and outbound of materials based on big data analysis 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 prediction method for material inbound and outbound operations based on big data analysis. This invention fully utilizes big data analysis, material business association modeling, and time-series status prediction technologies. It details the complete implementation process from material inbound and outbound business data collection, material business object association, similar material identification, inbound / outbound flow chain construction, to the prediction using an improved Lattice LSTM model. This enables intelligent prediction of the future inbound quantity, outbound quantity, inventory balance, replenishment demand, and inventory risk status of target materials. This invention comprehensively considers the business impact relationships between procurement, requisition, replenishment, stockouts, and production tasks, avoiding the prediction lag caused by relying solely on historical averages or fixed thresholds. It possesses advantages such as high prediction accuracy, timely inventory risk identification, reasonable replenishment scheduling, and a high degree of intelligent material management.

[0006] According to an embodiment of the present invention, a method for intelligent prediction of material inbound and outbound operations based on big data analysis includes:

[0007] Collect data on material inbound and outbound operations, and preprocess the data to obtain standardized material inbound and outbound data.

[0008] Based on standardized material inbound and outbound data, materials are grouped according to the business relationships of procurement, requisition, replenishment and shortage, generating a set of material business objects, and constructing a two-part relationship diagram of material-business objects;

[0009] Based on the material-business object bipartite association graph, the SimRank algorithm is executed to calculate the structural similarity between the target material and the remaining material nodes at the procurement object layer, the requisition object layer, the replenishment object layer, and the shortage object layer, respectively, to obtain the business layer structure similarity matrix, and to filter the procurement similar materials, requisition similar materials, replenishment similar materials, and shortage similar materials corresponding to the target material.

[0010] Construct the target material's own circulation chain based on the inbound and outbound circulation sequence of the target material, and construct a business-layered similar grid chain set based on the circulation sequence of similar materials purchased, similar materials issued, similar materials replenished, and similar materials out of stock;

[0011] Write the corresponding structural similarity in the business layer structure similarity matrix into the business layer similarity grid chain set, and use it as the grid chain propagation weight to obtain the implicit structure grid chain sequence.

[0012] An improved Lattice LSTM model is constructed based on implicit structure lattice chain sequences. State propagation is performed along the target material's own flow chain and the set of similar lattice chains in the business hierarchy. Gating fusion is then performed based on the lattice chain propagation weights to obtain the inbound and outbound prediction state representation.

[0013] Based on the inbound and outbound forecast status, the system generates inbound and outbound forecast results for target materials, inventory risk warnings, and inbound and outbound scheduling suggestions.

[0014] Optionally, the material inbound and outbound business data specifically includes basic material information data, historical inbound record data, historical outbound record data, inventory balance data, safety stock data, purchase order data, requisition application data, replenishment application data, stockout record data, production task data, and business collection timestamp data.

[0015] Optionally, the preprocessing of material inbound and outbound business data specifically includes material coding unification, timestamp alignment, quantity unit conversion, abnormal record removal, and missing field completion.

[0016] Optionally, the material-business object binary association diagram includes:

[0017] Based on the purchase order number, requisition application number, production task identifier, replenishment application number, business collection timestamp, inventory balance and safety stock quantity, standardized material inbound and outbound data are aggregated to generate a set of purchase objects, a set of requisition objects, a set of replenishment objects and a set of out-of-stock objects.

[0018] The sets of procurement objects, requisition objects, replenishment objects, and shortage objects are merged into a set of material business objects. Material node sets and business object node sets are constructed based on material codes and the set of material business objects, respectively.

[0019] Based on the set of material nodes, the set of business object nodes, and the relationship between them, a connection edge is established to obtain the material-business object bipartite association graph.

[0020] Optionally, the similar materials for procurement, similar materials for requisition, similar materials for replenishment, and similar materials for shortage corresponding to the target materials include:

[0021] The SimRank algorithm is executed to divide the material-business object bipartite association graph into the procurement object layer, the requisition object layer, the replenishment object layer, and the shortage object layer. The business path between the target material node and the remaining material node is extracted, and the corresponding inbound and outbound flow phases are identified.

[0022] Based on standardized material inbound and outbound data, semantic tags are applied to business paths to obtain heterogeneous path semantic sets. Priority, lag, and synchronization relationships are identified according to the business transmission order of requisition, stockout, replenishment, and procurement to generate business transmission relationships.

[0023] Based on standardized material inbound and outbound data, causal gating is performed on heterogeneous path semantic sets, inbound and outbound flow phases, and business transmission relationships to obtain gating path semantic sets. SimRank iterative calculation is then performed to obtain the structural similarity of the procurement layer, the requisition layer, the replenishment layer, and the shortage layer.

[0024] Based on the closed-loop flow relationship between outbound, stockout, replenishment and inbound, the similarity of the procurement layer structure, the requisition layer structure, the replenishment layer structure, and the stockout layer structure are enhanced in a closed loop. Conflict suppression is achieved by combining alternative outbound relationships, reverse inventory change relationships, and alternating requisition relationships, resulting in a business layer structure similarity matrix.

[0025] Based on the business hierarchical structure similarity matrix, the target materials are filtered to identify similar materials for procurement, requisition, replenishment, and shortage.

[0026] Optionally, the step of constructing a business-layered similar grid chain set based on the flow sequence of similar materials purchased, similar materials issued, similar materials replenished, and similar materials out of stock includes:

[0027] Based on standardized material inbound and outbound data, the system reads historical inbound records, historical outbound records, inventory balance data, replenishment request data, and stockout records of the target material according to the material code and business collection timestamp, generates the inbound and outbound flow sequence of the target material, and connects them in time sequence to obtain the flow chain of the target material itself.

[0028] Based on standardized material inbound and outbound data, historical inbound records, historical outbound records, inventory balance data, replenishment application data, and shortage record data for similar materials are extracted according to material codes and business collection timestamps. Corresponding similar material circulation sequences are generated and aligned according to the time sequence of the target material's own circulation chain. Procurement similar grid chains, requisition similar grid chains, replenishment similar grid chains, and shortage similar grid chains are constructed respectively to obtain a business-layered similar grid chain set.

[0029] Optionally, obtaining the implicit lattice chain sequence includes:

[0030] Based on the business hierarchical structure similarity matrix, the structural similarity of each similar material in the business hierarchical similar grid chain set is matched according to the material code and object layer type to generate grid chain similarity mapping results;

[0031] Based on the business collection timestamp, the grid chain similarity mapping results are written into the business layer similar grid chain set, and the structural similarity within the same object layer is normalized to obtain the grid chain propagation weight.

[0032] The target material's own circulation chain, the set of similar grid chains in business layering, and the grid chain propagation weights are serialized and combined according to the business collection timestamp to obtain an implicit structure grid chain sequence.

[0033] Optionally, obtaining the inbound / outbound prediction status representation includes:

[0034] An improved Lattice LSTM model is constructed, which includes a lattice chain input unit, a phase transition pulse reconstruction unit, an auction contribution correction unit, and a closed-loop responsibility output unit.

[0035] The grid chain input unit parses the implicit structure grid chain sequence to obtain the target material's own circulation chain, the business layer similar grid chain set, and the grid chain propagation weight. Based on standardized material inbound and outbound data, it identifies the inventory stage, phase transition boundary, and inventory change pulse. It then segments the target material's own circulation chain and the business layer similar grid chain set to obtain the target material phase transition circulation segment and the business layer similar grid chain segment set.

[0036] The phase change pulse reconstruction unit reconnects the similar grid chains for procurement, requisition, replenishment, and shortage based on inventory stages, phase change boundaries, and inventory change pulses. It also propagates the state along the phase change flow segments of the target material and the set of similar grid chain segments in the business layer to obtain the main chain memory state of the target material and the reconstructed grid chain propagation state.

[0037] The auction contribution correction unit sorts the write priority of the purchase gate, requisition gate, replenishment gate and stockout gate based on the grid chain propagation weight, inventory stage, historical prediction contribution and conflict suppression results, and controls the corresponding similar grid chains to perform main write, auxiliary write, propagation shielding and reverse memory suppression to obtain the corrected grid chain memory state.

[0038] The closed-loop responsibility output unit performs closed-loop inversion and write-back on the main chain memory state of the target material and the corrected grid chain memory state based on the closed-loop sequence of outbound, inventory decrease, stockout, replenishment and inbound recovery. It also performs gating fusion on the output states of the procurement gate, requisition gate, replenishment gate and stockout gate based on the grid chain propagation weight to obtain the inbound and outbound prediction state representation.

[0039] The improved Lattice LSTM model was trained using inbound / outbound prediction error, inventory balance prediction error, replenishment demand prediction error, and inventory risk identification error as joint optimization objectives. The parameters of the lattice input unit, phase transition pulse reconstruction unit, auction contribution correction unit, and closed-loop responsibility output unit were optimized. The training was considered to have converged when the decrease of the joint optimization objective was less than 0.001 in 10 consecutive training rounds, or when the number of training rounds reached 200.

[0040] Optionally, the pulse-driven reconnection of the procurement similar grid chain, the requisition similar grid chain, the replenishment similar grid chain, and the out-of-stock similar grid chain includes:

[0041] Based on the phase change flow segments of target materials, the set of similar grid chain segments in business layering, and inventory change pulses, the pulse triggering position is identified and the corresponding similar grid chain is matched to obtain the pulse grid chain matching result;

[0042] Based on the pulse grid chain matching results, the outbound pulse is connected to the requisition similar grid chain and the shortage similar grid chain, the inventory decrease pulse is connected to the shortage similar grid chain and the replenishment similar grid chain, the replenishment pulse is connected to the replenishment similar grid chain and the procurement similar grid chain, and the inbound pulse is connected to the procurement similar grid chain and the target material's own circulation chain, thus generating a pulse-driven reconnection relationship.

[0043] The grid chain connection is updated based on the pulse-driven reconnection relationship. The out-of-stock similar grid chain status, replenishment similar grid chain status, and procurement similar grid chain status are sequentially written into the replenishment similar grid chain, the procurement similar grid chain, and the target material's own circulation chain to obtain the reconnected business-layered similar grid chain set.

[0044] Optionally, the step of generating inbound / outbound forecast results, inventory risk warnings, and inbound / outbound scheduling suggestions for target materials based on the inbound / outbound forecast status includes:

[0045] Based on the inbound and outbound forecast status representation, combined with the historical inbound record data, historical outbound record data, inventory balance data and business collection timestamp data of the target materials, the forecasted inbound quantity, forecasted outbound quantity and forecasted inventory balance of the target materials within the forecast time range are generated, and the inbound and outbound forecast results are obtained.

[0046] Based on inbound / outbound forecasts, safety stock data, replenishment request data, and stockout records, the system identifies target materials as being in a state where their inventory is below safety stock, in a state of continuous outbound consumption, and in a state of insufficient replenishment response, and generates inventory risk warnings.

[0047] Based on the inbound / outbound forecast results and inventory risk warnings, combined with purchase order data, requisition application data, replenishment application data and production task data, suggestions are generated for the replenishment quantity of target materials, inbound time, outbound preparation and production requisition scheduling, resulting in inbound / outbound scheduling suggestions.

[0048] The beneficial effects of this invention are:

[0049] This invention discloses an intelligent prediction method for material inbound and outbound operations based on big data analysis. By collecting business data such as basic material information, historical inbound records, historical outbound records, inventory balance, safety stock, purchase orders, requisition requests, replenishment requests, stockout records, and production tasks, and by uniformly processing data from different sources, the prediction of material inbound and outbound operations no longer relies solely on single inventory quantities or human experience. Instead, it can extract more complete prediction basis from the business relationships between procurement, requisition, replenishment, stockouts, and production tasks, thereby improving the sufficiency of data utilization and the reliability of the prediction basis.

[0050] This invention constructs a two-part association diagram of materials and business objects, distinguishing between the procurement object layer, the requisition object layer, the replenishment object layer, and the shortage object layer. By combining structural similarity calculation, closed-loop flow enhancement, and conflict suppression, it can more accurately identify similar materials with real business relationships with the target materials, reduce misjudgments caused by substitute outbound, reverse inventory changes, and alternating requisition, and improve the accuracy of similar material screening results.

[0051] This invention constructs a target material's own circulation chain and a set of business-layered similar grid chains, and uses an improved Lattice LSTM model to predict inbound and outbound status. This allows the improved Lattice LSTM model to simultaneously express the temporal changes of the target material itself and the business transmission impact of similar materials. Through phase-change pulse reconstruction, gated write correction, and closed-loop responsibility output, this invention can promptly capture the continuous changing relationships between increases in outbound shipments, decreases in inventory, stockout triggers, replenishment responses, and inbound recovery. This improves the accuracy of predicting inbound quantity, outbound quantity, inventory balance, replenishment demand, and inventory risk, and provides a more rational decision-making basis for replenishment quantity, inbound time, outbound preparation, and production requisition scheduling. Attached Figure Description

[0052] 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:

[0053] Figure 1 This is a flowchart of an intelligent prediction method for material inbound and outbound operations based on big data analysis proposed in this invention.

[0054] Figure 2 This is a data flow diagram of the SimRank algorithm for a smart prediction method for material inbound and outbound operations based on big data analysis proposed in this invention.

[0055] Figure 3 This is a schematic diagram of the improved Lattice LSTM model, which is a method for intelligent prediction of material inbound and outbound operations based on big data analysis proposed in this invention. Detailed Implementation

[0056] 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.

[0057] refer to Figure 1 , Figure 2 and Figure 3 A method for intelligent prediction of material inbound and outbound operations based on big data analysis, comprising:

[0058] Collect data on material inbound and outbound operations, and preprocess the data to obtain standardized material inbound and outbound data.

[0059] Based on standardized material inbound and outbound data, materials are grouped according to the business relationships of procurement, requisition, replenishment and shortage, generating a set of material business objects, and constructing a two-part relationship diagram of material-business objects;

[0060] Based on the material-business object bipartite association graph, the SimRank algorithm is executed to calculate the structural similarity between the target material and the remaining material nodes at the procurement object layer, the requisition object layer, the replenishment object layer, and the shortage object layer, respectively, to obtain the business layer structure similarity matrix, and to filter the procurement similar materials, requisition similar materials, replenishment similar materials, and shortage similar materials corresponding to the target material.

[0061] Construct the target material's own circulation chain based on the inbound and outbound circulation sequence of the target material, and construct a business-layered similar grid chain set based on the circulation sequence of similar materials purchased, similar materials issued, similar materials replenished, and similar materials out of stock;

[0062] Write the corresponding structural similarity in the business layer structure similarity matrix into the business layer similarity grid chain set, and use it as the grid chain propagation weight to obtain the implicit structure grid chain sequence.

[0063] An improved Lattice LSTM model is constructed based on implicit structure lattice chain sequences. State propagation is performed along the target material's own flow chain and the set of similar lattice chains in the business hierarchy. Gating fusion is then performed based on the lattice chain propagation weights to obtain the inbound and outbound prediction state representation.

[0064] Based on the inbound and outbound forecast status, the system generates inbound and outbound forecast results for target materials, inventory risk warnings, and inbound and outbound scheduling suggestions.

[0065] In this embodiment, the material inbound and outbound business data specifically includes basic material information data, historical inbound record data, historical outbound record data, inventory balance data, safety stock data, purchase order data, requisition application data, replenishment application data, stockout record data, production task data, and business collection timestamp data.

[0066] In this embodiment, the preprocessing of material inbound and outbound business data specifically includes material code unification, timestamp alignment, quantity unit conversion, abnormal record removal, and missing field completion, wherein:

[0067] Quantity unit conversion refers to converting the quantities recorded in the material inbound and outbound business data, which are measured in boxes, packages, pieces, kilograms, tons, meters, and rolls, into the corresponding base unit of measurement for the materials in the basic material information data.

[0068] In this embodiment, the material-business object two-part relationship diagram includes:

[0069] Based on purchase order number, requisition number, production task identifier, replenishment request number, business data collection timestamp, inventory balance, and safety stock quantity, standardized material inbound and outbound data are aggregated to generate sets of procurement objects, requisition objects, replenishment objects, and shortage objects, among which:

[0070] The business data for standardized materials entering and leaving the warehouse is collected and aggregated, specifically as follows:

[0071] Read the purchase order number from the purchase order data, group records with the same purchase order number into a purchase record group by material code, purchase quantity, and business collection timestamp, and generate a purchase object using the purchase order number as an identifier. Summarize all purchase objects to form a purchase object set. Read the requisition application data and production task data, group records with the same requisition application number and production task identifier into a requisition record group by material code, requisition quantity, and business collection timestamp, and generate a requisition object using a combined identifier. Summarize all requisition objects to form a requisition object set. Read the replenishment application data, group records with the same replenishment application number into a replenishment record group by material code, replenishment quantity, and business collection timestamp, and generate a replenishment object set. Read the inventory balance data and safety stock data, filter records with inventory balance less than safety stock, and group them into a shortage record group by collection date, and generate a shortage object set.

[0072] The sets of procurement objects, requisition objects, replenishment objects, and shortage objects are merged into a set of material business objects. Based on the material code and the set of material business objects, sets of material nodes and business object nodes are constructed respectively, where:

[0073] The construction of the material node set and the business object node set is as follows:

[0074] The material codes in the standardized material inbound and outbound data are deduplicated and merged, and the material name, benchmark unit of measurement and material category are retained to generate a set of material nodes. For the set of procurement objects, the set of requisition objects, the set of replenishment objects and the set of out-of-stock objects, business object nodes are established one by one, and the business object type, identifier, timestamp and corresponding material code are written to generate a set of business object nodes.

[0075] Based on the set of material nodes, the set of business object nodes, and the attribution relationship between them, connection edges are established to obtain a binary association graph of material-business object, where:

[0076] The relationship diagram between the materials and business objects is as follows:

[0077] Using the set of material nodes as the first-level nodes and the set of business object nodes as the second-level nodes, the material code in the business object node is read, the corresponding material node is found in the set of material nodes and a connection edge is established. The business object type, business collection timestamp and the number of connection edges are written into the connection edge. The number of connection edges for procurement objects is the procurement quantity, the number of connection edges for requisition objects is the requisition quantity, the number of connection edges for replenishment objects is the replenishment quantity, and the number of connection edges for shortage objects is the inventory gap quantity obtained by subtracting the inventory balance from the safety stock quantity. After all connection edges are established, a two-part association graph of material-business object is generated.

[0078] In this embodiment, the screening of similar materials for procurement, similar materials for requisition, similar materials for replenishment, and similar materials for shortage corresponding to the target materials includes:

[0079] The SimRank algorithm is executed to divide the material-business object bipartite relationship graph into a procurement object layer, a requisition object layer, a replenishment object layer, and a shortage object layer. The business paths between the target material node and the remaining material nodes are extracted, and the corresponding inbound / outbound flow phases are identified.

[0080] The material-business object relationship diagram is divided into four layers: procurement object layer, requisition object layer, replenishment object layer, and shortage object layer, specifically:

[0081] Read the business object type of each business object node, and classify the procurement object node and its connected material nodes and connecting edges into the procurement object layer, the requisition object node and its connected material nodes and connecting edges into the requisition object layer, the replenishment object node and its connected material nodes and connecting edges into the replenishment object layer, and the out-of-stock object node and its connected material nodes and connecting edges into the out-of-stock object layer, thus obtaining 4 object layers.

[0082] Extract the business path between the target material node and the remaining material nodes, specifically as follows:

[0083] Starting from the target material node, read the business object nodes connected to the target material node in the four object layers, and then read the remaining material nodes connected to the same business object node. Arrange the target material node, business object nodes and remaining material nodes into a business path according to the connection order, and write the business object type, business object identifier, business collection timestamp and number of connection edges to obtain the business path between the target material node and the remaining material nodes.

[0084] Identify the corresponding inbound / outbound flow phases, specifically:

[0085] Using one collection day as the phase judgment cycle, the system reads historical inbound records, historical outbound records, inventory balance data, safety stock data, replenishment request data, and shortage record data for the target materials and remaining materials. A stable inventory phase is marked when the inventory balance is greater than or equal to 1.2 times the safety stock quantity and the outbound quantity on that day is less than 1.1 times the average outbound quantity of the past 7 days. A depletion phase is marked when the outbound quantity on that day is greater than or equal to 1.1 times the average outbound quantity of the past 7 days and the inventory balance is greater than the safety stock quantity. A critical inventory phase is marked when the inventory balance is less than the safety stock quantity or there is a shortage record. A replenishment request data is present and the inventory balance is less than 1.2 times the safety stock quantity. An inbound recovery phase is marked when historical inbound records exist and the inventory balance changes from being lower than the safety stock quantity to being greater than or equal to the safety stock quantity.

[0086] Based on standardized material inbound and outbound data, semantic tagging is performed on business paths to obtain a heterogeneous path semantic set. Then, prior, lagging, and synchronous relationships are identified according to the business transmission sequence of requisition, stockout, replenishment, and procurement, generating business transmission relationships, where:

[0087] Semantic tagging of business paths, specifically:

[0088] Read the business object type and business collection timestamp in each business path, mark the path corresponding to the procurement object as the procurement path, the path corresponding to the requisition object as the requisition path, the path corresponding to the replenishment object as the replenishment path, the path corresponding to the out-of-stock object as the out-of-stock path, and the requisition path is marked as the production task traction path when it contains a production task identifier. When the same material has an out-of-stock path and a replenishment path in sequence within 14 collection days, mark the corresponding path as the out-of-stock replenishment response path, and summarize to obtain a heterogeneous path semantic set.

[0089] The business transmission relationship is generated as follows:

[0090] Following the business transmission sequence of requisition, stockout, replenishment, and procurement, the business collection timestamps of the target material node and the remaining material node in the heterogeneous path semantic set are read. When the same type of path of the remaining material node is 1 to 14 collection days earlier than the target material node, it is marked as a leading relationship; when it is 1 to 14 collection days later than the target material node, it is marked as a lagging relationship; when the time difference is no more than 1 collection day, it is marked as a synchronous relationship. The business transmission relationship is then generated by summarizing the material code, business object type, and business collection timestamp.

[0091] Based on standardized material inbound and outbound data, causal gating is performed on heterogeneous path semantic sets, inbound / outbound flow phases, and business transmission relationships to obtain a gated path semantic set. SimRank iterative calculations are then performed to obtain the structural similarity of the procurement layer, requisition layer, replenishment layer, and stockout layer. Among these:

[0092] Causal gating is applied to heterogeneous path semantic sets, inbound / outbound flow phases, and business transmission relationships, specifically as follows:

[0093] Read each business path in the heterogeneous path semantic set one by one, obtain the path type, the corresponding material inbound / outbound flow phase, and the business collection timestamp. When the requisition path corresponds to the outbound consumption phase, the stockout path corresponds to the stockout trigger phase, the replenishment path corresponds to the replenishment response phase, and the procurement path corresponds to the inbound recovery phase, and the time interval between business paths is within 1 to 14 collection days, the corresponding business path is retained. When the path type of the business path does not correspond to the inbound / outbound flow phase, or the time interval between business paths is less than 1 collection day or greater than 14 collection days, the corresponding business path is removed. Summarize the retained business paths to obtain the gated path semantic set.

[0094] The SimRank iteration calculation is performed as follows:

[0095] In the procurement object layer, requisition object layer, replenishment object layer, and shortage object layer, the initial structural similarity between the same material node and itself is set to 1, and the initial structural similarity between different material nodes is set to 0. In each iteration, the adjacent business objects of the target material node and the remaining material nodes in the gated path semantic set are read. The sum of the similarity between the adjacent business objects in the previous round is divided by the product of the number of adjacent business objects of the two, and then multiplied by the decay coefficient 0.8 to obtain the structural similarity of the current round. The four object layers are iterated for 10 rounds respectively. The iteration stops when the maximum change value of the structural similarity between two adjacent rounds is less than 0.001. The structural similarity of the procurement layer, requisition layer, replenishment layer, and shortage layer are obtained.

[0096] Based on the closed-loop flow relationship between outbound, stockout, replenishment, and inbound, closed-loop enhancement is performed on the structural similarity of the purchasing layer, requisition layer, replenishment layer, and stockout layer. Furthermore, conflict suppression is achieved by combining alternative outbound relationships, reverse inventory change relationships, and alternating requisition relationships, resulting in a business layer structure similarity matrix, where:

[0097] The closed-loop flow relationship between outbound, stockout, replenishment and inbound refers to the business status correspondence of the same material in consecutive collection days, which is: an increase in outbound quantity, inventory balance below safety stock, replenishment request generated, and an increase in inbound quantity that restores inventory balance to above safety stock.

[0098] Closed-loop enhancement is performed on the structural similarity of the procurement layer, the requisition layer, the replenishment layer, and the stockout layer, specifically as follows:

[0099] Read the historical outbound record data, stockout record data, replenishment application data, historical inbound record data, and business collection timestamp data of the target material node and the remaining material node, and determine whether both have formed a closed-loop flow relationship of outbound, stockout, replenishment, and inbound within 14 collection days. When both have formed a closed-loop flow relationship and the difference between the start time of the closed loop does not exceed 3 collection days, multiply the structural similarity of the requisition layer and the structural similarity of the procurement layer by 1.10, and multiply the structural similarity of the stockout layer and the structural similarity of the replenishment layer by 1.20. When the enhanced structural similarity is greater than 1, correct the structural similarity value to 1 to obtain the 4-layer structural similarity after closed-loop enhancement.

[0100] Conflict suppression is achieved by combining substitution outbound relationships, reverse inventory change relationships, and alternating requisition relationships, specifically as follows:

[0101] Read historical outbound records, inventory balance data, requisition data, production task data, and business collection timestamp data for both the target material node and the remaining material node. When the target material inventory balance is lower than the safety stock level and the outbound quantity of the remaining material increases by more than 30% compared to the average outbound quantity of the previous 7 days, it is marked as a substitution outbound relationship. When the target material inventory balance decreases and the remaining material inventory balance increases, and the opposite direction of their inventory changes occurs for 3 consecutive collection days, it is marked as a reverse inventory change relationship. When the target material and the remaining material alternately appear in the requisition record under the same production task identifier, it is marked as a substitution outbound relationship. When there is only one type of material requisition record in the application data and there is only one type of material requisition record on the same collection day, it is marked as an alternating requisition relationship. For the remaining material nodes that have at least one of the above relationships, the structural similarity of the four layers is multiplied by 0.60 to obtain the structural similarity after conflict suppression. Using the material code as the row identifier and column identifier, a similarity matrix for the procurement layer, a similarity matrix for the requisition layer, a similarity matrix for the replenishment layer, and a similarity matrix for the shortage layer are established. The structural similarity after conflict suppression is written into the intersection position of the target material code and the remaining material code, and the four matrices are merged to obtain the business layer structure similarity matrix.

[0102] Based on the business hierarchical structure similarity matrix, similar materials for procurement, requisition, replenishment, and shortage are selected for the target materials, among which:

[0103] Based on the business hierarchical structure similarity matrix, the target materials are filtered to include similar materials for procurement, requisition, replenishment, and shortage. Specifically:

[0104] Read the structural similarity of the target material node in the business layer structure similarity matrix for the procurement layer, requisition layer, replenishment layer, and shortage layer. Within each object layer, sort the remaining material nodes according to structural similarity from high to low, and retain material nodes with structural similarity greater than or equal to 0.65. When the number of material nodes retained in a single object layer exceeds 10, only the top 10 material nodes are retained. Material nodes retained in the procurement object layer are considered procurement-similar materials, material nodes retained in the requisition object layer are considered requisition-similar materials, material nodes retained in the replenishment object layer are considered replenishment-similar materials, and material nodes retained in the shortage object layer are considered shortage-similar materials.

[0105] In this embodiment, the step of constructing a business-layered similar grid chain set based on the flow sequence of similar materials purchased, similar materials issued, similar materials replenished, and similar materials out of stock includes:

[0106] Based on standardized material inbound and outbound data, the system reads historical inbound records, historical outbound records, inventory balance data, replenishment request data, and stockout records of the target material according to the material code and business collection timestamp. This generates the inbound and outbound flow sequence of the target material, which is then sequentially connected to obtain the target material's own flow chain.

[0107] The target material's own circulation chain is obtained, specifically:

[0108] Read the inbound records, outbound records, inventory balance, replenishment requests, stockout records, and business collection timestamps corresponding to the target material code. Using one collection day as a time node, sum the inbound and outbound quantities within the same collection day. Read the inventory balance for the day and generate replenishment request status and stockout status based on whether there are replenishment requests and stockout records. Merge the material code, collection day, inbound quantity, outbound quantity, inventory balance, replenishment request status, and stockout status into a target material circulation node. Connect adjacent target material circulation nodes from earliest to latest according to the collection day. Write the difference in inbound quantity, outbound quantity, and inventory balance between the current collection day and the previous collection day into the connection relationship to obtain the target material's own circulation chain.

[0109] Based on standardized material inbound and outbound data, historical inbound records, historical outbound records, inventory balance data, replenishment application data, and shortage record data for similar materials are extracted according to material codes and business collection timestamps. Corresponding similar material circulation sequences are generated and aligned according to the time sequence of the target material's own circulation chain. Procurement similar grid chains, requisition similar grid chains, replenishment similar grid chains, and shortage similar grid chains are constructed respectively to obtain a business-layered similar grid chain set.

[0110] In this embodiment, obtaining the implicit lattice chain sequence includes:

[0111] Based on the business hierarchical structure similarity matrix, the structural similarity of each similar material in the business hierarchical similarity grid chain set is matched according to the material code and object layer type to generate grid chain similarity mapping results, where:

[0112] The lattice chain similarity mapping results are generated as follows:

[0113] Read the material code, object layer type, and business collection timestamp of similar material nodes in the business layer similar grid chain set. Enter the similarity matrix of the procurement layer, requisition layer, replenishment layer, or shortage layer according to the object layer type. Read the structural similarity with the target material code as the row identifier and the material code of the similar material node as the column identifier. Organize the similar material node identifier, material code, object layer type, business collection timestamp, and structural similarity into mapping records and summarize to generate grid chain similarity mapping results.

[0114] Based on the business data collection timestamp, the grid chain similarity mapping results are written into the business-layered similar grid chain set, and the structural similarity within the same object layer is normalized to obtain the grid chain propagation weight, where:

[0115] The propagation weights of the lattice chain are obtained as follows:

[0116] Read the grid chain similarity mapping results according to the business collection timestamp, write the structural similarity into the similar material nodes in the business layer similar grid chain set that have the same material code, object layer type and business collection timestamp, and calculate the total structural similarity of the same collection day in the four object layers. When the total is greater than 0, divide the structural similarity of the similar material node by the total structural similarity of the same layer and the same day to obtain the grid chain propagation weight. When the total is equal to 0, set the grid chain propagation weight of the same layer and the same day to 0.

[0117] The target material's own circulation chain, the set of similar grid chains in business layers, and the grid chain propagation weights are serialized and combined according to the business collection timestamp to obtain an implicit structure grid chain sequence, where:

[0118] The implicit lattice chain sequence is obtained as follows:

[0119] Read the target material flow nodes and similar grid chain nodes for procurement, requisition, replenishment, and shortage within the same collection day, as well as their corresponding grid chain propagation weights, in order of business collection timestamp from earliest to latest. Take the target material flow nodes as the main chain nodes, organize the four types of similar grid chain nodes and their corresponding grid chain propagation weights into the grid chain input items of the main chain nodes, and arrange the main chain nodes and grid chain input items in the order of collection days to obtain the implicit structure grid chain sequence.

[0120] In this embodiment, obtaining the inbound / outbound prediction status representation includes:

[0121] An improved Lattice LSTM model is constructed, comprising a lattice chain input unit, a phase transition pulse reconstruction unit, an auction contribution correction unit, and a closed-loop responsibility output unit, wherein;

[0122] The improved Lattice LSTM model is constructed as follows:

[0123] Based on the traditional Lattice LSTM model, improvements are made by adding business-layered lattice parsing and lattice propagation weight reading processing to the original lattice input structure to obtain the lattice input unit. Inventory phase change recognition and pulse-driven lattice reconnection processing are added between the original main chain memory propagation structure and the lattice input structure to obtain the phase change pulse reconstruction unit. Purchase gate, requisition gate, replenishment gate, stockout gate and write priority correction processing are added to the original lattice gate to obtain the auction contribution correction unit. Closed-loop state write-back processing of outbound, stockout, replenishment and inbound is added to the original hidden state output structure to obtain the closed-loop responsibility output unit, resulting in the improved Lattice LSTM model.

[0124] The lattice input unit parses the implicit structure lattice sequence to obtain the target material's own circulation chain, the business-layered similar lattice chain set, and the lattice propagation weight. Based on standardized material inbound and outbound data, it identifies inventory stages, phase transition boundaries, and inventory change pulses. It then segments the target material's own circulation chain and the business-layered similar lattice chain set to obtain the target material's phase transition circulation segment and the business-layered similar lattice chain segment set, where:

[0125] The lattice input unit includes:

[0126] Main chain node register: stores the material code, collection date, quantity entering the warehouse, quantity leaving the warehouse, inventory balance, replenishment application status and stockout status of the target material flow node;

[0127] Grid Chain Node Register: Stores similar grid chain nodes for procurement, similar grid chain nodes for requisition, similar grid chain nodes for replenishment, and similar grid chain nodes for stockouts;

[0128] Weighted index table: stores the correspondence between similar material codes, object layer types, collection dates, and grid chain propagation weights;

[0129] Time-aligned cache: Align main chain nodes, grid chain nodes, and grid chain propagation weights according to the collection date;

[0130] The target material's own circulation chain, the set of similar grid chains at the business level, and the grid chain propagation weights are obtained, specifically:

[0131] The grid input unit reads the main chain nodes and grid input items in the implicit structure grid sequence, sorts and connects the main chain nodes according to the collection date to obtain the target material's own circulation chain, reads the object layer type, similar material code, collection date, inbound quantity, outbound quantity, inventory balance, replenishment application status and shortage status from the grid input items, classifies them according to the procurement object layer, requisition object layer, replenishment object layer and shortage object layer, and connects them according to the similar material code and collection date to obtain the procurement similar grid chain, requisition similar grid chain, replenishment similar grid chain and shortage similar grid chain, summarizes them to obtain the business layer similar grid chain set, reads the grid propagation weights written in the grid input items, establishes a corresponding relationship according to the similar material code, object layer type and collection date, and extracts the grid propagation weights;

[0132] Based on standardized material inbound and outbound data, we identify inventory stages, phase transition boundaries, and inventory change pulses, specifically:

[0133] The system reads the target material's inbound and outbound records, inventory balance, safety stock, replenishment requests, shortage records, and data collection timestamps. Using one data collection day as the judgment period, a stable phase is marked when the inventory balance is greater than or equal to 1.2 times the safety stock and the outbound quantity is less than 1.1 times the average outbound quantity of the past 7 days. A consumption phase is marked when the outbound quantity is greater than or equal to 1.1 times the average outbound quantity of the past 7 days and the inventory balance is higher than the safety stock. A critical phase is marked when the inventory balance is greater than or equal to the safety stock but less than 1.2 times the safety stock. A critical phase is marked when the inventory balance is lower than the safety stock or there is a shortage. When a stock is recorded, it is marked as a stockout stage. When there is a replenishment request and the remaining inventory is less than 1.2 times the safety stock, it is marked as a replenishment stage. When there is an inbound record and the remaining inventory recovers from less than the safety stock to greater than or equal to the safety stock, it is marked as an inbound recovery stage. When the inventory stages of two adjacent collection days are different, the later collection day is determined as the phase transition boundary. When the outbound quantity increases by more than 30% compared to the average outbound quantity of the past 7 days, the remaining inventory decreases by more than 20% compared to the previous collection day, the replenishment request status changes from none to yes, or the inbound quantity increases by more than 30% compared to the previous collection day, the corresponding collection day is marked as an inventory change pulse.

[0134] The target material's own circulation chain and the set of similar grid chains for business layers are segmented into segments, specifically:

[0135] Using the collection date corresponding to the phase transition boundary as the dividing point, the target material's own circulation chain is divided according to the collection date order to obtain stable segments, consumption segments, critical segments, stockout segments, replenishment segments, and warehousing recovery segments, forming target material phase transition circulation segments. Procurement similar grid chains, requisition similar grid chains, replenishment similar grid chains, and stockout similar grid chains are simultaneously divided according to the same dividing point, keeping the collection date range of similar grid chain segments consistent with that of target material phase transition circulation segments, to obtain a set of business-layered similar grid chain segments;

[0136] The phase transition pulse reconstruction unit, based on inventory stages, phase transition boundaries, and inventory change pulses, performs pulse-driven reconnection of the procurement similar grid chain, requisition similar grid chain, replenishment similar grid chain, and stockout similar grid chain. It then propagates the state along the target material phase transition flow segment and the set of business-layered similar grid chain segments to obtain the target material main chain memory state and the reconstruction grid chain propagation state, where:

[0137] The phase-change pulse reconstruction unit includes:

[0138] Inventory Stage Register: Stage markers for stable inventory stage, consumption stage, critical stage, stockout stage, replenishment stage, and inbound recovery stage;

[0139] Phase transition boundary register: stores the next acquisition day corresponding to a change in the inventory stage between adjacent acquisition days;

[0140] Pulse Trigger Register: Inventory Outgoing Pulse, Inventory Decrease Pulse, Replenishment Pulse, and Incoming Pulse, along with their corresponding acquisition dates;

[0141] Pulse-driven reconnection relationship cache: Stores the reconnection relationship between the requisition similar grid chain and the out-of-stock similar grid chain, the out-of-stock similar grid chain and the replenishment similar grid chain, the replenishment similar grid chain and the procurement similar grid chain, and the procurement similar grid chain and the target material's own circulation chain.

[0142] State propagation is performed along the target material phase change flow segment and the set of similar lattice chain segments in the business hierarchy, specifically as follows:

[0143] The phase-change pulse reconstruction unit reads the target material flow nodes in the order of the collection date, uses the main chain memory value of the previous collection date as the historical status field, and uses the inbound quantity, outbound quantity, inventory balance, replenishment application status, shortage status, inventory stage and inventory change pulse of the current collection date as the current status field to generate the main chain memory value of the current collection date. The main chain memory status of the target material is then summarized. At the same time, it reads similar grid chain segments of procurement, requisition, replenishment and shortage, and writes the status of the reconnection starting node to the reconnection ending node according to the connection direction in the cross-grid reconnection structure, according to the grid chain propagation weight. The propagation value of similar grid chain nodes is updated along the order of the collection date to obtain the reconstructed grid chain propagation status.

[0144] The auction contribution correction unit prioritizes the write operations of the purchase gate, requisition gate, replenishment gate, and stockout gate based on the lattice propagation weight, inventory stage, historical predicted contribution, and conflict suppression results. It then controls the corresponding similar lattice chains to perform primary writes, secondary writes, propagation masking, and reverse memory suppression to obtain the corrected lattice memory state, where:

[0145] The auction contribution correction unit includes:

[0146] Procurement gate: Stores the state of similar procurement grid chains, the propagation weight of procurement grid chains, and the output state of the procurement gate;

[0147] Requisition gate: Stores the requisition state of similar grid chains, the propagation weight of the requisition grid chains, and the output state of the requisition gate;

[0148] Replenishment gate: the state of similar grid chains for replenishment, the propagation weight of replenishment grid chains, and the output state of the replenishment gate;

[0149] Out-of-stock gate: similar grid chain status of out-of-stock items, propagation weight of out-of-stock grid chain, and output status of out-of-stock gate;

[0150] Contribution score cache: Stores the percentage of times each similar grid chain reduced the prediction error within the first 10 collection days;

[0151] Conflict suppression cache: Stores conflict suppression results corresponding to substitution outbound relationships, reverse inventory change relationships, and alternating requisition relationships;

[0152] Memory write cache: stores main write state, auxiliary write state, propagation masking results, and reverse suppression memory state;

[0153] The Procurement Gate, Issuance Gate, Replenishment Gate, and Stockout Gate refer to the four gate control structures that correspond to the Procurement Similarity Grid Chain, Issuance Similarity Grid Chain, Replenishment Similarity Grid Chain, and Stockout Similarity Grid Chain, respectively.

[0154] The write priority is sorted for the purchase, requisition, replenishment, and stockout gates, as follows:

[0155] Read the grid propagation weight, inventory stage, historical prediction contribution, and conflict suppression results of four similar grid chains within the same collection day. The historical prediction contribution is the value obtained by dividing the number of times the corresponding similar grid chain reduced the prediction error in the previous 10 collection days by 10. Multiply the grid propagation weight of each gate by 0.6 and the historical prediction contribution by 0.4, and add them together to obtain the basic priority score of the gate. In the consumption stage, the score of the requisition gate is increased by 0.2. In the stockout stage, the score of the stockout gate is increased by 0.2. In the replenishment stage, the score of the replenishment gate is increased by 0.2. In the inbound recovery stage, the score of the procurement gate is increased by 0.2. When there is a conflict suppression result, the corresponding gate score is reduced by 0.3. Arrange the four gates from high to low according to the final score to obtain the writing priority sorting result.

[0156] Controlling the execution of primary write, secondary write, propagation masking, and reverse memory suppression for corresponding similar lattice chains is specifically as follows:

[0157] Read the priority sorting results, write the similar grid chain status corresponding to the first-ranked gate to the grid chain memory position of the current collection day, and write the similar grid chain status corresponding to the second-ranked gate to the auxiliary memory position. When there is a conflict suppression result for the corresponding similar grid chain and the grid chain propagation weight is less than 0.20, the corresponding similar grid chain status is blocked and not written to the grid chain memory position. When the inventory change direction of the similar grid chain and the target material is opposite in the same collection day, reverse memory suppression is performed. The reverse memory suppression coefficient is 0.60. The inbound quantity, outbound quantity, inventory balance and grid chain propagation weight in the corresponding similar grid chain status are multiplied by 0.60 and written to the reverse suppression memory position. The main write status, auxiliary write status, propagation blocking result and reverse suppression memory position are summarized to obtain the corrected grid chain memory status.

[0158] The closed-loop responsibility output unit, based on the closed-loop sequence of outbound, inventory decrease, stockout, replenishment, and inbound recovery, performs closed-loop inversion and write-back of the target material's main chain memory state and the corrected grid chain memory state. It also performs gating fusion of the output states of the purchasing, requisition, replenishment, and stockout gates based on the grid chain propagation weights to obtain the inbound / outbound prediction state representation, where:

[0159] The closed-loop responsibility output unit includes:

[0160] Closed-loop sequence register: the closed-loop sequence of inventory release, inventory decrease, stockout, replenishment, and inbound recovery;

[0161] Inversion write-back register: stores the status of purchase gate, replenishment gate, stockout gate and requisition gate written back to the corresponding main chain memory location;

[0162] Gated fusion cache: Stores the output status of the procurement gate, the requisition gate, the replenishment gate, and the out-of-stock gate, multiplied by their respective grid chain propagation weights, resulting in a weighted output status;

[0163] Predictive status output cache: Inbound trend, Outbound trend, Inventory balance trend, Replenishment demand trend, and Stockout risk trend;

[0164] The closed-loop sequence of outbound shipment, inventory decrease, stockout, replenishment, and inbound recovery refers to the sequence of the same material within a consecutive collection day, in which the outbound quantity increases, the inventory balance decreases, the inventory balance falls below the safety stock level, a replenishment request is generated, and the inbound quantity increases and the inventory balance recovers to above the safety stock level.

[0165] The closed-loop inversion and write-back of the target material's main chain memory state and the corrected lattice chain memory state is performed as follows:

[0166] The closed-loop responsibility output unit reads the target material main chain memory status and the corrected grid chain memory status in reverse order of inbound recovery, replenishment, stockout, inventory decrease and outbound. When the inbound recovery status corresponds to the purchase gate output status, the purchase gate output status is written back to the main chain memory position of the replenishment stage. When the replenishment status corresponds to the replenishment gate output status, the replenishment gate output status is written back to the main chain memory position of the stockout stage. When the stockout status corresponds to the stockout gate output status, the stockout gate output status is written back to the main chain memory position of the inventory decrease stage. When the inventory decrease status corresponds to the requisition gate output status, the requisition gate output status is written back to the main chain memory position of the outbound stage, thus obtaining the main chain memory status after closed-loop write-back.

[0167] The inbound / outbound forecast status is represented as follows:

[0168] Read the main chain memory state after closed-loop write-back, the corrected grid chain memory state, the purchase gate output state, the requisition gate output state, the replenishment gate output state, the stockout gate output state, and the corresponding grid chain propagation weights. Extract the inbound component, outbound component, inventory balance component, replenishment component, and stockout component from the main chain memory state after closed-loop write-back. Multiply the four gate output states by the corresponding grid chain propagation weights to obtain the purchase weighted output state, the requisition weighted output state, the replenishment weighted output state, and the stockout weighted output state.

[0169] The following steps are combined to form an inbound trend: First, the inbound component in the main chain's inbound state, the inbound component in the procurement weighted output state, and the inbound-related component in the replenishment weighted output state. Second, the outbound component in the main chain's outbound state, the outbound component in the requisition weighted output state, and the outbound-related component in the stockout weighted output state are combined to form an outbound trend. Third, the inventory balance component in the main chain's inventory remaining quantity, the inventory replenishment component in the procurement weighted output state, the replenishment replenishment component in the replenishment weighted output state, the outbound consumption component in the requisition weighted output state, and the stockout consumption component in the stockout weighted output state are combined to form an inventory remaining quantity trend. Fourth, the replenishment component in the main chain's replenishment state, the replenishment component in the replenishment weighted output state, and the stockout trigger component in the stockout weighted output state are combined to form a replenishment demand trend. Fifth, the stockout component in the main chain's stockout state, the stockout component in the stockout weighted output state, the outbound consumption component in the requisition weighted output state, and the reverse inhibition memory component in the corrected grid chain memory state are combined to form a stockout risk trend. Finally, these five trends are merged into a state vector based on the same collection date to obtain the inbound / outbound prediction state representation.

[0170] The improved Lattice LSTM model was trained using inbound / outbound prediction error, inventory balance prediction error, replenishment demand prediction error, and inventory risk identification error as joint optimization objectives. The parameters of the lattice input unit, phase transition pulse reconstruction unit, auction contribution correction unit, and closed-loop responsibility output unit were optimized. Training was considered convergent when the decrease in the joint optimization objective was less than 0.001 over 10 consecutive training rounds, or when the training rounds reached 200.

[0171] The improved Lattice LSTM model is trained as follows:

[0172] Training samples are constructed from standardized material inbound and outbound data according to material codes and business collection timestamps. The target material's own circulation chain, business layer similar grid chain set, and grid chain propagation weight for 30 consecutive collection days are used as training inputs. The actual inbound quantity, actual outbound quantity, actual inventory balance, actual replenishment request quantity, and actual inventory risk status on the 31st collection day are used as training labels. The actual inventory risk status is determined based on the actual inventory balance, safety stock quantity, and stockout records on the 31st collection day. The actual inventory balance is set to 1 when it is lower than the safety stock quantity or there is a stockout record, and 0 otherwise.

[0173] The training input is fed into the improved Lattice LSTM model to obtain the inbound / outbound prediction state representation. Then, based on this representation, inbound / outbound prediction processing is performed to obtain the trained predicted inbound quantity, trained predicted outbound quantity, trained predicted inventory balance, trained predicted replenishment demand quantity, and trained inventory risk status. The inbound error is equal to the squared difference between the trained predicted inbound quantity and the actual inbound quantity. The outbound error is equal to the squared difference between the trained predicted outbound quantity and the actual outbound quantity. The total inbound / outbound prediction error is equal to the sum of the inbound and outbound errors divided by 2. The inventory balance prediction error is equal to the training predicted inventory level. The square of the difference between the remaining inventory and the actual remaining inventory is used to calculate the replenishment demand forecast error. The square of the difference between the trained predicted replenishment demand quantity and the actual replenishment request quantity is used to calculate the replenishment demand error. The training inventory risk status is set to 1 when there is risk and 0 when there is no risk. The inventory risk identification error is set to the square of the difference between the training inventory risk status value and the actual inventory risk status value. The joint optimization objective is the sum of the inbound / outbound forecast error multiplied by 0.35, the remaining inventory forecast error multiplied by 0.25, the replenishment demand forecast error multiplied by 0.25, and the inventory risk identification error multiplied by 0.15.

[0174] The parameters of the lattice chain input unit, phase transition pulse reconstruction unit, auction contribution correction unit, and closed-loop responsibility output unit are updated based on the joint optimization objective. When the decrease of the joint optimization objective is less than 0.001 in 10 consecutive training rounds, or when the number of training rounds reaches 200, the training is considered to have converged.

[0175] In this embodiment, the pulse-driven reconnection of the procurement similar grid chain, the requisition similar grid chain, the replenishment similar grid chain, and the out-of-stock similar grid chain includes:

[0176] Based on the target material phase change flow segment, the business-layered similar grid chain segment set, and inventory change pulses, the pulse trigger position is identified, and the corresponding similar grid chain is matched to obtain the pulse grid chain matching result, where:

[0177] Identify the pulse trigger position, specifically:

[0178] Read the pulse type and corresponding collection date from the inventory change pulses. Determine the collection date that contains outbound pulses, inventory decrease pulses, replenishment pulses, or inbound pulses as the pulse trigger collection date. Find the target material segment containing the pulse trigger collection date in the target material phase change flow segment, and establish a correspondence between the pulse type, the pulse trigger collection date, and the target material segment to obtain the pulse trigger position.

[0179] Matching corresponding similar grid chains, specifically:

[0180] Read the business layer similar grid chain fragment set according to pulse type. Outbound pulses match the requisition similar grid chain fragments and out-of-stock similar grid chain fragments within the same collection day. Inventory decrease pulses match the out-of-stock similar grid chain fragments and replenishment similar grid chain fragments within the same collection day. Replenishment pulses match the replenishment similar grid chain fragments and procurement similar grid chain fragments within the same collection day. Inbound pulses match the procurement similar grid chain fragments and target material's own circulation chain fragments within the same collection day.

[0181] The pulse lattice chain matching results are as follows:

[0182] The pulse type, pulse trigger acquisition date, target material fragment, matched similar grid chain fragment, similar material code, and grid chain propagation weight are organized into matching records. All matching records are summarized to obtain the pulse grid chain matching result.

[0183] Based on the pulse grid chain matching results, outbound pulses are connected to similar grid chains for requisition and stockout; inventory decrease pulses are connected to similar grid chains for stockout and replenishment; replenishment pulses are connected to similar grid chains for replenishment and procurement; and inbound pulses are connected to similar grid chains for procurement and the target material's own circulation chain, generating pulse-driven reconnection relationships, where:

[0184] Generate pulse-driven reconnection relationships, specifically as follows:

[0185] Read the pulse grid chain matching results, determine the reconnection direction according to the pulse type, generate the reconnection direction from the requisition similar grid chain segment to the out-of-stock similar grid chain segment, generate the reconnection direction from the out-of-stock similar grid chain segment to the replenishment similar grid chain segment, generate the reconnection direction from the replenishment similar grid chain segment to the procurement similar grid chain segment, generate the reconnection direction from the procurement similar grid chain segment to the target material's own circulation chain segment. Organize the reconnection start point, reconnection end point, pulse type, pulse trigger acquisition date and grid chain propagation weight into reconnection records, summarize all reconnection records, and generate pulse-driven reconnection relationships;

[0186] The grid chain connection is updated based on pulse-driven reconnection, and the out-of-stock similar grid chain status, replenishment similar grid chain status, and procurement similar grid chain status are sequentially written into the replenishment similar grid chain, the procurement similar grid chain, and the target material's own flow chain, resulting in a reconnected business-layered similar grid chain set, where:

[0187] The lattice link connection is updated based on the pulse-driven reconnection relationship, specifically as follows:

[0188] Read each pulse-driven reconnection relationship, locate the node corresponding to the pulse trigger acquisition day in the reconnection start segment and reconnection end segment respectively, retain the original grid chain connection connected according to the acquisition day, and add a cross-grid chain connection edge between the reconnection start node and the reconnection end node. Write the pulse type, pulse trigger acquisition day and grid chain propagation weight into the cross-grid chain connection edge to generate a cross-grid chain reconnection structure.

[0189] The status of similar grid chains for stockouts, replenishment, and procurement refers to the status record in the corresponding similar grid chain node, which consists of material code, collection date, quantity received, quantity shipped, inventory balance, replenishment application status, stockout status, and grid chain propagation weight.

[0190] The reconnected business layer similarity lattice chain set is obtained as follows:

[0191] Read the cross-chain reconnection structure and pulse-driven reconnection relationship. According to the writing direction from the out-of-stock similar chain to the replenishment similar chain, the replenishment similar chain to the procurement similar chain, and the procurement similar chain to the target material's own circulation chain, locate the endpoint node corresponding to the same pulse-triggered collection day. Write the out-of-stock similar chain status to the source status field of the replenishment similar chain endpoint node, write the replenishment similar chain status to the source status field of the procurement similar chain endpoint node, and write the procurement similar chain status to the source status field of the target material's own circulation chain endpoint node. Summarize the procurement similar chain, requisition similar chain, replenishment similar chain, and out-of-stock similar chain after writing the source status field to obtain the reconnected business-layered similar chain set.

[0192] In this embodiment, the step of generating inbound / outbound forecast results, inventory risk warnings, and inbound / outbound scheduling suggestions for target materials based on the inbound / outbound forecast status includes:

[0193] Based on the inbound / outbound forecast status representation, and combining historical inbound record data, historical outbound record data, inventory balance data, and business data collection timestamp data of the target materials, the predicted inbound quantity, predicted outbound quantity, and predicted inventory balance of the target materials within the forecast time range are generated, resulting in the inbound / outbound forecast results, where:

[0194] Generate the predicted inbound quantity, predicted outbound quantity, and predicted inventory balance of the target materials within the predicted time range, specifically:

[0195] The system reads the inbound trend, outbound trend, and inventory balance trend from the inbound / outbound forecast status representation. It also reads the historical inbound record data, historical outbound record data, and inventory balance data for the seven collection days prior to the current collection date. Using the seven collection days after the current collection date as the forecast time range, it calculates the average inbound quantity over the past seven collection days as the inbound baseline quantity. It adds the inbound baseline quantity to the inbound quantity correction value corresponding to the inbound trend to obtain the forecast inbound quantity. If the result is less than 0, it is corrected to 0. The system then calculates the average outbound quantity over the past seven collection days as the outbound baseline quantity. It adds the outbound baseline quantity to the outbound quantity correction value corresponding to the outbound trend to obtain the forecast outbound quantity. If the result is less than 0, it is corrected to 0. Using the current inventory balance as the starting inventory, it adds the forecast inventory balance of the previous forecast collection day to the forecast inbound quantity of the current day, subtracts the forecast outbound quantity of the current day, and adds the inventory correction value corresponding to the inventory balance trend to obtain the forecast inventory balance of the current day. If the result is less than 0, it is corrected to 0. The system calculates the data for the seven forecast collection days each day and summarizes them to obtain the inbound / outbound forecast results.

[0196] Based on inbound / outbound forecasts, safety stock data, replenishment request data, and stockout records, the system identifies target materials' inventory levels as below safety stock, continuously consumed through outbound shipments, and insufficient replenishment response, generating inventory risk warnings.

[0197] Generate inventory risk warnings, specifically:

[0198] Read the predicted inventory balance, predicted outbound quantity, safety stock data, replenishment request data, and stockout record data. The safety stock quantity is the minimum inventory guarantee quantity recorded in the safety stock data. When the predicted inventory balance is lower than the safety stock quantity, an inventory below safety stock status is generated. When the predicted outbound quantity for three consecutive forecast collection days is greater than 1.2 times the average outbound quantity for the past seven collection days, a continuous outbound consumption status is generated. When there is a replenishment request and the predicted inventory balance is still lower than the safety stock quantity within three collection days after the replenishment request is generated, or when stockout records occur more than twice in the past seven collection days, a replenishment response inadequate status is generated. Write each status into the risk record according to the target material code and the forecast collection date to generate an inventory risk warning.

[0199] Based on inbound / outbound forecasts and inventory risk warnings, and combined with purchase order data, requisition data, replenishment request data, and production task data, suggestions are generated for the replenishment quantity, inbound time, outbound preparation, and production requisition scheduling of target materials, resulting in inbound / outbound scheduling suggestions, among which:

[0200] The following inbound / outbound scheduling suggestions were received:

[0201] The system reads inbound / outbound forecasts, inventory risk warnings, purchase order data, requisition data, replenishment request data, and production task data. When inventory is below safety stock, the system subtracts the minimum predicted inventory balance from the safety stock quantity to obtain the inventory gap quantity, and adds the average outbound quantity over the past three data collection days to generate a replenishment quantity suggestion. When there are incomplete purchase orders, the system compares the planned arrival time with the data collection date on which the predicted inventory balance first falls below the safety stock quantity. If the planned arrival time is later than the data collection date, the inbound time suggestion is set to one data collection date prior to the corresponding data collection date; otherwise, the planned arrival time is retained. When the predicted outbound quantity is above the safety stock quantity for three consecutive data collection days... When the daily outbound quantity exceeds 1.2 times the average outbound quantity of the past 7 data collection days, the quantity and time of the requisition application are read. The larger value between the application quantity and the predicted outbound quantity is taken as the stock preparation quantity, and the data collection day before the earliest application time is taken as the stock preparation time. An outbound stock preparation suggestion is generated. When there is a risk of stockout for materials corresponding to production task data, the production task demand quantity, task start time, and predicted inventory balance are read. Production tasks whose predicted inventory balance can cover the demand quantity are ranked first, and those whose predicted inventory balance cannot cover the demand quantity are ranked last. They are then sorted by task start time to generate production requisition scheduling suggestions. The four types of suggestions are summarized to obtain inbound and outbound scheduling suggestions.

[0202] Example 1: To verify the feasibility of this invention in practice, it was applied to a production and warehousing material inbound / outbound prediction scenario. In this scenario, the target material is coded M017, the base unit of measurement is a piece, and the safety stock quantity is 500 pieces, mainly used for production task requisition. Traditional methods use the average outbound quantity of the past 7 data collection days as the future outbound prediction value and generate a replenishment reminder when the inventory balance is lower than the safety stock quantity. This traditional method is simple to implement, but it cannot identify the transmission relationship between the increase in similar material requisition, the lag in replenishment response, and the rapid decline in inventory, and it is easy to underestimate the replenishment demand in continuous outbound scenarios.

[0203] In the simulated data, the system collects basic material information, historical inbound records, historical outbound records, inventory balance, safety stock data, purchase order data, requisition application data, replenishment application data, stockout record data, production task data, and business collection timestamp data for 30 consecutive collection days for material M017. In the original data, the same material has three codes: M017, M-017, and Material 017. The system unifies them all as M017. In the purchase order, 15 boxes of materials are converted to 300 pieces per box (20 pieces per box). On collection day 12, there was a missing inventory balance. The system replenished it to 854 pieces by adding the previous collection day's inventory balance of 980 pieces, the current day's inbound quantity of 0 pieces, and subtracting the current day's outbound quantity of 126 pieces. On collection day 18, a record with a negative outbound quantity was removed. After processing, standardized material inbound and outbound data are generated. Among them, the outbound quantities from the 24th to the 30th of the collection day M017 were 132, 148, 151, 166, 173, 181 and 190 respectively, and the remaining inventory were 1126, 978, 827, 661, 488, 307 and 117 respectively. The average outbound quantity over the past 7 collection days was 163, and the inventory has been declining continuously.

[0204] The system generates procurement objects, requisition objects, replenishment objects, and stockout objects based on standardized data, and constructs a two-part relationship diagram between materials and business objects. On collection day 25, procurement object P042 contains 300 units of M017 and 200 units of M021. On collection day 27, requisition object L118 contains 166 units of M017 and 94 units of M019. On collection day 29, replenishment object R036 contains 700 units of M017 replenishment requests. On collection day 30, stockout object D014 is formed, with a shortage of 383 units (500 units of safety stock minus 117 units of remaining stock). The system calculates structural similarity at the procurement object, requisition object, replenishment object, and stockout object levels, obtaining a structural similarity of 0.84 for the requisition layer of M019, 0.76 for the procurement layer of M021, and 0.71 for the replenishment layer. Although another material, M033, alternates with M017 in terms of outbound shipments, when the inventory of M017 decreases, the inventory of M033 increases, and this state occurs for three consecutive data collection days. The system marks M033 as having a reverse inventory change relationship and suppresses the structural similarity from 0.68 to 0.41. Therefore, it is not included in the prediction as a similar material.

[0205] Subsequently, the system constructs its own circulation chain for M017, with each collection day as a node. The node records for collection day 28 show: 0 units received, 173 units shipped, 488 units remaining in inventory, no replenishment requests, and stockout status. The node records for collection day 29 show: 0 units received, 181 units shipped, 307 units remaining in inventory, and stockout status. The node records for collection day 30 show: 0 units received, 190 units shipped, 117 units remaining in inventory, and stockout status. The system records the change in inventory balance from collection day 28 to collection day 29 as -181 units, and the change in inventory balance from collection day 29 to collection day 30 as -190 units. Meanwhile, M019 had outbound quantities of 172, 178, and 186 items from collection date 28 to collection date 30, and was constructed as a similar grid chain for requisition. M021 had purchase and replenishment records from collection date 26 to collection date 30, and was constructed as a similar grid chain for purchase and replenishment. The propagation weight of the M019 requisition grid chain was 0.42, the propagation weight of the M021 replenishment grid chain was 0.33, and the propagation weight of the M021 purchase grid chain was 0.25.

[0206] The improved Lattice LSTM model receives the target material's own circulation chain, the business-layered similar grid chain set, and the grid chain propagation weights. It identifies collection day 28 as the stockout phase, collection day 29 as the replenishment phase, and collection day 30 as the stockout continuation phase. Since the number of items shipped on collection day 30 (190 pieces) is 1.1 times higher than the average number of items shipped over the past 7 collection days (163 pieces), and the remaining inventory is lower than the safety stock level, the system identifies the outbound pulse and the inventory decrease pulse, and establishes pulse-driven reconnection relationships from the requisition similar grid chain to the stockout similar grid chain, and from the stockout similar grid chain to the replenishment similar grid chain. The auction contribution correction unit reads the grid chain propagation weights and historical prediction contributions. The M019 requisition similar grid chain reduced the prediction error 7 times in the previous 10 collection days, with a historical prediction contribution of 0.70. The M021 replenishment similar grid chain reduced the prediction error 6 times, with a historical prediction contribution of 0.60. During the stockout phase, the stockout gate priority score increases by 0.2, and the replenishment gate participates in auxiliary writing based on the replenishment response relationship. The M033 corresponding grid chain is masked due to conflict suppression results. The closed-loop responsibility output unit writes back the status in the reverse order of inbound recovery, replenishment, stockout, inventory decrease, and outbound to obtain the inbound / outbound prediction status representation.

[0207] In the training samples, the system uses collection days 1 to 30 as input windows and collection day 31 as the label. In sample A, the total number of items received in window M017 is 900, the total number of items shipped is 2282, the inventory balance on collection day 30 is 117, and the labels for collection day 31 are: actual receipts 0, actual shipments 202, actual inventory balance 0, actual replenishment requests 700, and actual inventory risk status is risky. The model predicts a shipment quantity of 196 on collection day 31 with an error of 6, and a predicted inventory balance of 0, consistent with the actual figures. Sample B uses collection days 2 to 31 as input, with labels for collection day 32: actual receipts 600, actual shipments 184, and actual inventory balance 416. The model predicts a receipt quantity of 580 with an error of 20, a shipment quantity of 187 with an error of 3, and a predicted inventory balance of 393 with an error of 23. After 10 consecutive rounds of training, the joint optimization objective decreased from 0.087 to 0.0864, ​​with the decrease being less than 0.001, indicating that the model had converged.

[0208] In the prediction comparison, the current collection date is set as collection date 30, and predictions are made for collection dates 31 to 37. Traditional methods consistently predict an outbound quantity of 163 units, while the method of this invention predicts 196, 187, 172, 168, 160, 154, and 151 units respectively; the actual outbound quantities are 202, 184, 176, 165, 158, 150, and 149 units respectively. The average absolute error of the traditional method over the seven collection days is 21.86 units, while that of the method of this invention is 3.86 units. Regarding inventory balance prediction, the traditional method predicts an inventory balance of 72 units on collection date 31, but the actual inventory is 0 units, while the method of this invention predicts 0 units, consistent with the actual result. Regarding replenishment recommendations, the traditional method, based on a safety stock of 500 units minus the remaining inventory of 117 units on collection day 30, provides a replenishment recommendation of 383 units. The method of this invention reads the lowest predicted remaining inventory value of 0 units and overlays it with the average outbound quantity of 181 units over the past three collection days, providing a replenishment recommendation of 681 units, which more closely approximates the actual consumption demand from collection day 31 to collection day 33. This simulation case demonstrates that this invention can identify inventory risks in advance when continuous outbound shipments, inventory declines, and replenishment lags occur simultaneously, improving the accuracy of outbound quantity prediction, remaining inventory prediction, and replenishment recommendations.

[0209] 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 prediction of material inbound and outbound operations based on big data analysis, characterized in that, include: Collect data on material inbound and outbound operations, and preprocess the data to obtain standardized material inbound and outbound data. Based on standardized material inbound and outbound data, materials are grouped according to the business relationships of procurement, requisition, replenishment and shortage, generating a set of material business objects, and constructing a two-part relationship diagram of material-business objects; Based on the material-business object bipartite association graph, the SimRank algorithm is executed to calculate the structural similarity between the target material and the remaining material nodes at the procurement object layer, the requisition object layer, the replenishment object layer, and the shortage object layer, respectively, to obtain the business layer structure similarity matrix, and to filter the procurement similar materials, requisition similar materials, replenishment similar materials, and shortage similar materials corresponding to the target material. Construct the target material's own circulation chain based on the inbound and outbound circulation sequence of the target material, and construct a business-layered similar grid chain set based on the circulation sequence of similar materials purchased, similar materials issued, similar materials replenished, and similar materials out of stock; Write the corresponding structural similarity in the business layer structure similarity matrix into the business layer similarity grid chain set, and use it as the grid chain propagation weight to obtain the implicit structure grid chain sequence. An improved Lattice LSTM model is constructed based on implicit structure lattice chain sequences. State propagation is performed along the target material's own flow chain and the set of similar lattice chains in the business hierarchy. Gating fusion is then performed based on the lattice chain propagation weights to obtain the inbound and outbound prediction state representation. Based on the inbound and outbound forecast status, the system generates inbound and outbound forecast results for target materials, inventory risk warnings, and inbound and outbound scheduling suggestions.

2. The intelligent prediction method for material inbound and outbound operations based on big data analysis according to claim 1, characterized in that, The material inbound and outbound business data specifically includes basic material information data, historical inbound record data, historical outbound record data, inventory balance data, safety stock data, purchase order data, requisition application data, replenishment application data, stockout record data, production task data, and business collection timestamp data.

3. The intelligent prediction method for material inbound and outbound operations based on big data analysis according to claim 1, characterized in that, The preprocessing of material inbound and outbound business data specifically includes material coding standardization, timestamp alignment, quantity unit conversion, abnormal record removal, and missing field completion.

4. The intelligent prediction method for material inbound and outbound operations based on big data analysis according to claim 1, characterized in that, The material-business object two-part relationship diagram includes: Based on the purchase order number, requisition application number, production task identifier, replenishment application number, business collection timestamp, inventory balance and safety stock quantity, standardized material inbound and outbound data are aggregated to generate a set of purchase objects, a set of requisition objects, a set of replenishment objects and a set of out-of-stock objects. The sets of procurement objects, requisition objects, replenishment objects, and shortage objects are merged into a set of material business objects. Material node sets and business object node sets are constructed based on material codes and the set of material business objects, respectively. Based on the set of material nodes, the set of business object nodes, and the relationship between them, a connection edge is established to obtain the material-business object bipartite association graph.

5. The intelligent prediction method for material inbound and outbound operations based on big data analysis according to claim 1, characterized in that, The similar materials for procurement, requisition, replenishment, and shortage corresponding to the target materials for screening include: The SimRank algorithm is executed to divide the material-business object bipartite association graph into the procurement object layer, the requisition object layer, the replenishment object layer, and the shortage object layer. The business path between the target material node and the remaining material node is extracted, and the corresponding inbound and outbound flow phases are identified. Based on standardized material inbound and outbound data, semantic tags are applied to business paths to obtain heterogeneous path semantic sets. Priority, lag, and synchronization relationships are identified according to the business transmission order of requisition, stockout, replenishment, and procurement to generate business transmission relationships. Based on standardized material inbound and outbound data, causal gating is performed on heterogeneous path semantic sets, inbound and outbound flow phases, and business transmission relationships to obtain gating path semantic sets. SimRank iterative calculation is then performed to obtain the structural similarity of the procurement layer, the requisition layer, the replenishment layer, and the shortage layer. Based on the closed-loop flow relationship between outbound, stockout, replenishment and inbound, the similarity of the procurement layer structure, the requisition layer structure, the replenishment layer structure, and the stockout layer structure are enhanced in a closed loop. Conflict suppression is achieved by combining alternative outbound relationships, reverse inventory change relationships, and alternating requisition relationships, resulting in a business layer structure similarity matrix. Based on the business hierarchical structure similarity matrix, the target materials are filtered to identify similar materials for procurement, requisition, replenishment, and shortage.

6. The intelligent prediction method for material inbound and outbound operations based on big data analysis according to claim 1, characterized in that, The process of constructing a hierarchical similarity grid chain set based on the flow sequence of similar materials procured, similar materials issued, similar materials replenished, and similar materials out of stock includes: Based on standardized material inbound and outbound data, the system reads historical inbound records, historical outbound records, inventory balance data, replenishment request data, and stockout records of the target material according to the material code and business collection timestamp, generates the inbound and outbound flow sequence of the target material, and connects them in time sequence to obtain the flow chain of the target material itself. Based on standardized material inbound and outbound data, historical inbound records, historical outbound records, inventory balance data, replenishment application data, and shortage record data for similar materials are extracted according to material codes and business collection timestamps. Corresponding similar material circulation sequences are generated and aligned according to the time sequence of the target material's own circulation chain. Procurement similar grid chains, requisition similar grid chains, replenishment similar grid chains, and shortage similar grid chains are constructed respectively to obtain a business-layered similar grid chain set.

7. The intelligent prediction method for material inbound and outbound operations based on big data analysis according to claim 1, characterized in that, The obtained implicit lattice chain sequence includes: Based on the business hierarchical structure similarity matrix, the structural similarity of each similar material in the business hierarchical similar grid chain set is matched according to the material code and object layer type to generate grid chain similarity mapping results; Based on the business collection timestamp, the grid chain similarity mapping results are written into the business layer similar grid chain set, and the structural similarity within the same object layer is normalized to obtain the grid chain propagation weight. The target material's own circulation chain, the set of similar grid chains in business layering, and the grid chain propagation weights are serialized and combined according to the business collection timestamp to obtain an implicit structure grid chain sequence.

8. The intelligent prediction method for material inbound and outbound operations based on big data analysis according to claim 1, characterized in that, The obtained inbound / outbound prediction status representation includes: An improved Lattice LSTM model is constructed, which includes a lattice chain input unit, a phase transition pulse reconstruction unit, an auction contribution correction unit, and a closed-loop responsibility output unit. The grid chain input unit parses the implicit structure grid chain sequence to obtain the target material's own circulation chain, the business layer similar grid chain set, and the grid chain propagation weight. Based on standardized material inbound and outbound data, it identifies the inventory stage, phase transition boundary, and inventory change pulse. It then segments the target material's own circulation chain and the business layer similar grid chain set to obtain the target material phase transition circulation segment and the business layer similar grid chain segment set. The phase change pulse reconstruction unit reconnects the similar grid chains for procurement, requisition, replenishment, and shortage based on inventory stages, phase change boundaries, and inventory change pulses. It also propagates the state along the phase change flow segments of the target material and the set of similar grid chain segments in the business layer to obtain the main chain memory state of the target material and the reconstructed grid chain propagation state. The auction contribution correction unit sorts the write priority of the purchase gate, requisition gate, replenishment gate and stockout gate based on the grid chain propagation weight, inventory stage, historical prediction contribution and conflict suppression results, and controls the corresponding similar grid chains to perform main write, auxiliary write, propagation shielding and reverse memory suppression to obtain the corrected grid chain memory state. The closed-loop responsibility output unit performs closed-loop inversion and write-back on the main chain memory state of the target material and the corrected grid chain memory state based on the closed-loop sequence of outbound, inventory decrease, stockout, replenishment and inbound recovery. It also performs gating fusion on the output states of the procurement gate, requisition gate, replenishment gate and stockout gate based on the grid chain propagation weight to obtain the inbound and outbound prediction state representation. The improved Lattice LSTM model was trained using inbound / outbound prediction error, inventory balance prediction error, replenishment demand prediction error, and inventory risk identification error as joint optimization objectives. The parameters of the lattice input unit, phase transition pulse reconstruction unit, auction contribution correction unit, and closed-loop responsibility output unit were optimized. The training was considered to have converged when the decrease of the joint optimization objective was less than 0.001 in 10 consecutive training rounds, or when the number of training rounds reached 200.

9. The intelligent prediction method for material inbound and outbound operations based on big data analysis according to claim 8, characterized in that, The pulse-driven reconnection of the procurement similar grid chain, the requisition similar grid chain, the replenishment similar grid chain, and the out-of-stock similar grid chain includes: Based on the phase change flow segments of target materials, the set of similar grid chain segments in business layering, and inventory change pulses, the pulse triggering position is identified and the corresponding similar grid chain is matched to obtain the pulse grid chain matching result; Based on the pulse grid chain matching results, the outbound pulse is connected to the requisition similar grid chain and the shortage similar grid chain, the inventory decrease pulse is connected to the shortage similar grid chain and the replenishment similar grid chain, the replenishment pulse is connected to the replenishment similar grid chain and the procurement similar grid chain, and the inbound pulse is connected to the procurement similar grid chain and the target material's own circulation chain, thus generating a pulse-driven reconnection relationship. The grid chain connection is updated based on the pulse-driven reconnection relationship. The out-of-stock similar grid chain status, replenishment similar grid chain status, and procurement similar grid chain status are sequentially written into the replenishment similar grid chain, the procurement similar grid chain, and the target material's own circulation chain to obtain the reconnected business-layered similar grid chain set.

10. The intelligent prediction method for material inbound and outbound operations based on big data analysis according to claim 1, characterized in that, The method for generating inbound / outbound forecast results, inventory risk warnings, and inbound / outbound scheduling suggestions for target materials based on inbound / outbound forecast status includes: Based on the inbound and outbound forecast status representation, combined with the historical inbound record data, historical outbound record data, inventory balance data and business collection timestamp data of the target materials, the forecasted inbound quantity, forecasted outbound quantity and forecasted inventory balance of the target materials within the forecast time range are generated, and the inbound and outbound forecast results are obtained. Based on inbound / outbound forecasts, safety stock data, replenishment request data, and stockout records, the system identifies target materials as being in a state where their inventory is below safety stock, in a state of continuous outbound consumption, and in a state of insufficient replenishment response, and generates inventory risk warnings. Based on the inbound / outbound forecast results and inventory risk warnings, combined with purchase order data, requisition application data, replenishment application data and production task data, suggestions are generated for the replenishment quantity of target materials, inbound time, outbound preparation and production requisition scheduling, resulting in inbound / outbound scheduling suggestions.

Images