An inventory management system based on order data

Abstract

The present invention discloses an order data-based inventory management system, which relates to the field of intelligent warehousing technology. The system includes a priority assessment module that calculates transfer weights and inventory transfer priority scores between warehouses based on an inventory gap list and generates a QUBO matrix. An inventory transfer module uses a quantum annealing algorithm to search for a global optimal solution based on the QUBO matrix, decodes the solution using encoding rules, generates an initial inventory transfer instruction set, and simultaneously optimizes the system using MILP to generate and execute the optimal inventory transfer instructions. By using a quantum annealing algorithm to search for a global optimal solution and combining it with MILP for secondary optimization, the present invention ensures that the generated transfer instructions minimize overall costs while meeting complex business constraints.

Description

Technical Field

[0001] The present invention relates to the field of intelligent warehousing technology, and in particular to an inventory management system based on order data. Background Art

[0002] In the field of intelligent warehousing and supply chain management, conventional technical solutions usually build demand forecasting models based on the multi-dimensional collection and feature engineering of order data. Existing methods generally adopt standardized data preprocessing processes, extract order trend characteristics through time series analysis, combine geographic information system (GIS) coding to achieve spatial modeling of logistics networks, and build category association maps based on the commodity classification system. In terms of data collaboration, traditional systems often adopt a centralized feature pool architecture, realize cross-enterprise data interaction through encrypted transmission and standardized interfaces, and use weighted averaging or principal component analysis (PCA) and other methods to complete feature alignment. The inventory optimization stage usually adopts a strategy that combines heuristic rules with linear programming, generates allocation plans based on the transportation cost matrix and inventory level constraints, and introduces an in-transit inventory correction mechanism to improve the real-time decision-making.

[0003] Existing technologies face the dual challenges of feature semantic gaps and computational complexity in multi-party collaborative optimization scenarios. Traditional feature fusion methods are limited by the differences in the distribution of heterogeneous data, making it difficult to achieve fine-grained feature alignment while protecting data privacy. This results in limited accuracy in extracting spatiotemporal correlation patterns from cross-enterprise order data. At the inventory transfer decision-making level, classical optimization algorithms require multiple relaxation approximations when dealing with nonlinear constraints and are unable to effectively coordinate the spatiotemporal coupling relationships in large-scale logistics networks, resulting in an optimization paradox between transportation resource utilization and inventory turnover. Summary of the Invention

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

[0005] Therefore, the present invention provides an inventory management system based on order data to solve the problem of feature alignment and inventory allocation in multi-party collaborative optimization in the prior art.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, the present invention provides an inventory management system based on order data, which includes a data acquisition module, which collects the original order data of each participant and performs preprocessing, and extracts the three-dimensional features of each order. The original order data includes the order timestamp, product SKU code, order quantity and delivery address; a feature fusion module, which jointly encrypts the three-dimensional features of different orders through a federated learning algorithm to generate a homomorphic encryption data packet, and uses a federated gradient descent algorithm for collaborative training to generate a cross-domain feature similarity matrix, and finally generates a global feature vector through gated attention feature fusion; a stock-out prediction module, which is based on a dynamic warehouse relationship diagram, a time series GRU layer, and an empty space. The inter-warehouse graph convolution layer and the cross-warehouse attention layer are used to build a distributed time series prediction model. Based on the global feature vector, the model predicts the demand for goods in each warehouse in the future. Combined with real-time inventory data, the expected shortage of goods in each warehouse is calculated to form an inventory gap list. The priority evaluation module calculates the transfer weights and inventory transfer priority scores between warehouses based on the inventory gap list and generates a QUBO matrix. The inventory transfer module uses the quantum annealing algorithm to search for the global optimal solution based on the QUBO matrix, decodes it through encoding rules, generates an initial inventory transfer instruction set, and uses MILP for optimization to generate and execute the optimal inventory transfer instructions.

[0008] As a preferred solution of the order data-based inventory management system described in the present invention, the three-dimensional features of each order include time dimension features, space dimension features and category dimension features.

[0009] As a preferred solution of the inventory management system based on order data of the present invention, the steps of generating the global feature vector are as follows:

[0010] The three-dimensional features of each participant's order are jointly encrypted using the Paillier homomorphic encryption algorithm to generate a homomorphic encrypted data packet.

[0011] The secure inner product protocol is used to calculate the multi-dimensional feature similarity of homomorphically encrypted data packets and generate a cross-domain feature similarity matrix.

[0012] The cross-domain feature similarity matrix is used as the weight constraint of the federated gradient descent algorithm to collaboratively train the 3D features of each participant's order and generate a shared projection matrix with cross-enterprise feature alignment.

[0013] Using the shared projection matrix as the linear transformation operator, the three-dimensional features of the orders of each participant are mapped to a unified dimension, and the three-dimensional features of the orders of each participant are fused through gated attention feature fusion to generate a global feature vector.

[0014] As a preferred solution of the inventory management system based on order data of the present invention, the steps of constructing a distributed time series prediction model and predicting the demand for goods in each warehouse in the future period of time based on the global feature vector are as follows:

[0015] Taking each warehouse as a node and the geographical distance between warehouses as an edge, we use the inverse distance weighting method to construct a weighted adjacency matrix, and use ST-GCN to aggregate spatiotemporal features to generate a dynamic warehouse relationship graph.

[0016] Build a distributed time series prediction model based on a dynamic warehouse relationship graph, a time series GRU layer, a spatial graph convolution layer, and a cross-warehouse attention layer;

[0017] The global feature vector is input into the distributed time series forecasting model to predict the demand for goods in each warehouse in the future.

[0018] As a preferred solution of the inventory management system based on order data of the present invention, wherein: the expected out-of-stock quantity of goods in each warehouse is calculated by combining real-time inventory data to form an inventory gap list, the steps are as follows:

[0019] Based on the commodity demand, real-time inventory data, and in-transit inventory of each warehouse in the future, the expected out-of-stock quantity of each warehouse's commodities is calculated using the threshold zeroing method;

[0020] Based on the commodity demand of each warehouse and the expected shortage of commodities in each warehouse in the future period, an inventory gap list containing warehouse number, commodity code, shortage date and expected shortage of commodities in each warehouse is generated.

[0021] As a preferred solution of the inventory management system based on order data of the present invention, the steps of calculating the transfer weights and inventory transfer priority scores between warehouses and generating the QUBO matrix are as follows:

[0022] Based on the warehouse location and expected shortage quantity of each warehouse in the inventory gap list, the transfer weights between warehouses are calculated using the inverse distance weighting method with time and space constraints.

[0023] Based on the expected stockout quantity and historical turnover rate of each warehouse in the inventory gap list, the inventory allocation priority score is calculated through a dynamic index coupling algorithm and a turnover rate deviation amplification mechanism;

[0024] Based on the transfer weights and inventory transfer priority scores between warehouses, the QUBO matrix is integrated and generated through the quadratic unconstrained binary optimization modeling method.

[0025] As a preferred solution of the order data-based inventory management system described in the present invention, the use of a quantum annealing algorithm to search for a global optimal solution refers to initializing quantum annealing parameters and performing an energy minimization search through quantum tunneling effects and thermal fluctuations to generate a binary solution set after quantum annealing.

[0026] As a preferred solution of the inventory management system based on order data of the present invention, the steps of generating and executing the optimal inventory transfer instruction are as follows:

[0027] Define coding rules based on warehouse quantity, commodity type, and transfer quantity range in historical transfer data;

[0028] Decode the binary solution set into transfer decision variables according to the encoding rules, and extract the corresponding warehouse number, product and transfer quantity through hash search method, and finally integrate them to generate the initial inventory transfer instruction set;

[0029] MILP is used to optimize the initial inventory transfer instruction set, generate the optimal inventory transfer instruction and execute it.

[0030] As a preferred solution of the order data-based inventory management system described in the present invention, the original order data includes an order timestamp, a product SKU code, and an order quantity.

[0031] As a preferred solution of the order data-based inventory management system described in the present invention, the preprocessing includes data cleaning, dimension alignment and normalization processing of the order data.

[0032] The beneficial effects of the present invention are as follows: the three-dimensional characteristics of time, space and category of orders are extracted respectively through the STL algorithm, Geohash coding and TF-IDF weighted coding, achieving a comprehensive characterization of the multi-dimensional characteristics of order data, enabling subsequent prediction models to capture richer spatiotemporal and product association patterns; the quantum annealing algorithm is used to optimize the QUBO matrix, and the quantum tunneling effect and thermal fluctuation mechanism are used to perform a global optimal search in an ultra-large-scale solution space. At the same time, MILP is combined with secondary optimization to ensure that the generated allocation instructions achieve overall cost minimization while meeting complex business constraints. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1Schematic diagram of an inventory management system based on order data.

[0035] Figure 2 Flowchart for generating global feature vectors.

[0036] Figure 3 Flowchart for generating inventory gap list.

[0037] Figure 4 A flowchart for generating and executing optimal stock transfer instructions. DETAILED DESCRIPTION

[0038] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0039] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0040] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0041] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides an inventory management system based on order data, including the following steps:

[0042] The data acquisition module collects the original order data of each participant and pre-processes it to extract the three-dimensional features of each order;

[0043] The original order data includes the order timestamp, product SKU code, order quantity, and delivery address.

[0044] Furthermore, raw order data is acquired in real time through the online trading platform. The order timestamp records the precise time of the transaction. The product SKU code is obtained by matching the product master database. The order quantity is directly derived from the shopping cart settlement information, and the delivery address is extracted from the logistics information entered by the user. The order timestamp is stored in the ISO 8601 standard format, the product SKU code is strictly consistent with the ERP inventory database, the order quantity is stored in the transaction log after tamper-proof verification, and the delivery address is converted to latitude and longitude coordinates through geocoding and verified by matching with the administrative division database. During the data collection process, the order timestamp, product SKU code, order quantity, and delivery address are all transmitted through an encrypted channel, formatted and standardized in memory, and written to distributed file storage.

[0045] Preprocessing includes data cleaning, dimension alignment, and normalization of order data.

[0046] Furthermore, order data preprocessing begins with data cleansing. Order timestamps are checked for time format validity and time zone offset correction. Product SKU codes are verified for encoding rules and illegal characters are removed. Negative and abnormally large values are filtered in the order quantity field. Shipping addresses are standardized and missing province / city information is completed. During the dimension alignment phase, order timestamps are uniformly converted to UTC timestamps, product SKU codes are mapped to a standard classification system, order quantities are converted to integer units of measurement, and shipping addresses are broken down into three-level administrative codes: province, city, and district. During the normalization phase, order timestamps are converted to relative timestamps, product SKU codes are one-hot encoded, order quantities are scaled to the range 0-1 using Min-Max scaling, and shipping address latitude and longitude coordinates are converted to planar coordinates using Gaussian projection. Data cleansing combines regular expression matching with business rule validation. Dimension alignment is performed based on a predefined encoding comparison table. Normalization uses linear algebra to ensure data distribution consistency.

[0047] Use STL algorithm to extract the time dimension features of order data;

[0048] Furthermore, when the STL algorithm extracts the time dimension features of order data, it first aggregates the order timestamps into a time series at hourly granularity, then splits the series into trend terms, seasonal terms, and residual terms through seasonal decomposition, and finally selects the slope of the trend term as the time dimension feature.

[0049] Use Geohash encoding to extract the spatial dimension features of order data;

[0050] Furthermore, when Geohash encoding extracts the spatial dimension features of order data, the latitude and longitude coordinates of the delivery address are first converted into a Geohash string, and the encoding length is set to 7 bits to balance accuracy and computational efficiency. Then, the Geohash prefixes of adjacent orders are clustered, and finally the Geohash value of the cluster center point is used as the spatial dimension feature.

[0051] TF-IDF weighted coding is used to extract the category dimension features of order data.

[0052] Furthermore, when TF-IDF weighted coding is used to extract the category dimension features of order data, the frequency of occurrence of all product SKU codes in the order is first counted, the TF-IDF value of each SKU is calculated, and then the top 20 SKU codes with the highest TF-IDF values are selected as keywords. Finally, the TF-IDF values of the keywords are combined into category dimension features through sparse matrix vectorization method.

[0053] It should be noted that the above data collection process is subject to user consent and is used for legitimate purposes.

[0054] The feature fusion module jointly encrypts the three-dimensional features of different orders through a federated learning algorithm to generate a homomorphically encrypted data packet. It then uses a federated gradient descent algorithm for collaborative training to generate a cross-domain feature similarity matrix. Finally, it generates a global feature vector through gated attention feature fusion.

[0055] The three-dimensional features of each participant's order are jointly encrypted using the Paillier homomorphic encryption algorithm to generate a homomorphic encrypted data packet.

[0056] Furthermore, when the Paillier homomorphic encryption algorithm processes the three-dimensional features of each participant's order, it first generates a separate public-private key pair for each participant. The public key is used for encryption operations, while the private key is kept by the data owner. The time dimension features, spatial dimension features, and category dimension features are each converted to fixed-point representations. Then, the numerical value of each feature is encrypted element by element using the Paillier public key. During the encryption process, the time dimension features use the integer portion of the millisecond timestamp as input, the spatial dimension features convert the Geohash string into an ASCII code value, and the category dimension features use the scaled integer value of the TF-IDF value. All encrypted three-dimensional features are grouped and packaged by feature category, forming a homomorphically encrypted data package with complete structural information. The data package retains the original feature's dimensional information and encryption order mark to ensure the dimensional position in subsequent calculations.

[0057] The secure inner product protocol is used to calculate the multi-dimensional feature similarity of homomorphically encrypted data packets and generate a cross-domain feature similarity matrix.

[0058] Furthermore, when processing homomorphically encrypted data packets, the secure inner product protocol first parses the ASN.1 DER formatted homomorphically encrypted data packet to obtain encrypted feature vectors. The time, space, and category dimension features remain encrypted during the computation. During the secure inner product protocol, the encrypted feature vectors of the two participants are point-multiplied in the ciphertext space, leveraging the additive homomorphic properties of Paillier homomorphic encryption to perform the encrypted vector inner product calculation. The computation is grouped by feature dimension. Time dimension features use sliding time window matching to calculate temporal correlation, spatial dimension features use Geohash prefix matching to assess spatial similarity, and category dimension features use encrypted TF-IDF vector cosine similarity to measure category association. The similarity scores for each dimension are weighted and combined to generate the final similarity value. The similarity values for all order pairs from the participating parties are arranged in rows and columns to form a cross-domain feature similarity matrix. Matrix elements are stored in IEEE 754 double-precision floating-point format, with row and column indices strictly corresponding to the original order IDs.

[0059] The cross-domain feature similarity matrix is used as the weight constraint of the federated gradient descent algorithm to collaboratively train the 3D features of each participant's order and generate a shared projection matrix with cross-enterprise feature alignment.

[0060] Furthermore, the cross-domain feature similarity matrix serves as input to the federated gradient descent algorithm. The matrix is first symmetrically normalized so that the sums of rows and columns are all 1, and then decomposed into a feature vector space as the initial projection basis. During the iterative process of the federated gradient descent algorithm, each participant calculates gradients based on the three-dimensional features of their local orders. The time dimension feature gradients are obtained through time series residual backpropagation, the spatial dimension feature gradients are adjusted using the Geohash distance decay function, and the category dimension feature gradients are calculated using TF-IDF weighted cross entropy. During the gradient aggregation phase, the cross-domain feature similarity matrix serves as a weight constraint matrix to perform a similarity-weighted average of the gradients from different participants to ensure feature space alignment. The parameters of the shared projection matrix are updated after each iteration. The rows of the projection matrix correspond to the feature dimensions, and the columns correspond to the latent space dimensions. Matrix elements are stored as double-precision floating-point numbers. The shared projection matrix is generated at the end of training. The shared projection matrix maps the heterogeneous features of different participants into a unified latent semantic space, achieving alignment and comparability of features across enterprise data.

[0061] Using the shared projection matrix as the linear transformation operator, the three-dimensional features of the orders of each participant are mapped to a unified dimension, and the three-dimensional features of the orders of each participant are fused through gated attention feature fusion to generate a global feature vector, which is expressed as:

[0062] ;

[0063] in, It is The global feature vector of orders, is the index of the three-dimensional features of the order (1 = time dimension feature, 2 = space dimension feature, 3 = category dimension feature), This is the order The gating weight matrix of the dimension features, It is Order No. dimensional features, is the shared projection matrix, is the Sigmoid activation function, is the Hadamard product operator.

[0064] Furthermore, the shared projection matrix As a linear transformation operator, first 3D features of orders Perform matrix multiplication to obtain basic feature representation At the same time, the characteristics of each dimension Respectively with the corresponding gating weight matrix After multiplication, the Sigmoid activation function is used Generates a gating coefficient between 0 and 1 During the calculation process, the time dimension feature and The product retains the temporal pattern and spatial dimension features The transformation results maintain geographical relevance and category dimension characteristics Finally, gated fusion is performed on the three dimensions: the gating coefficient of each dimension is Hadamard producted with the basic feature representation of the corresponding dimension, and then summed along the feature dimension to generate a global feature vector that combines spatiotemporal and category information. .

[0065] It should be noted that the order The gating weight matrix of the dimension feature Through federated learning training, each participant calculates the gradient based on local data, and the central server aggregates the gradient and updates Parameters, and finally form a unified gate weight matrix. During the training process, the characteristics of time, space, and category dimensions correspond to independent matrix.

[0066] The out-of-stock quantity prediction module builds a distributed time series prediction model based on a dynamic warehouse relationship graph, a time series GRU layer, a spatial graph convolution layer, and a cross-warehouse attention layer. Based on the global feature vector, it predicts the demand for goods in each warehouse over a period of time. Combined with real-time inventory data, it calculates the expected out-of-stock quantity of goods in each warehouse and forms an inventory gap list.

[0067] Based on the dynamic warehouse relationship graph, time series GRU layer, spatial graph convolution layer and cross-warehouse attention layer, a distributed time series prediction model is constructed as follows:

[0068] Taking each warehouse as a node and the geographical distance between warehouses as an edge, we use the inverse distance weighting method to construct a weighted adjacency matrix, and use ST-GCN to aggregate spatiotemporal features to generate a dynamic warehouse relationship graph.

[0069] Furthermore, the latitude and longitude coordinates of all warehouses are first obtained, and the spherical distances between each pair are calculated as the initial edge weights. The original distances are processed using the inverse distance weighting method, and the edge weights are converted into values inversely proportional to the distances, forming the initial version of the weighted adjacency matrix. During the ST-GCN operation, the weighted adjacency matrix is input together with the time series features output by the time series GRU layer. The graph convolution kernel is approximated using Chebyshev polynomial expansion, and feature transformation is performed in the spectral domain. During the spatiotemporal feature aggregation stage, the edge weights of the adjacency matrix are dynamically adjusted so that they are updated as the actual frequency of goods transfers between warehouses changes. The resulting dynamic warehouse relationship graph is able to adaptively capture real-time changes in logistics relationships between warehouses, effectively improving the prediction accuracy of subsequent commodity demand forecasts.

[0070] The time series GRU layer is based on the time dimension features in the global feature vector, and performs linear transformation and gating mechanism processing through the GRU unit to generate time series features with temporal dependencies.

[0071] Furthermore, when the time series GRU layer processes the temporal features in the global feature vector, the GRU unit first receives a feature sequence arranged by time step. The update gate calculates the ratio coefficient for retaining the historical state, and the reset gate determines the fusion weight of the current input and the historical state. The candidate hidden state combines the current input and the reset historical state to generate new memory content. The final hidden state is balanced by the update gate based on the ratio of the historical state to the candidate state. During the calculation process, the temporal features undergo a linear transformation and are then fed into the update gate, reset gate, and candidate state calculation path. The parameters of each path are optimized through backpropagation during training. The GRU unit outputs a hidden state sequence that preserves long-term dependencies, which serves as a temporal feature with temporal dependencies. The gating mechanism uses a sigmoid activation function to generate a gating value between 0 and 1, and the candidate state uses a Tanh activation function to ensure gradient stability. The dimensions of the temporal features are consistent with the temporal features of the input, facilitating processing by the subsequent spatial graph convolution layer.

[0072] The spatial graph convolution layer uses Chebyshev polynomial approximation to combine temporal features with the dynamic warehouse relationship graph and generates spatial correlation feature vectors through graph convolution operations.

[0073] Furthermore, when processing time series features and the dynamic warehouse relationship graph, the spatial graph convolution layer first performs eigendecomposition on the normalized Laplacian matrix of the dynamic warehouse relationship graph. The graph convolution kernel is approximated using Chebyshev polynomial expansion, with the expansion order set to 2 to balance computational efficiency and accuracy. Time series features are input as node features and then matrix multiplied with Chebyshev polynomial terms. Each order of polynomial term corresponds to independent, trainable convolution kernel parameters. During the computation, the zero-order term preserves the node's own characteristics, the first-order term aggregates features of directly adjacent nodes, and the second-order term captures broader neighborhood information. The weighted sum of the results of each order is then transformed nonlinearly using an activation function to generate a spatial correlation feature vector that incorporates spatiotemporal information. Edge weights in the dynamic warehouse relationship graph adjust the contribution of feature aggregation in real time, ensuring that the spatial correlation feature vector reflects the latest logistics network status. The final output spatial correlation feature vector maintains the same dimensionality as the input time series features, facilitating subsequent processing by the cross-warehouse attention layer.

[0074] The cross-warehouse attention layer predicts the demand for goods in each warehouse in the future through the cross-node attention mechanism based on the spatial correlation feature vector.

[0075] Furthermore, when processing spatially correlated feature vectors, the cross-warehouse attention layer first inputs the feature vectors of each warehouse node as the query vector and key vector into a multi-head attention calculation. The dot-product attention score measures the strength of demand correlation between warehouses. The attention score is normalized using Softmax and used as a weight coefficient to perform weighted aggregation on the warehouse features represented by the value vector. During the calculation process, each attention head independently learns different feature interaction patterns, and the temporal features output by the time series GRU layer are incorporated into the query vector as positional encoding. Finally, the output features of multiple attention heads are concatenated and linearly transformed to generate a forecast reflecting cross-warehouse synergy, directly outputting the product demand for each warehouse over a period of time.

[0076] The global feature vector is input into the distributed time series forecasting model to predict the demand for goods in each warehouse in the future. The expression is:

[0077] ;

[0078] in, Represents a warehouse Products in The quantity of goods demanded at time t+m is: is the Sigmoid activation function, is the attention head number, It's a warehouse The set of neighbor nodes of It is A warehouse for attention head calculation With warehouse The demand association weight, is the order index of the Chebyshev polynomial expansion (value range 0≤ ≤2), is the Chebyshev polynomial The convolution kernel parameters of order, It is Chebyshev polynomials of order in the normalized Laplace matrix The calculation results on Indicates at time Time Warehouse The global eigenvector of Indicates the current moment, Represents the predicted future time and the current moment time interval, is the normalized Laplacian matrix.

[0079] Furthermore, when calculating the demand for warehouse goods, we first normalize the Laplace matrix Perform Chebyshev polynomial expansion and calculate the zero-order , first-order and second-order Polynomial terms. Each warehouse The global eigenvector of Multiply them with the third-order Chebyshev polynomial terms and then perform weighted summation. The convolution kernel parameters The cross-warehouse attention layer uses 4 independent attention heads for parallel calculation, each attention head Computing Warehouse Its neighbor warehouse Demand association weight , processed by Softmax normalization. Each attention head multiplies the associated weight with the corresponding polynomial transformation feature and aggregates it. The results of the four heads are added together and activated by the Sigmoid function. Output final goods demand .

[0080] It should be noted that The calculation results of the Chebyshev polynomial of order are used to approximate the frequency domain response of the graph convolution kernel, for example, ∈{0, 1, 2} corresponds to the recursive value of 1, x, 2x²-1 respectively; warehouse The set of neighbor nodes Refers to all warehouses The set of adjacent warehouse nodes with logistics transfer relationship; Chebyshev polynomial The convolution kernel parameters of the order.

[0081] Based on the commodity demand, real-time inventory data, and in-transit inventory of each warehouse in the future, the expected out-of-stock quantity of each warehouse is calculated using the threshold zeroing method. The expression is:

[0082] ;

[0083] in, Represents a warehouse Products in In the Expected stockout in days, Represents a warehouse Products in In the Daily demand, Indicates the number of days, Represents a warehouse Products in Actual available inventory, Represents a warehouse Products in of in-transit inventory;

[0084] Furthermore, when calculating the expected out-of-stock quantity of warehouse goods, first obtain the future Tiancangku Medium Commodity Demand , which is derived from the output of the demand forecasting model. Real-time inventory data Record the current warehouse Medium Commodity The actual available inventory quantity, inventory in transit Contains the quantity of goods that have been ordered but not delivered. The calculation process will be demand Subtract actual available inventory and in-transit inventory If the result is negative, take zero, if it is positive, keep the original result as the warehouse The expected out-of-stock quantity G of the product.

[0085] It should be noted that real-time inventory data is obtained by scanning the actual number of goods currently stored in the warehouse and is checked against ERP records daily to ensure accuracy. In-transit inventory data is extracted from logistics transportation documents, summarizing the number of all goods that have been shipped but have not yet arrived at the warehouse, combined with the arrival time of the transportation plan.

[0086] The threshold zeroing method is a mathematical method that calculates the expected shortage of warehouse goods by taking the demand Subtract actual available inventory and in-transit inventory If the difference is negative, output zero. If it is non-negative, retain the original value as the final out-of-stock quantity. This method ensures that the expected stock-out quantity is always a non-negative real number, which meets the actual business constraint that the stock-out quantity cannot be negative in the inventory management scenario. The calculation process is strictly based on the input demand quantity. , Actual available inventory and In-transit Library Three parameters, no additional corrections or adjustments are introduced. Directly reflects the number of out-of-stocks that may occur in the future under the current inventory situation.

[0087] Based on the commodity demand of each warehouse and the expected shortage of commodities in each warehouse in the future period, an inventory gap list containing warehouse number, commodity code, shortage date and expected shortage of commodities in each warehouse is generated.

[0088] Furthermore, firstly, the demand for goods in the future days Expected out-of-stock quantity for the corresponding warehouse-product combination Match by warehouse number and product code Create an index relationship. Then for each expected out-of-stock quantity A record is created for a warehouse-product combination greater than zero, which includes the warehouse number, product code, and out-of-stock date. and the calculated expected stockout Value. All records sorted by out-of-stock date Sort in ascending order, records with the same date are sorted by warehouse number and product code Sorting, and finally forming a structured inventory gap list.

[0089] The priority assessment module calculates the transfer weights and inventory transfer priority scores between warehouses based on the inventory gap list and generates a QUBO matrix;

[0090] Based on the warehouse location and expected shortage quantity of each warehouse in the inventory gap list, the transfer weights between warehouses are calculated using the inverse distance weighted method with time and space constraints. The expression is:

[0091] ;

[0092] in, It's a warehouse To the warehouse The allocation weight when allocating goods (the larger the value, the higher the allocation priority), Represents a warehouse The current expected out-of-stock quantity of the product, Indicates the current warehouse arrive The actual transport distance, is the distance attenuation coefficient, Indicates the estimated transportation time obtained in real time through TMS. represents the mean transportation time of goods, Indicates the transport time tolerance, Indicates the current average inventory level of the product. Represents a warehouse The current available inventory of is the benchmark time for commodity transportation (the value range is (0, +∞), and the specific value is determined by statistical analysis of historical transportation time data);

[0093] Further, calculate the inter-warehouse transfer weight When first obtaining the warehouse based on the inventory gap list Expected out-of-stock quantity of products , while extracting warehouses from the logistics database arrive Actual transport distance Distance term Attenuation coefficient over distance Exponential transformation processing to form the distance attenuation factor The time factor calculation part obtains the estimated transportation time through the transportation management system , and the average transportation time of goods After comparison, the Gaussian kernel function is used Quantify the impact of time deviation. Inventory factors are partially taken from warehouses Current available inventory With warehouse Expected stockout the smaller of the current average inventory level of the item , and obtain the inventory adjustment factor Final allocation weight It is composed of three products: As the basic weight, the time factor is used as the time penalty item, the inventory factor is used as the inventory adjustment item, and the product of the three is divided by the distance attenuation factor. .

[0094] The transportation time tolerance should be stated This indicates the acceptable range of fluctuations in actual transit time from the standard transit time. The transit time tolerance is determined through statistical analysis of historical transit time data and is calculated as the standard deviation of all shipments recorded along the same route. The value range is zero to positive infinity.

[0095] The distance attenuation coefficient & ranges from (0, +∞). In practical applications, it is usually between 1.0 and 2.0. The specific value is determined based on the sensitivity of transportation costs and business scenario requirements.

[0096] Based on the expected stockout quantity and historical turnover rate of each warehouse in the inventory gap list, the inventory transfer priority score is calculated through the dynamic index coupling algorithm and the turnover rate deviation amplification mechanism. The expression is:

[0097] ;

[0098] in, It's a warehouse The inventory transfer priority score, It's a warehouse Inventory turnover cycle, It's a warehouse The historical turnover rate, is the average turnover rate of all warehouses, is the turnover elasticity coefficient, represents the average out-of-stock quantity of all warehouses, is the standard deviation of stockouts, is the stock-out elasticity coefficient;

[0099] Furthermore, first the warehouse Expected stockout Divided by the inventory turnover period , the result is the out-of-stock elasticity coefficient The basic scoring item is obtained by power. At the same time, the warehouse is calculated Historical turnover rate and average turnover rate The ratio of turnover elasticity coefficient The power is multiplied by the deviation amplification factor, which is determined by the warehouse Out of stock Average out-of-stock quantity The absolute difference divided by twice the standard deviation Add 1 to form. Finally, the warehouse Expected stockout and historical turnover rate The square root of the product of is used as a supplementary scoring item. The three results are added together to get the final priority score The higher the score, the better the warehouse The higher the allocation priority.

[0100] It should be noted that the turnover elasticity coefficient The value range of is (0, +∞). In practical applications, it is usually between 0.5 and 2.0, which is used to control the impact of turnover rate on priority score.

[0101] Out-of-stock elasticity coefficient The value range of is (0, +∞), and in practical applications it is usually between 1.0 and 3.0, which is used to adjust the nonlinear amplification effect of out-of-stock quantity on priority score.

[0102] Based on the transfer weights and inventory transfer priority scores between warehouses, the QUBO matrix is integrated and generated through the quadratic unconstrained binary optimization modeling method.

[0103] Furthermore, when generating the QUBO matrix, the transfer weights between warehouses are first calculated. and the corresponding receiving warehouse Inventory transfer priority score Multiplying them together gives the basic allocation benefit term. At the same time, constraints are constructed to ensure that the outbound volume of each warehouse does not exceed the available inventory and the inbound volume does not exceed the out-of-stock volume. Penalty coefficients are imposed when the constraints are violated. The basic allocation benefit term and the constraint penalty term are added together to form the objective function. The coefficients of the quadratic terms in the objective function are directly filled into the diagonal and off-diagonal positions of the QUBO matrix. The rows and columns of the matrix correspond to the allocation decision variables of each warehouse, and the values of the matrix elements reflect the overall benefit of the allocation plan and the degree of constraint satisfaction. The dimension of the QUBO matrix is equal to the total number of all possible allocation decision variables, and the symmetry of the matrix ensures the solvability of the optimization problem. The resulting QUBO matrix can be directly input into the quantum annealing algorithm to solve the optimal inventory allocation instructions.

[0104] The inventory transfer module uses the quantum annealing algorithm based on the QUBO matrix to search for the global optimal solution, decodes it through the encoding rules, generates the initial inventory transfer instruction set, and uses MILP for optimization to generate the optimal inventory transfer instructions and execute them.

[0105] Initialize quantum annealing parameters based on historical allocation data;

[0106] It should be noted that the quantum annealing parameters include annealing time and temperature profile.

[0107] Based on the initialized quantum annealing parameters, an energy minimization search is performed through quantum tunneling effect and thermal fluctuations to generate a binary solution set after quantum annealing.

[0108] Furthermore, at the beginning of the quantum annealing process, the quantum processor maps the QUBO matrix to the qubit's coupling strength and bias magnetic field based on the initialized annealing time and temperature curve parameters. During the annealing process, the qubit initially resides in a superposition of all possible states. As the temperature curve evolves from high to low temperatures, quantum tunneling enables the qubit to traverse energy barriers and explore energy states associated with different combinations of allocation decisions. Thermal fluctuations help the qubit escape local minima, facilitating the discovery of a global optimal solution. When the temperature drops below a critical point, the qubit gradually collapses to a classical state, the quantum tunneling effect weakens, the impact of thermal fluctuations decreases, and the qubit state stabilizes. Finally, the qubit state is converted to a binary solution through quantum measurement.

[0109] Define coding rules based on warehouse quantity, commodity type, and transfer quantity range in historical transfer data;

[0110] Furthermore, we first analyzed the warehouse distribution, product classification, and actual transfer quantity ranges in historical transfer data to determine the required binary representation length for each element. Warehouse numbers were sequentially coded to correspond to actual warehouse IDs, product codes were hierarchically organized by category, and transfer quantity values were represented using standard binary format. Each field was combined into a complete coding unit of a fixed length, ensuring that each warehouse-product combination had a unique corresponding binary sequence.

[0111] Decode the binary solution set into transfer decision variables according to the encoding rules, and extract the corresponding warehouse number, product and transfer quantity through hash search method, and finally integrate them to generate the initial inventory transfer instruction set;

[0112] Furthermore, when decoding the binary solution set, the quantum bit state sequence is first divided into several decision units according to predefined encoding rules, with each unit corresponding to an allocation decision variable. The activated binary bits are searched in the pre-built allocation decision mapping table using a hash search method, and the activated binary bits are converted into a specific warehouse number and product code combination. The allocation quantity value is determined by the length of continuously activated binary bits, and Gray code encoding is used to ensure numerical continuity. After all activated decision units have undergone integrity verification, they are sorted in ascending order by warehouse number. Allocation instructions for the same warehouse are sorted by product code to generate a structured initial inventory allocation instruction set. The hash search process uses a perfect hash function to avoid conflicts and ensure that each binary combination uniquely corresponds to an allocation decision.

[0113] MILP is used to optimize the initial inventory transfer instruction set, generate the optimal inventory transfer instruction and execute it.

[0114] Furthermore, a mathematical model consisting of an integer transfer quantity variable and a binary transportation route variable was first established. The MILP objective function was constructed as a linear combination of transportation cost, inventory cost, and stockout cost. The MILP constraints included the inventory balance equation, transportation capacity inequality, and time window constraint inequality. A branch-and-cut approach was used to solve the MILP problem. This approach used linear programming relaxation to obtain a lower bound, applied a branching strategy to handle integer constraints, and added valid inequalities to tighten the feasible domain. The MILP solver outputs an optimal solution that satisfies all integer constraints, generating an optimal inventory transfer instruction that includes a precise transfer quantity, a clear transportation route, and a specific time schedule.

[0115] In summary, the present invention uses the STL algorithm, Geohash coding, and TF-IDF weighted coding to extract the three-dimensional features of orders, namely time, space, and category, respectively. This achieves a comprehensive characterization of the multidimensional characteristics of order data, enabling subsequent prediction models to capture richer spatiotemporal and product association patterns. Furthermore, the quantum annealing algorithm is used to optimize the QUBO matrix, and a global optimal search is performed in an ultra-large-scale solution space through the quantum tunneling effect and thermal fluctuation mechanism. Furthermore, secondary optimization is combined with MILP to ensure that the generated transfer instructions minimize overall costs while satisfying complex business constraints.

[0116] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An inventory management system based on order data, characterized by: include, The data collection module collects and pre-processes the original order data of each participant to extract the three-dimensional features of each order. The original order data includes the order timestamp, product SKU code, order quantity, and delivery address. The three-dimensional features of each order include time dimension features, space dimension features and category dimension features; The feature fusion module jointly encrypts the three-dimensional features of different orders through a federated learning algorithm to generate a homomorphically encrypted data packet. It then uses a federated gradient descent algorithm for collaborative training to generate a cross-domain feature similarity matrix. Finally, it generates a global feature vector through gated attention feature fusion. The out-of-stock quantity prediction module builds a distributed time series prediction model based on a dynamic warehouse relationship graph, a time series GRU layer, a spatial graph convolution layer, and a cross-warehouse attention layer. Based on the global feature vector, it predicts the demand for goods in each warehouse over a period of time. Combined with real-time inventory data, it calculates the expected out-of-stock quantity of goods in each warehouse and forms an inventory gap list. The priority assessment module calculates the transfer weights and inventory transfer priority scores between warehouses based on the inventory gap list and generates a QUBO matrix; The inventory transfer module uses the quantum annealing algorithm based on the QUBO matrix to search for the global optimal solution, decodes the encoded rules, generates the initial inventory transfer instruction set, and uses MILP optimization to generate the optimal inventory transfer instructions and execute them. The steps of constructing a distributed time series forecasting model and predicting the demand for goods in each warehouse in the future period of time based on the global feature vector are as follows: Build a distributed time series prediction model based on a dynamic warehouse relationship graph, a time series GRU layer, a spatial graph convolution layer, and a cross-warehouse attention layer; Taking each warehouse as a node and the geographical distance between warehouses as an edge, we use the inverse distance weighting method to construct a weighted adjacency matrix, and use ST-GCN to aggregate spatiotemporal features to generate a dynamic warehouse relationship graph. The time series GRU layer generates time series features with temporal dependencies by performing linear transformation and gating processing through GRU units based on the time dimension features in the global feature vector; The spatial graph convolution layer uses Chebyshev polynomial approximation to combine time series features with the dynamic warehouse relationship graph, and generates spatial correlation feature vectors through graph convolution operations; The cross-warehouse attention layer predicts the demand for goods in each warehouse in the future through the cross-node attention mechanism based on the spatial correlation feature vector.

2. The order data-based inventory management system according to claim 1, wherein: The steps of generating the global feature vector are as follows: The three-dimensional features of each participant's order are jointly encrypted using the Paillier homomorphic encryption algorithm to generate a homomorphic encrypted data packet. The secure inner product protocol is used to calculate the multi-dimensional feature similarity of homomorphically encrypted data packets and generate a cross-domain feature similarity matrix. The cross-domain feature similarity matrix is used as the weight constraint of the federated gradient descent algorithm to collaboratively train the 3D features of each participant's order and generate a shared projection matrix with cross-enterprise feature alignment. The shared projection matrix is used as a linear transformation operator to map the three-dimensional features of the orders of each participant to a unified dimension. The three-dimensional features of the orders of each participant are fused through gated attention feature fusion to generate a global feature vector.

3. The order data-based inventory management system according to claim 2, wherein: The above steps are as follows: Based on the commodity demand, real-time inventory data, and in-transit inventory of each warehouse in the future, the expected out-of-stock quantity of each warehouse's commodities is calculated using the threshold zeroing method; Based on the commodity demand of each warehouse and the expected shortage of commodities in each warehouse in the future period, an inventory gap list containing warehouse number, commodity code, shortage date and expected shortage of commodities in each warehouse is generated.

4. The order data-based inventory management system according to claim 1, wherein: The steps to calculate the transfer weights between warehouses and the inventory transfer priority scores and generate the QUBO matrix are as follows: Based on the warehouse location and expected shortage quantity of each warehouse in the inventory gap list, the transfer weights between warehouses are calculated using the inverse distance weighting method with time and space constraints. Based on the expected stockout quantity and historical turnover rate of each warehouse in the inventory gap list, the inventory allocation priority score is calculated through a dynamic index coupling algorithm and a turnover rate deviation amplification mechanism; Based on the transfer weights and inventory transfer priority scores between warehouses, the QUBO matrix is integrated and generated through the quadratic unconstrained binary optimization modeling method.

5. The order data-based inventory management system according to claim 4, characterized in that: The use of the quantum annealing algorithm to search for the global optimal solution refers to initializing quantum annealing parameters, and performing energy minimization search through quantum tunneling effect and thermal fluctuations to generate a binary solution set after quantum annealing.

6. The order data-based inventory management system according to claim 5, characterized in that: The steps to generate and execute the optimal inventory transfer instruction are as follows: Define coding rules based on warehouse quantity, commodity type, and transfer quantity range in historical transfer data; Decode the binary solution set into transfer decision variables according to the encoding rules, and extract the corresponding warehouse number, product and transfer quantity through hash search method, and finally integrate them to generate the initial inventory transfer instruction set; MILP is used to optimize the initial inventory transfer instruction set, generate the optimal inventory transfer instruction and execute it.

7. The order data-based inventory management system according to claim 1, wherein: The original order data includes the order timestamp, product SKU code and order quantity.

8. The order data-based inventory management system according to claim 1, wherein: The preprocessing includes data cleaning, dimension alignment and normalization of the order data.

Images