A library inventory management method and system

Abstract

The present invention provides a library inventory management method and system, which relates to the field of book management technology, including obtaining metadata of the library, including book information and user borrowing records, constructing a knowledge graph based on these data, and extracting user query intentions using a heterogeneous graph matching algorithm to obtain a target node set. Then, semantic representation information of user intentions is extracted based on multi-resolution wavelet transform and frequency domain analysis. A sparse Bayesian learning algorithm is used to predict future borrowing demand and generate demand prediction results. An inventory allocation plan is generated by combining inventory allocation and logistics costs through a multi-objective programming algorithm and a convex optimization method, and the plan is optimized through a Lagrangian relaxation method. Finally, the inventory allocation plan is adjusted in real time in combination with a dynamic system state estimation method. The present invention can significantly improve the accuracy and flexibility of library inventory management, reduce operating costs, and improve resource utilization.

Description

Technical Field

[0001] The present invention relates to the technical field of book management, and in particular to an inventory management method and system for a library. Background Art

[0002] With the continuous growth of library collections and user access needs, traditional inventory management methods are facing more and more challenges. Existing technologies usually rely on fixed rules or simple statistical analysis of historical data to make inventory replenishment and allocation decisions. For example, by analyzing borrowing history records, a predetermined inventory threshold is set, and replenishment is triggered when the inventory level falls below the threshold. However, this method cannot capture the dynamic changes in user demand in real time, resulting in delayed resource allocation, relying only on historical data, lacking a deep understanding of user behavior, and prone to supply and demand imbalances. In addition, existing technologies still have shortcomings such as insufficient prediction accuracy, high optimization costs, and poor dynamic adaptability when solving inventory management problems.

[0003] Therefore, there is an urgent need for a library inventory management method and system to solve the above problems. Summary of the invention

[0004] The purpose of the present invention is to provide a library inventory management method and system to improve the above problems. In order to achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0005] In a first aspect, the present application provides a library inventory management method, comprising:

[0006] Obtaining metadata of the library, wherein the metadata of the library includes book information and user's borrowing records;

[0007] A knowledge graph is constructed based on the library metadata, and the knowledge graph is processed using a heterogeneous graph matching algorithm, wherein a target node set on the knowledge graph is obtained by extracting and analyzing the user query intention;

[0008] Capturing local semantic features on the knowledge graph based on multi-resolution wavelet transform and target node set, and extracting feature embedding through frequency domain analysis to obtain semantic representation information of user intention;

[0009] Based on the semantic representation information of the user's intention, the user's borrowing record and a sparse Bayesian learning algorithm, the borrowing demand at a preset time point is predicted to obtain a demand prediction result for each book;

[0010] The inventory allocation plan is constructed based on the demand forecast results. The convex optimization method and multi-objective programming algorithm are used to balance the inventory allocation and logistics costs, and the inventory allocation plan is obtained through the Lagrangian relaxation method.

[0011] The inventory allocation plan and the metadata of the library acquired in real time are processed based on a dynamic system state estimation method, and the inventory allocation plan is adjusted in real time according to the processing results to obtain a final inventory management plan.

[0012] In a second aspect, the present application also provides a library inventory management system, including:

[0013] An acquisition unit, used for acquiring metadata of a library, wherein the metadata of the library includes book information and borrowing records of users;

[0014] A first processing unit is used to construct a knowledge graph based on the library metadata and process the knowledge graph using a heterogeneous graph matching algorithm, wherein a target node set on the knowledge graph is obtained by extracting and analyzing the user query intention;

[0015] A second processing unit is used to capture local semantic features on the knowledge graph based on multi-resolution wavelet transform and target node set, and extract feature embedding through frequency domain analysis to obtain semantic representation information of user intention;

[0016] A third processing unit is used to predict the borrowing demand at a preset time point based on the semantic representation information of the user's intention, the user's borrowing record and a sparse Bayesian learning algorithm to obtain a demand prediction result for each book;

[0017] A fourth processing unit is used to construct an inventory allocation plan according to the demand forecast result, wherein a convex optimization method and a multi-objective programming algorithm are used to balance inventory allocation and logistics costs, and the inventory allocation plan is obtained through a Lagrangian relaxation method;

[0018] The fifth processing unit is used to process the inventory allocation plan and the metadata of the library acquired in real time based on the dynamic system state estimation method, and adjust the inventory allocation plan in real time according to the processing result to obtain the final inventory management plan.

[0019] The beneficial effects of the present invention are:

[0020] The present invention models user behavior based on the knowledge graph, captures user query intentions by mining the potential relationships between graph nodes, and uses a path extraction algorithm to accurately locate target nodes, thereby improving the depth and accuracy of demand forecasting; captures local features of the knowledge graph through multi-resolution wavelet transform, generates a semantic representation of user intentions in combination with frequency domain analysis, and enhances the understanding of user needs; uses a sparse Bayesian learning algorithm to dynamically model and optimize the prediction of user needs, effectively improving prediction accuracy and reducing computational complexity; balances inventory allocation and logistics costs based on convex optimization and multi-objective planning methods, and adjusts the allocation plan in real time in combination with a dynamic state estimation method, ensuring the efficiency, accuracy and flexibility of inventory management.

[0021] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or be understood by implementing the embodiments of the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0023] Figure 1 A schematic diagram of a library inventory management method flow chart according to an embodiment of the present invention;

[0024] Figure 2 It is a schematic diagram of the structure of the library inventory management system described in an embodiment of the present invention.

[0025] In the figure: 701, acquisition unit; 702, first processing unit; 703, second processing unit; 704, third processing unit; 705, fourth processing unit; 706, fifth processing unit. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0027] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0028] Embodiment 1:

[0029] This embodiment provides an inventory management method for a library.

[0030] See also Figure 1 , the figure shows that the method includes step S1, step S2, step S3, step S4, step S5 and step S6.

[0031] Step S1, obtaining metadata of the library, wherein the metadata of the library includes book information and user's borrowing records;

[0032] It can be understood that this step is to build a high-quality data foundation to provide support for the subsequent construction of the knowledge graph and user behavior analysis. The metadata of the library mainly includes two categories: book information (such as book title, author, classification, publication year, keywords, collection status, etc.) and user borrowing records (such as user ID, borrowing time, borrowing frequency, book return time, etc.) Through this step, the system builds a complete and standardized metadata foundation, which can not only support the construction of the subsequent knowledge graph, but also lay a data foundation for user behavior analysis, ensuring the accuracy and reliability of the data-driven method. The technical effect is reflected in: on the one hand, the utilization efficiency of metadata is improved by integrating multi-source heterogeneous data; on the other hand, the data quality is improved through data cleaning and normalization, providing strong support for the operation of subsequent algorithms.

[0033] Step S2: construct a knowledge graph based on the library metadata, and process the knowledge graph using a heterogeneous graph matching algorithm, wherein a target node set on the knowledge graph is obtained by extracting and analyzing the user query intention;

[0034] It can be understood that this step, on the one hand, reveals the complex relationship between books and users by constructing a multi-dimensional knowledge graph and applying a heterogeneous graph matching algorithm, accurately reflecting the interests and needs of users; on the other hand, the introduction of the path extraction algorithm enables the system to accurately capture the user's query intention from the huge graph, thereby improving the accuracy and depth of demand prediction. In this step, step S2 includes step S21, step S22, step S23 and step S24.

[0035] Step S21, constructing a heterogeneous graph structure according to the library metadata, and generating a sparse heterogeneous graph based on a graph weighted downsampling screening method to obtain an initial skeleton of the heterogeneous graph;

[0036] It can be understood that this step first models the library metadata as a heterogeneous graph containing multiple types of nodes and edges based on the diversity of the library metadata. Among them, the nodes can be divided into book nodes, user nodes, and related attribute nodes (such as classification, keywords, authors, etc.). The edges reflect the relationship between different nodes, such as the relationship between users borrowing books, the relationship between books sharing the same classification, and the relationship between users rating books. By introducing multiple types of nodes and edges, the heterogeneous graph can fully express the multi-dimensional relationship between books and users.

[0037] In addition, in this step, the weighted downsampling screening method of the graph adopts a distributed weighted downsampling algorithm based on local information, such as the local quadratic approximation method to perform local analysis on the neighborhood information of each node, and find the edge with the least impact on the global structure for screening by minimizing the quadratic loss function. This method can retain the core structural features of the graph while significantly reducing the computational complexity.

[0038] Step S22, an implicit relationship mining algorithm based on random walk is used to capture the potential connections between nodes of the initial skeleton of the heterogeneous graph, and edge weights are expanded through path information to obtain an enhanced heterogeneous graph;

[0039] It can be understood that this step first defines the transition probability of the node and designs the weighted jump rules according to the node and edge types of the heterogeneous graph. In order to improve the pertinence of the walk, a bias factor (such as node importance or path length constraint) is introduced. A set of specific meta-paths (such as "user→book→user" or "book→category→book") is defined to guide the walk to focus on specific semantic associations along these paths. A large number of random walk trajectories are counted, and the co-occurrence probability between each pair of nodes is calculated to discover the implicit relationship between nodes that are not directly connected by edges; then, according to the implicit relationship captured by the random walk, the edge weights in the graph are expanded and optimized. Finally, the importance of the path is used as the basis for adjusting the edge weight. For example, the weight of the original edge is weighted and updated according to the number of co-occurrences of the path, the path length, and the meta-path type. By mining the implicit relationship in the heterogeneous graph through random walks, this step effectively makes up for the node-to-node connection that may be missed in the initial skeleton graph, and at the same time enhances the semantic expression ability of the edge weight. The enhanced heterogeneous graph finally generated can better reveal the complex potential connections between users and books, improve the accuracy of subsequent path extraction and user intent analysis, and significantly improve the system's adaptability and processing capabilities for complex query scenarios.

[0040] Step S23: clustering the enhanced heterogeneous graph based on a node clustering algorithm, and generating node categories through the clustering results to obtain preliminary semantic partitions in the graph;

[0041] It can be understood that in a heterogeneous graph environment, the relationships between nodes are diverse and heterogeneous. In this step, first, the edge weight information of the enhanced heterogeneous graph is used to construct an adjacency matrix, and the normalized Laplacian matrix is ​​generated in combination with the node degree to capture the global structural characteristics of the graph. Subsequently, the Laplacian matrix is ​​eigendecomposed, and the eigenvectors are extracted to form a feature space after dimensionality reduction, and the k-means clustering algorithm is applied in this space to divide the graph nodes into several categories. The nodes of each category are defined as a semantic partition, in which the nodes have a high semantic similarity or association strength, such as clustering books that users have borrowed or books with the same subject into the same partition. Through this clustering method, not only the structured semantic partitioning of complex heterogeneous graphs is achieved, but also the implicit associations across node types are revealed, while reducing the computational complexity of high-dimensional graphs. This preliminary semantic partitioning provides a solid foundation for subsequent path extraction and accurate parsing of user query target nodes.

[0042] Step S24: A path extraction algorithm based on multi-level graph search performs a preliminary semantic partition search path in the graph to obtain a set of target nodes for the user query.

[0043] It can be understood that in this step, the key semantic information in the user query is first analyzed to locate the semantic partition most relevant to the query content, narrowing the search scope to improve the efficiency of the algorithm. Subsequently, the Hierarchical Depth-First Search (HDFS) algorithm is applied to the selected semantic partition, combining the topological characteristics of the graph to expand the search scope layer by layer, and dynamically evaluate the matching degree between the path weight and the user's intention.

[0044] In order to avoid path redundancy and computational waste during the search process, a heuristic pruning strategy is adopted to eliminate paths that are irrelevant to the query semantics or have low weights during the search process. At the same time, a path expansion method based on a priority queue is used to prioritize the exploration of paths with higher matching degrees, thereby quickly converging to the optimal result. Finally, the highly relevant paths obtained by the search are deconstructed into a set of target nodes, including all important nodes directly or indirectly pointed to by the user query. This method not only achieves efficient path extraction in complex knowledge graphs, but also improves the accuracy of user intent parsing, laying a data foundation for subsequent semantic feature extraction and demand prediction.

[0045] Step S3, capturing local semantic features on the knowledge graph based on multi-resolution wavelet transform and target node set, and extracting feature embedding through frequency domain analysis to obtain semantic representation information of user intention;

[0046] It can be understood that this step combines multi-resolution wavelet transform and frequency domain analysis, which not only ensures the keen capture of local features, but also realizes the efficient embedding of global features, effectively improves the depth and accuracy of user intention representation, and provides key input data for subsequent demand prediction. In this step, step S3 includes step S31, step S32, step S33 and step S34.

[0047] Step S31, extracting a local subgraph based on the knowledge graph and the target node set, wherein the local subgraph is captured based on a hierarchical clustering algorithm to obtain a local graph related to the user intention;

[0048] It can be understood that in this step, the target node set is used as the initial input, and the connection relationship, node features and edge weights in the graph are used to recursively identify and aggregate nodes and edges with strong semantic association with the target node. Among them, the hierarchical clustering algorithm relies on the dual clues of topology and semantics to capture subgraphs. First, starting from the target node, an initial neighborhood is constructed, and the first-level subgraph boundary is screened by edge weights and semantic similarity metrics between nodes (such as cosine similarity or Euclidean distance based on embedded vectors). Subsequently, the neighborhood is gradually expanded using an iterative aggregation method until the connectivity between nodes in the subgraph decreases significantly or the external connection weight exceeds the set threshold. The segmentation strategy also supports multi-scale adjustment, that is, a trade-off between local compactness and global coverage is made to ensure that the captured subgraph is moderate in scale, which can fully reflect the context of the target node and avoid information redundancy. This method effectively reduces the computational complexity of subsequent semantic feature extraction, so that the processing focuses on the subset directly related to the user's intention; enhances the integrity of the semantic context of the target node, and makes the representation of the user's query intention in the local graph more logically consistent.

[0049] Step S32, performing multi-resolution wavelet transform processing according to the local graph related to the user intention, wherein local semantic features are captured by decomposing the graph signal to obtain a multi-scale feature representation;

[0050] It can be understood that the first step of graph signal decomposition is to generate the spectral basis of the graph based on the graph Laplacian operator. These spectral bases reflect the global and local connectivity patterns of the graph and provide a frequency domain reference framework for signal processing. Subsequently, a multi-resolution graph wavelet kernel function is constructed in combination with the local graph characteristics related to user intention. The wavelet kernel can dynamically enhance or smooth the signal locally by adjusting the expansion scale parameter, thereby resolving the low-frequency (global characteristics) and high-frequency (local characteristics) components in the local graph. The execution of this multi-resolution wavelet transform is divided into multiple steps: first, the graph signal of the local graph is initialized and set to the function value of the node features or edge weights; second, the signal at each layer of resolution is deconstructed by applying the wavelet transform layer by layer; finally, the feature components at each resolution are extracted, and multi-scale feature representations are generated by feature vector aggregation. These feature representations can reveal hidden semantic patterns in the local graph, such as community distribution characteristics or relationship networks between key nodes.

[0051] The flexibility of wavelet transform allows adaptive processing of local graphs of different scales, ensuring the universality of the algorithm for heterogeneous graph data; and the resulting multi-scale feature representation provides a strong foundation for subsequent frequency domain analysis and semantic decoding, helping to more accurately refine user intent.

[0052] Step S33, performing frequency domain analysis according to the multi-scale feature representation, wherein global semantic features are extracted based on sparse Fourier transform spectrum embedding to obtain frequency domain feature embedding;

[0053] It can be understood that in this step, frequency domain analysis is a way to understand the structure of a graph from a global perspective. In graph analysis, frequency domain information helps to reveal potential global connections between graph nodes, which may not be reflected by simple local structures. Through frequency domain analysis, graph signals can be converted from the time domain (or node domain) to the frequency domain, which helps to capture the global patterns or trends of the graph at different scales, which are often difficult to directly reflect in the adjacency matrix of the graph. In this step, the sparse Fourier transform is based on the Laplace matrix of the graph or the eigenvalue decomposition of the graph. By operating the eigenvectors in the frequency domain, the most important frequency components in the graph are identified. Among them, for each node, the frequency domain information related to it is extracted to obtain an embedding that reflects the global semantic features. After the sparse Fourier transform, the feature representation of each node in the frequency domain will be embedded in a new feature space. These frequency domain features reflect the global information of the graph structure and can capture the deep semantic relationship in the graph from a macro level. For example, in the library scenario, these features can represent the global connection between books, users, and queries, which helps to more accurately understand the user's query intention.

[0054] Step S34: Generate semantic representation based on frequency domain feature embedding, wherein the semantic representation is derived based on a graph decoding method based on variational inference to obtain semantic information of the user's intention.

[0055] It can be understood that in this step, the graph decoding method based on variational inference is used to derive semantic representation from frequency domain feature embedding. The graph decoding method generally involves the decoding process of the graph, that is, the potential semantics of each node in the graph is derived based on the feature information of the nodes and edges. This process infers the hidden information of the graph through variational inference, and finally obtains the semantic representation related to the user's query intention. In the process of graph decoding, variational inference is combined with frequency domain feature embedding (which has reflected the global semantic information of the graph) to decode the nodes, so as to obtain the specific semantics of each node in the context of user intention. This semantic representation is not only a local feature, but also combines the global structure and frequency domain information, which can more accurately capture the user's query goals and needs. After the process of variational inference and graph decoding, the system can obtain the semantic information of the user's intention. This semantic information is comprehensively generated based on user behavior, historical borrowing data, query patterns and frequency domain feature embedding, and has a high degree of personalization and deep semantic connotation.

[0056] Step S4, predicting the borrowing demand at a preset time point based on the semantic representation information of the user's intention, the user's borrowing record and a sparse Bayesian learning algorithm, and obtaining a demand prediction result for each book;

[0057] It can be understood that in this step, the system will integrate the user's intention semantic representation information (semantic representation obtained by graph decoding) and the user's borrowing record (such as historical borrowing frequency, borrowing time, etc.) as input features. Then use the sparse Bayesian learning algorithm to model the user's needs. Sparse Bayesian learning is an efficient probabilistic inference method that can find important features in high-dimensional data and model them. In this scenario, the Bayesian model generates a prediction model by calculating the borrowing probability of different books and combining the user's semantic intention and borrowing record. The sparsity feature enables the model to ignore irrelevant or noise information and only focus on the most influential features, thereby improving the prediction efficiency and accuracy. The application of the sparse Bayesian learning algorithm utilizes the sparse characteristics in the data, effectively screens out the most representative factors from a large amount of historical data, reduces the risk of overfitting, and improves the stability and accuracy of the prediction. In this step, step S4 includes step S41, step S42, step S43, and step S44.

[0058] Step S41: performing feature fusion processing according to the semantic representation information of user intention and the borrowing record, wherein the semantic representation information of user intention and the borrowing record are decomposed by non-negative matrix to generate a low-dimensional feature representation, and optimizing by minimizing the objective function to obtain a joint feature representation of user behavior;

[0059] It can be understood that the non-negative matrix decomposition in this step is an algorithm commonly used for dimensionality reduction and feature extraction. It decomposes a non-negative matrix (for example, the semantic representation information matrix of user intention and the borrowing record matrix) into two low-dimensional non-negative matrices. Through non-negative matrix decomposition, the system merges the semantic representation information of user intention (obtained through graph decoding) with the borrowing records (such as the number of historical borrowings, borrowing types, etc.). The decomposition results of the two matrices can extract the common potential patterns of user behavior and merge information from different data sources into a unified representation.

[0060] Step S42: perform time series modeling of future demand based on the joint feature representation of user behavior, predict demand at a preset time point in the future through Gaussian process regression, and obtain time series prediction features;

[0061] It can be understood that in this step, Gaussian process regression is used to model the time series of borrowing demand based on the joint feature representation of user behavior. These feature representations combine the user's historical borrowing records and semantic information, reflecting the changes in user interests at different time points. By taking these features as input, the Gaussian process regression model can effectively learn the trend and periodicity of demand changes. When performing time series modeling, the system first inputs the historical borrowing records and user behavior features as training data into the Gaussian process regression model. The core advantage of the Gaussian process is that it can flexibly handle nonlinearity and non-stationarity in the data, which is crucial for predicting changes in demand for books in the library. After the model training is completed, the Gaussian process regression can not only give the predicted value of borrowing demand at a certain point in the future, but also provide a confidence interval, that is, the uncertainty of the predicted value. This measure of uncertainty is particularly important for inventory management systems because it helps the system identify the reliability of the prediction results and then determine the inventory allocation strategy. The future demand predicted by Gaussian process regression is not just a simple point prediction, it also provides a confidence interval for each prediction point, allowing the system to better understand the range of future demand changes. Using these predicted values ​​and uncertainties as time series forecasting features can provide rich information for subsequent inventory optimization and scheduling, especially when facing demand fluctuations and uncertainties, so that more flexible response strategies can be adopted.

[0062] Step S43: establishing a demand forecasting model according to the time series forecasting characteristics, and optimizing the demand forecasting model weights through sparse Bayesian learning to obtain a sparse weight matrix of book borrowing demand;

[0063] It can be understood that this step establishes a demand forecasting model through time series prediction features, and uses sparse Bayesian learning to optimize the model weights to generate a sparse weight matrix that reflects the important features and priorities of book borrowing demand. In the demand forecasting model, by inputting time series features and related auxiliary information (such as historical borrowing data, user query trends, etc.), the dynamic trend of book demand is captured; combined with the advantages of sparse Bayesian learning, by introducing sparse priors to the model weights, a small number of important features are automatically screened for prediction, and the weight distribution is gradually optimized, thereby improving the accuracy and interpretability of the model. The sparse weight matrix finally obtained not only significantly reduces the computational complexity and the risk of model overfitting, but also improves the accuracy of feature selection and provides an intuitive feature importance assessment. This process provides an efficient decision-making basis for subsequent inventory optimization, while supporting real-time adjustment and personalized inventory allocation strategies, and realizing refined prediction and management of book demand.

[0064] Step S44: Send the sparse weight matrix of book borrowing demand to the demand prediction model to generate demand prediction results, and obtain the borrowing demand value of the book at a preset time point.

[0065] It can be understood that this step inputs the sparse weight matrix generated by sparse Bayesian learning into the demand forecasting model, and finally generates the specific borrowing demand value of the book at a preset time point. In this process, the model uses the structural characteristics of the sparse weight matrix to focus the prediction on books with significant demand characteristics, avoiding redundant processing of low-relevance or irrelevant data. Through the synergy of the sparse weight matrix and the demand forecasting model, the model can quickly focus on key features and generate efficient prediction results. At the same time, in order to improve the prediction accuracy, the model introduces multiple iterative solution processes to dynamically adjust the input sparse features to ensure that the generated demand value can truly reflect user needs and future trends. The borrowing demand value finally output has high credibility and refined characteristics, which provides a scientific basis for subsequent inventory management and allocation strategies, thereby optimizing library resource allocation and improving service quality and operational efficiency.

[0066] Step S5: construct an inventory allocation plan based on the demand forecast result, wherein a convex optimization method and a multi-objective programming algorithm are used to balance inventory allocation and logistics costs, and the inventory allocation plan is obtained through a Lagrangian relaxation method;

[0067] It can be understood that this step establishes a comprehensive optimization model of inventory allocation and logistics costs through a multi-objective programming algorithm. The model takes meeting user needs and optimizing logistics costs as its objective function, and sets a variety of constraints including storage capacity, delivery timeliness, and demand fluctuations. Then, the objective function is globally optimized using a convex optimization method to ensure that the model has unique solutions and rapid convergence. In order to further solve the optimization bottleneck that may be caused by complex constraints in multi-objective programming, the Lagrangian relaxation method is used to relax these complex constraints into the objective function, so that the optimization problem can be solved at a lower computational complexity. Through iterative optimization of the relaxed objective function, an inventory allocation plan that meets the actual constraints is generated. In this step, step S5 includes step S51, step S52, and step S53.

[0068] Step S51: perform inventory allocation and logistics cost modeling based on the demand forecast results, wherein the modeling is performed through a multi-objective programming algorithm to obtain the objective function and constraint conditions of the multi-objective programming model;

[0069] It can be understood that this step combines inventory demand forecast data and logistics dynamic information, and at the same time, a unified modeling of multiple mutually influencing objectives is performed through a multi-objective programming algorithm, laying the foundation for subsequent optimization solutions. The objective function in this step is as follows:

[0070] in, represents the minimization of the objective function value, and is the weight parameter, represents the inventory allocation balancing cost, Represents logistics cost.

[0071] The calculation formula of the inventory allocation balancing cost is as follows:

[0072] in, represents the inventory allocation balancing cost, Indicates the total warehouse volume, Indicates warehouses, Indicates Allocated inventory for warehouses, Indicates The demand forecast value corresponding to each warehouse.

[0073] Step S52, performing convex optimization processing on the objective function of the multi-objective programming model to obtain a convex optimization model;

[0074] It can be understood that this step convexifies the constraints. For example, capacity constraints or logistics capacity constraints are usually linear constraints, which themselves meet the requirements of convex optimization. However, for nonlinear constraints (such as path dependence or complex time window constraints), relaxation techniques or convex approximation methods can be used to make them meet the convex constraint form. The convexified objective function and constraints are integrated into a standard convex optimization problem, expressed as:

[0075] in, represents a convex objective function, represents the decision variable, represents a convex constraint function, each constraint must satisfy , represents the total number of constraints, Indicates constraints.

[0076] Step S53: Relax the complex constraints into the objective function based on the Lagrangian relaxation method, obtain the inventory allocation data and logistics cost data that meet the relaxed constraint conditions, and then generate an inventory allocation plan.

[0077] It can be understood that the Lagrangian relaxation method in this step is mainly used to transform the original optimization problem with complex constraints into a relatively simple optimization problem. By introducing Lagrangian multipliers, the constraints and the objective function are combined to form a Lagrangian function. Specifically, the constraints in the original problem will be relaxed, that is, they will be added to the objective function, but they will not be directly forced to be satisfied. Instead, these constraints will be "punished" through Lagrangian multipliers. In inventory management problems, you may encounter various constraints such as inventory capacity and logistics transportation capacity. After relaxing these constraints, construct a new objective function as described below:

[0078] in, represents the new objective function, represents the objective function of inventory allocation and logistics cost, represents the new Lagrange multiplier corresponding to each relaxed constraint, represents each relaxed constraint, Indicates After the relaxation of the constraints, Represents the total number of relaxed constraints.

[0079] It can be understood that after relaxing the constraints, the purpose of the Lagrangian relaxation method is to obtain an approximate optimal solution by solving the Lagrangian function. After Lagrangian relaxation, a solution that satisfies the relaxed constraints is obtained. In this way, the system can effectively balance inventory allocation and logistics costs. In the end, the obtained inventory allocation data and logistics cost data are the optimized inventory allocation plan. This plan reduces inventory and logistics costs as much as possible without completely enforcing constraints, while ensuring that most actual needs and constraints are met.

[0080] Step S6: Process the inventory allocation plan and the metadata of the library acquired in real time based on the dynamic system state estimation method, and adjust the inventory allocation plan in real time according to the processing results to obtain a final inventory management plan.

[0081] It can be understood that this step can continuously track and update the status of the inventory system through the dynamic system state estimation method, making the inventory allocation plan more accurate. This real-time adjustment mechanism effectively reduces the risk of excess or shortage of inventory, improves the efficiency and accuracy of inventory management, ensures that inventory can be flexibly adjusted according to changing demand and supply conditions, and avoids the lag and inaccuracy in traditional inventory management methods. At the same time, dynamic system state estimation can handle uncertainty, enhance the system's ability to predict future demand changes, and improve the intelligence level of the inventory management system. In this step, step S6 includes step S61, step S62 and step S63.

[0082] Step S61: establishing a dynamic state estimation model according to the library metadata and inventory allocation plan acquired in real time, wherein the dynamic state estimation model includes a state transition model and an observation model;

[0083] It is understandable that the state transition model describes the dynamic law of system state changes over time. In inventory management, the state transition model is mainly used to describe how state variables such as inventory quantity and borrowing demand change over time and according to external factors (such as user borrowing behavior, logistics distribution, inventory replenishment, etc.). For example, if the current inventory of a book is , then the inventory at the next moment is It may depend on the inventory at the previous moment, the borrowing demand of the book, and other factors (such as replenishment speed, transportation delay, etc.). The state transition model can capture this temporal dependency and provide predictions for the inventory status at future moments.

[0084] Among them, the state transition model is usually expressed as:

[0085] in, is the inventory at the next moment, It's time The inventory volume, It's time The control input, It's time The disturbance factor, It is the mapping function of state transition.

[0086] Among them, the observation model is usually expressed as:

[0087] in, Indicates time The observation data, Indicates time The actual inventory status or borrowing demand status, Indicates mapping the real state of the system to the actual observation value, which is a new observation model function. represents the relevant observation error parameter.

[0088] Step S62, iteratively updating the dynamic state estimation model based on Kalman filtering to obtain dynamically estimated book inventory and demand states;

[0089] Among them, Kalman filter is a recursive filtering algorithm that is suitable for state estimation problems in linear systems. In the context of inventory management, Kalman filter can effectively estimate the true state of the system from the observed data, especially when there is noise or uncertainty in inventory and demand. By continuously updating and correcting previous estimates, Kalman filter can make more accurate estimates of the inventory status and demand of books. In this step, the prediction formula is as follows:

[0090] in, It's time The predicted value of the state, It's time The predicted value of the state, is the state transition matrix, is the control matrix, It's time control input.

[0091] Among them, the update formula for combining the observed data with the predicted value and updating the estimated state is:

[0092] in, It's time The optimal estimate of the state, It's time The predicted value of the state, is the Kalman gain, Indicates time The observation data, is the observation model.

[0093] Step S63: adjust the inventory allocation plan in real time according to the dynamically estimated inventory and demand status, and verify whether the inventory allocation plan adjusted in real time meets the demand and warehouse capacity constraints. If so, use the inventory allocation plan adjusted in real time as the final inventory management plan.

[0094] It is understandable that the current inventory and demand status is obtained through the state estimation model. These data reflect the current inventory level and expected demand, and take into account various uncertain factors, such as borrowing trends, changes in user behavior, and logistics delivery time. At this time, the system adjusts the inventory according to the predicted demand data to ensure that each type of book can meet user needs in a timely manner while avoiding excessive inventory backlogs.

[0095] When adjusting the inventory allocation plan, the system will verify it through a series of constraints to ensure that the adjusted plan does not violate the following important restrictions:

[0096] Demand matching: The real-time adjusted inventory must match the actual or predicted demand. That is, the inventory of each book must be sufficient to support the borrowing demand in the current and future time periods.

[0097] Warehouse capacity limit: Considering the storage capacity of the warehouse, the adjusted inventory allocation plan needs to ensure that the total inventory does not exceed the carrying capacity of the warehouse. Warehouse capacity limit is an important constraint in inventory management, because exceeding the capacity limit may result in wasted space or inability to store other urgently needed books.

[0098] Logistics and delivery time: The inventory adjustment plan also needs to take into account the time window of logistics and delivery to ensure that books can be dispatched to the required location in the shortest time.

[0099] When the adjusted inventory allocation plan meets these needs and constraints, it will serve as the final inventory management plan.

[0100] Embodiment 2:

[0101] like Figure 2 As shown, this embodiment provides an inventory management system for a library, see Figure 2 The system includes an acquisition unit 701 , a first processing unit 702 , a second processing unit 703 , a third processing unit 704 , a fourth processing unit 705 and a fifth processing unit 706 .

[0102] An acquisition unit 701 is used to acquire metadata of a library, where the metadata of the library includes book information and user borrowing records;

[0103] The first processing unit 702 is used to construct a knowledge graph based on the library metadata and process the knowledge graph using a heterogeneous graph matching algorithm, wherein a target node set on the knowledge graph is obtained by extracting and analyzing the user query intention;

[0104] The second processing unit 703 is used to capture local semantic features on the knowledge graph based on multi-resolution wavelet transform and target node set, and extract feature embedding through frequency domain analysis to obtain semantic representation information of user intention;

[0105] The third processing unit 704 is used to predict the borrowing demand at a preset time point based on the semantic representation information of the user's intention, the user's borrowing record and the sparse Bayesian learning algorithm to obtain the demand prediction result of each book;

[0106] The fourth processing unit 705 is used to construct an inventory allocation plan according to the demand forecast result, wherein the inventory allocation and logistics cost are balanced by using a convex optimization method and a multi-objective programming algorithm, and the inventory allocation plan is obtained by a Lagrangian relaxation method;

[0107] The fifth processing unit 706 is used to process the inventory allocation plan and the metadata of the library acquired in real time based on the dynamic system state estimation method, and adjust the inventory allocation plan in real time according to the processing result to obtain a final inventory management plan.

[0108] It should be noted that, regarding the system in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0109] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0110] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A library inventory management method, characterized in that: include: Obtaining metadata of the library, wherein the metadata of the library includes book information and user's borrowing records; A knowledge graph is constructed based on the library metadata, and the knowledge graph is processed using a heterogeneous graph matching algorithm, wherein a target node set on the knowledge graph is obtained by extracting and analyzing the user query intention; Capturing local semantic features on the knowledge graph based on multi-resolution wavelet transform and target node set, and extracting feature embedding through frequency domain analysis to obtain semantic representation information of user intention; Based on the semantic representation information of the user's intention, the user's borrowing record and a sparse Bayesian learning algorithm, the borrowing demand at a preset time point is predicted to obtain a demand prediction result for each book; The inventory allocation plan is constructed based on the demand forecast results. The convex optimization method and multi-objective programming algorithm are used to balance the inventory allocation and logistics costs, and the inventory allocation plan is obtained through the Lagrangian relaxation method. The inventory allocation plan and the metadata of the library acquired in real time are processed based on a dynamic system state estimation method, and the inventory allocation plan is adjusted in real time according to the processing results to obtain a final inventory management plan.

2. The library inventory management method according to claim 1, characterized in that , constructing a knowledge graph based on the library metadata, and processing the knowledge graph using a heterogeneous graph matching algorithm, including: A heterogeneous graph structure is constructed based on the library metadata, and a sparse heterogeneous graph is generated based on a graph weighted downsampling screening method to obtain the initial skeleton of the heterogeneous graph. The implicit relationship mining algorithm based on random walk captures the potential connections between nodes of the initial skeleton of the heterogeneous graph, and expands the edge weights through path information to obtain an enhanced heterogeneous graph; Cluster the enhanced heterogeneous graph based on the node clustering algorithm, generate node categories based on the clustering results, and obtain the preliminary semantic partitioning in the graph; The path extraction algorithm based on multi-level graph search searches for paths in the graph through preliminary semantic partitioning to obtain the target node set of the user query.

3. The library inventory management method according to claim 1, characterized in that ,The ,capturing local semantic features on the knowledge graph based on the multi-resolution wavelet ,transform and the target node set and extracting the ,feature embedding through frequency domain analysis includes: Extract local subgraphs based on the knowledge graph and the target node set, where local subgraphs are captured based on a hierarchical clustering algorithm to obtain a local graph related to user intent; Multi-resolution wavelet transform processing is performed on the local graph related to user intention, in which local semantic features are captured by graph signal decomposition to obtain multi-scale feature representation; Perform frequency domain analysis based on multi-scale feature representation, wherein global semantic features are extracted based on sparse Fourier transform spectral embedding to obtain frequency domain feature embedding; Semantic representation is generated based on frequency domain feature embedding, where a graph decoding method based on variational inference is used to derive the semantic representation and obtain the semantic information of the user's intention.

4. The library inventory management method according to claim 1, characterized in that , based on the semantic representation information of the user's intention, the user's borrowing record and the sparse Bayesian learning algorithm, predicting the borrowing demand at a preset time point, including: Perform feature fusion processing based on the semantic representation information of user intention and borrowing records, wherein the semantic representation information of user intention and borrowing records are decomposed by non-negative matrix to generate low-dimensional feature representation, and then optimized by minimizing the objective function to obtain the joint feature representation of user behavior; Based on the joint feature representation of user behavior, the time series modeling of future demand is carried out, and the demand at a preset time point in the future is predicted through Gaussian process regression to obtain the time series prediction features; A demand forecasting model is established based on the time series forecasting characteristics, and the weights of the demand forecasting model are optimized through sparse Bayesian learning to obtain a sparse weight matrix of book borrowing demand; The sparse weight matrix of book borrowing demand is sent to the demand prediction model to generate demand prediction results, and the borrowing demand value of the book at a preset time point is obtained.

5. The library inventory management method according to claim 1, characterized in that ,According to the demand forecast results, the inventory allocation plan is constructed, in which the convex optimization method and multi-objective programming algorithm are used to balance the inventory allocation and logistics costs, and the inventory allocation plan is obtained through the Lagrangian relaxation method, including: Inventory allocation and logistics cost modeling are performed based on demand forecast results, where the modeling is performed using a multi-objective programming algorithm to obtain the objective function and constraints of the multi-objective programming model; The objective function of the multi-objective programming model is subjected to convex optimization processing to obtain a convex optimization model; Based on the Lagrangian relaxation method, the complex constraints are relaxed into the objective function to obtain the inventory allocation data and logistics cost data that meet the relaxed constraints, and then generate an inventory allocation plan.

6. A library inventory management system, characterized in that: include: An acquisition unit, used for acquiring metadata of a library, wherein the metadata of the library includes book information and borrowing records of users; A first processing unit is used to construct a knowledge graph based on the library metadata and process the knowledge graph using a heterogeneous graph matching algorithm, wherein a target node set on the knowledge graph is obtained by extracting and analyzing the user query intention; A second processing unit is used to capture local semantic features on the knowledge graph based on multi-resolution wavelet transform and target node set, and extract feature embedding through frequency domain analysis to obtain semantic representation information of user intention; A third processing unit is used to predict the borrowing demand at a preset time point based on the semantic representation information of the user's intention, the user's borrowing record and a sparse Bayesian learning algorithm to obtain a demand prediction result for each book; A fourth processing unit is used to construct an inventory allocation plan according to the demand forecast result, wherein a convex optimization method and a multi-objective programming algorithm are used to balance inventory allocation and logistics costs, and the inventory allocation plan is obtained through a Lagrangian relaxation method; The fifth processing unit is used to process the inventory allocation plan and the metadata of the library acquired in real time based on the dynamic system state estimation method, and adjust the inventory allocation plan in real time according to the processing result to obtain the final inventory management plan.

7. The library inventory management system according to claim 6, characterized in that: The first processing unit comprises: The first processing subunit is used to construct a heterogeneous graph structure according to the library metadata, and generate a sparse heterogeneous graph based on a graph weighted downsampling screening method to obtain an initial skeleton of the heterogeneous graph; The second processing sub-unit is used to capture the potential connections between nodes of the initial skeleton of the heterogeneous graph based on the implicit relationship mining algorithm of random walk, and expand the edge weights through the path information to obtain the enhanced heterogeneous graph; The third processing sub-unit is used to cluster the enhanced heterogeneous graph based on the node clustering algorithm, and generate node categories through the clustering results to obtain preliminary semantic partitions in the graph; The fourth processing sub-unit is used to perform a preliminary semantic partitioning search path in the graph based on a path extraction algorithm for multi-level graph search to obtain a user query target node set.

8. The library inventory management system according to claim 6, characterized in that: The second processing unit comprises: A fifth processing sub-unit is used to extract a local sub-graph according to the knowledge graph and the target node set, wherein the local sub-graph is captured based on a hierarchical clustering algorithm to obtain a local graph related to the user intention; a sixth processing subunit, configured to perform multi-resolution wavelet transform processing according to a local graph related to user intention, wherein local semantic features are captured by decomposing the graph signal to obtain a multi-scale feature representation; A first analysis subunit is used to perform frequency domain analysis based on the multi-scale feature representation, wherein global semantic features are extracted based on sparse Fourier transform spectrum embedding to obtain frequency domain feature embedding; The second analysis subunit is used to generate semantic representation based on frequency domain feature embedding, wherein the semantic representation is derived based on a graph decoding method based on variational inference to obtain semantic information of user intention.

9. The library inventory management system according to claim 6, characterized in that: The third processing unit comprises: The seventh processing subunit is used to perform feature fusion processing according to the semantic representation information of the user's intention and the borrowing record, wherein the semantic representation information of the user's intention and the borrowing record are decomposed by non-negative matrix to generate a low-dimensional feature representation, and the optimization is performed by minimizing the objective function to obtain a joint feature representation of the user behavior; The first prediction subunit is used to perform time series modeling of future demand based on the joint feature representation of user behavior, predict the demand at a preset time point in the future through Gaussian process regression, and obtain time series prediction features; The second prediction subunit is used to establish a demand prediction model according to the time series prediction characteristics, and optimize the weight of the demand prediction model through sparse Bayesian learning to obtain a sparse weight matrix of book borrowing demand; The third prediction subunit is used to send the sparse weight matrix of the book borrowing demand to the demand prediction model to generate the demand prediction result, and obtain the borrowing demand value of the book at a preset time point.

10. The library inventory management system according to claim 6, characterized in that: The fourth processing unit comprises: The first modeling subunit is used to perform inventory allocation and logistics cost modeling based on the demand forecast results, wherein the modeling is performed through a multi-objective programming algorithm to obtain the objective function and constraint conditions of the multi-objective programming model; The second modeling subunit is used to perform convex optimization on the objective function of the multi-objective planning model to obtain a convex optimization model; the third modeling subunit is used to relax the complex constraints into the objective function based on the Lagrangian relaxation method to obtain inventory allocation data and logistics cost data that meet the relaxed constraint conditions, and then generate an inventory allocation plan.

Images