An intelligent warehousing decision generation method and system, device, and storage medium

Abstract

The application provides a kind of intelligent warehousing decision generation method and system, equipment, storage medium, belong to intelligent warehousing technical field, this method includes: obtaining multimodal warehousing data, multimodal warehousing data includes inventory data, cargo attribute data, warehouse layout data and transport path data;According to data type, multimodal warehousing data is respectively extracted features, to obtain multiple warehousing features;Data type includes numerical type, category type and text type;Multiple warehousing features are mapped to uniform feature space, to obtain multiple target warehousing features, one target warehousing feature corresponds one modality warehousing data;Based on warehousing environment sensor data, multiple target warehousing features are attention weighted fusion, to obtain target feature vector;Based on target feature vector, generate warehousing management decision;Warehousing management decision includes warehousing layout strategy, inventory scheduling strategy and article storage location strategy.The application can comprehensively multimodal data to realize warehousing intelligent decision.

Description

Technical Field

[0001] This application belongs to the field of intelligent warehousing technology, and more specifically, relates to an intelligent warehousing decision generation method and system, equipment, and storage medium. Background Technology

[0002] With the rapid development of e-commerce and logistics, warehouse management faces increasingly complex challenges. Existing technologies struggle to meet the growing scale and dynamic demands of warehouses, resulting in low utilization efficiency of warehouse space and resources, high operating costs, and low management efficiency. In recent years, intelligent warehouse management methods based on machine learning and deep learning have gradually become a research hotspot, providing more accurate decision support for warehouse management. However, existing technologies still face problems such as insufficient information fusion and inadequate intelligent decision support, urgently requiring an intelligent decision-making method capable of optimizing warehouse layout, inventory scheduling, and item storage location in real time. Summary of the Invention

[0003] The purpose of this application is to provide an intelligent warehousing decision generation method, system, device, and storage medium to achieve intelligent warehousing decision-making by integrating multimodal data.

[0004] A first aspect of this application provides an intelligent warehousing decision generation method, comprising:

[0005] Acquire multimodal warehouse data, which includes inventory data, cargo attribute data, warehouse layout data, and transportation route data; extract features from the multimodal warehouse data according to the data type to obtain multiple warehouse features; the data types include numerical, categorical, and text.

[0006] Multiple warehouse features are mapped to a unified feature space to obtain multiple target warehouse features, with each target warehouse feature corresponding to a modal warehouse data. Attention-weighted fusion of multiple target warehouse features is then performed based on warehouse environment sensor data to obtain a target feature vector.

[0007] Warehouse management decisions are generated based on target feature vectors; warehouse management decisions include warehouse layout strategies, inventory scheduling strategies, and item storage location strategies.

[0008] A second aspect of this application provides an intelligent warehousing decision generation system, comprising:

[0009] The multimodal feature extraction module is used to acquire multimodal warehouse data, which includes inventory data, cargo attribute data, warehouse layout data, and transportation route data. Features are extracted from the multimodal warehouse data according to the data type, resulting in multiple warehouse features. Data types include numerical, categorical, and textual.

[0010] The feature attention fusion module is used to map multiple warehouse features to a unified feature space to obtain multiple target warehouse features, with each target warehouse feature corresponding to a modal warehouse data. Based on warehouse environment sensor data, attention weighted fusion is performed on multiple target warehouse features to obtain a target feature vector.

[0011] The decision generation module is used to generate warehouse management decisions based on target feature vectors; warehouse management decisions include warehouse layout strategies, inventory scheduling strategies, and item storage location strategies.

[0012] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the intelligent warehousing decision generation method described above.

[0013] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the intelligent warehouse decision generation method described above.

[0014] The beneficial effects of the intelligent warehousing decision generation method, system, device, and storage medium provided in this application are as follows: Traditional warehousing management methods typically rely on single-modal data and static rules, failing to effectively integrate heterogeneous information from different data sources and neglecting to fully consider the complex relationships between items, equipment, and paths within the warehouse, resulting in limited optimization effects. This application, by introducing an improved attention mechanism, can effectively integrate information from different data sources such as inventory data, cargo attributes, warehouse layout, and transportation routes, automatically learn the relationships between various modalities, and dynamically adjust the weights between modalities based on real-time environmental feedback. This adaptive adjustment method can automatically optimize decisions based on the latest data at every moment, avoiding the limitations of optimization under traditional static rules and significantly improving the accuracy and intelligence level of warehousing management decisions.

[0015] Meanwhile, this application combines the advantages of dynamic graph convolution and graph neural networks. Unlike the static graph structure of traditional graph neural networks, this application uses a dynamic graph convolution method, enabling the graph structure to update in real time as the warehouse environment changes, thus ensuring that the decision-making process can quickly respond to environmental changes. This dynamic graph convolution operation optimizes the relationships between nodes, accurately optimizing decisions such as warehouse layout, path scheduling, and inventory scheduling. It can respond to changes in warehouse demand in real time, automatically adjust decisions, avoid inventory surplus or shortage problems, effectively reduce operating costs, and improve management efficiency. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating an intelligent warehousing decision generation method provided in an embodiment of this application;

[0018] Figure 2 A structural block diagram of an intelligent warehousing decision generation system provided in an embodiment of this application;

[0019] Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0022] This application provides an intelligent warehouse decision generation method based on a multimodal cross-attention mechanism and graph neural networks. The method of this embodiment can be implemented based on a decision generation model. The decision generation model can be used to perform attention-weighted fusion of multiple target warehouse features based on warehouse environment sensor data to obtain a target feature vector; and to generate warehouse management decisions based on the target feature vector; the warehouse management decisions include warehouse layout strategies, inventory scheduling strategies, and item storage location strategies.

[0023] The decision generation model consists of multiple layers, primarily including a multimodal cross-attention mechanism, a dynamic graph convolutional neural network, and a decision generation layer. The multimodal cross-attention mechanism is used to perform attention-weighted fusion of multiple target warehouse features based on warehouse environment sensor data to obtain a target feature vector. The dynamic graph convolutional neural network and the decision generation layer work together to generate warehouse management decisions based on the target feature vector. The purpose of this application is to generate specific warehouse management optimization suggestions by integrating multimodal data and then optimizing warehouse management decisions based on graph neural networks.

[0024] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating an intelligent warehousing decision generation method according to an embodiment of this application. The method can be executed by an electronic device, and specifically, the method may include S101 to S103.

[0025] S101: Acquire multimodal warehousing data, which includes inventory data, cargo attribute data, warehouse layout data, and transportation route data; extract features from the multimodal warehousing data according to the data type to obtain multiple warehousing features; the data types include numerical, categorical, and textual.

[0026] In this embodiment, feature extraction is performed on the multimodal warehouse data according to the data type to obtain multiple warehouse features, including: for numerical warehouse data, the warehouse data is converted into a numerical feature vector; for categorical warehouse data, the warehouse data is one-hot encoded to obtain a binary feature representation, and the binary feature representation is used as the category feature vector corresponding to the warehouse data; for textual warehouse data, the warehouse data is converted into a text feature vector through a word embedding model.

[0027] In this embodiment, numerical warehousing data refers to directly quantifiable warehousing data, specifically including inventory quantity, inbound quantity, and outbound quantity in inventory data, volume and weight in cargo attributes, and transportation time in transportation routes. This numerical warehousing data can be directly converted into numerical feature vectors. Specifically, Z-score standardization can be used to eliminate dimensional differences, bringing different data indicators to the same scale for easier unified calculation in subsequent models. For example, the inventory quantity of a certain product (500) and volume (0.2 cubic meters) can be scaled proportionally to ensure the data distribution meets the model input requirements.

[0028] Categorical warehousing data is data with discrete categorical attributes, specifically including warehouse layout area divisions (such as area A and area B) and storage requirements (such as ambient temperature and refrigerated storage) in cargo attributes. This embodiment can perform one-hot encoding on each categorical type of warehousing data, such as representing "refrigerated" as [0, 1] and "ambient temperature" as [1, 0], forming a binary feature representation. This transforms non-numerical categorical information into a model-recognizable numerical form, using vector differences to reflect the discriminative power between categories, facilitating the model's learning of the impact of different categories on warehousing decisions.

[0029] Text-based warehousing data consists of descriptive text information, which may include descriptions of goods (e.g., "fragile items"), transportation route notes (e.g., "priority lane"), etc. This embodiment can employ a Word2Vec word embedding model to convert text-based warehousing data into fixed-length vectors, such as "fragile items" corresponding to a set of numerical vectors. The principle is to capture semantic relationships within the text through a pre-trained model; for example, "fragile items" and "handle with care" are close in vector space, enabling the model to understand the meaning behind the text and provide a basis for storage and transportation decisions.

[0030] Multiple warehouse features are key information extracted from various types of warehouse data. Numerical feature vectors represent the quantitative characteristics of the data, categorical feature vectors represent classification attributes, and textual feature vectors represent textual semantics. These three together constitute the features that the model can process, providing a foundation for subsequent multimodal data fusion and decision generation. Binary feature representation is a feature representation obtained after one-hot encoding of categorical warehouse data, composed of 0s and 1s, used to convert non-numerical categorical information into a numerical form that the model can recognize. For example, this embodiment can obtain warehouse management-related data from multiple sources, specifically including inventory data, goods attribute data, warehouse layout data, transportation route data, and environmental data. Regarding inventory data, this embodiment can collect information such as the inventory quantity, purchase records, outbound records, and historical sales data of each commodity in real time through a warehouse management system. Goods attribute data can include the volume, weight, and storage requirements (such as temperature and humidity) of the goods. Goods attribute data helps the model understand the storage requirements and transportation conditions of the goods. For warehouse layout data, this embodiment can collect spatial information of the warehouse, including various storage areas, shelf locations, and equipment distribution. Warehouse layout data ensures that the model can consider the physical space of the warehouse. For transportation route data, this embodiment can collect route information of equipment within the warehouse (such as automated guided vehicles, forklifts, etc.) and the transportation routes of materials. For environmental data, this embodiment can use environmental sensors pre-installed in the warehouse to monitor environmental conditions such as temperature, humidity, and vibration in real time, serving as a reference for dynamic adjustment decisions.

[0031] After data collection, this embodiment can perform cleaning, standardization, and formatting on the collected data. Specifically, this includes: cleaning the collected data, such as removing missing and outlier values ​​to ensure data quality; filling in missing inventory information using the mean or by inferring from other relevant information; and standardizing the cleaned data. For numerical data such as the volume and weight of items, this embodiment can use Z-score standardization to ensure all data are within the same dimension. The standardization formula is:

[0032]

[0033] Where X represents the raw data, such as inventory data, cargo attribute data, warehouse layout data, and transportation route data. The mean of the data. denoted as the standard deviation of the data.

[0034] The standardized data is formatted to transform heterogeneous data from different sources into a unified input format, facilitating subsequent weighted fusion processing via a multimodal cross-attention mechanism. Numerical data can be directly converted into a unified feature vector; categorical data requires one-hot encoding, transforming each categorical data point into a binary feature representation. Text data can be converted into a fixed-length vector representation using a Word2Vec pre-trained word embedding model.

[0035] For example, in addition to the methods described above, this embodiment can also perform separate feature extraction for each data modality in the following way: Let the input of each data be... (in (representing different data modes), passed through the encoder After processing, the result is Corresponding embedding feature representation The expression is as follows:

[0036]

[0037] The resulting embedded feature representations can be used as input for subsequent models.

[0038] S102: Map multiple warehouse features to a unified feature space to obtain multiple target warehouse features, with each target warehouse feature corresponding to a modal warehouse data; perform attention-weighted fusion of multiple target warehouse features based on warehouse environment sensor data to obtain a target feature vector.

[0039] In this embodiment, the unified feature space is a space of the same dimension to which different types of warehouse features are mapped, ensuring that all types of features can be processed uniformly by the model. After mapping multiple warehouse features to the unified feature space, the dimensional differences of heterogeneous features can be eliminated, allowing the model to compare the importance of different features in the same dimension.

[0040] Dimensionally aligned target warehouse features refer to feature vectors mapped to a unified space, retaining the core information of the original features while maintaining consistent dimensions. The process of mapping multiple warehouse features to a unified feature space can be achieved through linear transformations or embedding layers to adjust dimensions. For example, text vectors can be transformed from 300 dimensions to 256 dimensions to maintain consistency with numerical and categorical features, laying the foundation for subsequent fusion.

[0041] Warehouse environment sensor data can include real-time monitoring data such as temperature, humidity, and light intensity within the warehouse, used to dynamically adjust the weights of various modal features. This data reflects the impact of environmental changes on warehousing decisions; for example, in high-temperature environments, the weights of features related to temperature-sensitive goods need to be increased. Attention-weighted fusion refers to calculating the weights of each target warehousing feature based on the environmental sensor data, assigning high weights to important features, and then obtaining the target feature vector through weighted summation. This embodiment can first analyze the correlation between sensor data and features (such as temperature and refrigerated goods features), adjust the weights, and then fuse them, allowing the model to focus more on key information in the current environment and improve the targeting of decisions.

[0042] In this embodiment, various features are first mapped to the same-dimensional space through an embedding layer; then, the weights of each feature are calculated based on the data from warehouse environment sensors, such as increasing the weight of refrigerated goods attribute features when the temperature is high; finally, the features of each modality are fused according to the weights to generate a target feature vector that comprehensively reflects multimodal information.

[0043] For example, this embodiment fuses information from different modalities and transforms it into a unified feature vector for subsequent model processing. Feature fusion ensures the effective combination of different data modalities for use in subsequent model training. Specifically, this embodiment can employ alignment techniques to ensure spatiotemporal consistency among features from multiple modalities. Features from different modalities are mapped to a unified feature space. Then, this embodiment uses a cross-attention mechanism to weight and fuse the feature vectors from different sources, forming a high-dimensional feature vector. For example, information such as inventory quantity, cargo volume, and warehouse location can be concatenated into a single feature vector (i.e., the target feature vector).

[0044] For example, the core of the cross-attention mechanism lies in effectively modeling the relationships between different modalities. Assume that the features of each modality are... ,in This represents the number of modalities. The cross-attention mechanism can weightedly fuse these features by calculating the relationships between queries, keys, and values, as shown in the following formula:

[0045]

[0046] in, As input to the attention module, It is the output of the attention module. Q is the transpose of the mapping result obtained by mapping the input of the attention module based on the key matrix parameters; Q is the mapping result obtained by mapping the input of the attention module based on the query matrix parameters; V is the mapping result obtained by mapping the input of the attention module based on the value matrix parameters. This refers to the input data dimension. The self-attention processing here is implemented through the attention module in the transformer. For attention mechanisms, is the activation function. Through this multimodal feature fusion approach incorporating cross-attention, the model can adaptively fuse information from different modalities, ensuring that each modality contributes more reasonably to the decision.

[0047] However, in practical applications, the multimodal data in warehouse management systems is dynamically changing, with frequent fluctuations in warehouse items, environment, and demand. Therefore, this embodiment introduces an adaptive cross-attention mechanism for dynamic adjustment, enabling the model to dynamically adjust the weights between modalities based on current input features and environmental feedback each time data is input. Furthermore, to respond to changes in the warehouse environment, this embodiment designs a dynamic weighted adjustment mechanism, using warehouse environmental sensor data to control the dynamic updating of different modal weights.

[0048] S103: Generate warehouse management decisions based on target feature vectors; warehouse management decisions include warehouse layout strategies, inventory scheduling strategies, and item storage location strategies.

[0049] In this embodiment, the decision generation model includes a dynamic graph convolutional neural network; the warehouse management decision is determined based on the target feature vector and the dynamic graph convolutional neural network in the following way: the adjacency matrix corresponding to the initial graph structure in the dynamic graph convolutional neural network is updated based on the target feature vector to obtain the updated target adjacency matrix; the initial graph structure is a graph structure modeled based on the initial items, initial equipment, and initial transportation paths in the warehouse, where the nodes in the initial graph structure represent the initial items and initial equipment, and the edges in the initial graph structure represent the initial transportation paths between nodes; the target adjacency matrix includes node features corresponding to multiple nodes;

[0050] Multiple graph convolution operations are performed on the target adjacency matrix. In each graph convolution operation, the updated node features are obtained through the graph convolution update formula of that operation.

[0051] Warehouse management decisions are generated based on the updated node characteristics;

[0052] Among them, the The graph convolution update formula for the nth graph convolution operation is:

[0053]

[0054] in, For the nodes in the target adjacency matrix After the first The node features obtained after the layer graph convolution operation are updated, i.e., the node... In the Layer node table; For nodes in the target adjacency matrix One of the neighboring nodes, For the nodes in the target adjacency matrix In the Layer node characteristics, For the nodes in the target adjacency matrix The set of neighboring nodes, For the first The weight matrix of the layer, For activation function, For the first The bias term of the layer. The bias term, working in conjunction with the adjacency matrix update, enables the graph neural network to dynamically propagate node features at each time step t through the interaction between graph convolution and the adjacency matrix. Specifically, as the warehouse state changes, the adjacency matrix update reflects the changes in the relationships between nodes, and the bias term helps the network make rapid non-linear adjustments to these changes. For example, when a device malfunctions, related edges are removed, causing a change in the adjacency matrix; the bias term can then adjust the non-linear relationships of node features, ensuring the accurate propagation of node features in the new environment.

[0055] For example, in When = 0, this embodiment performs the first graph convolution operation on the target adjacency matrix, where the first... Graph convolution operations The graph convolution update formula is:

[0056]

[0057] in, For the nodes in the target adjacency matrix The node features are updated after the first layer of graph convolution operation. For the nodes in the target adjacency matrix The initial node characteristics, For the first The weight matrix of the layer, This is the bias term for level 1.

[0058] In this embodiment, a dynamic graph convolutional neural network (DCNN) refers to a neural network model that can update the graph structure in real time with warehouse environment data. The DCNN can optimize decisions by dynamically adjusting the relationships between nodes and edges. The parameters of the DCNN can include the number of nodes (i.e., the total number of items and equipment), the edge connection rules (i.e., transportation paths), and the number of graph convolutional layers. By modeling warehouse elements as a graph structure, updating graph relationships through real-time data, and then using graph convolution to propagate features, the DCNN can acquire the dynamic dependencies between items, equipment, and transportation paths. Considering that traditional static graphs cannot cope with changes such as item movement and equipment failure, the dynamic graph convolutional neural network set in this embodiment can improve the timeliness of decision-making.

[0059] The target feature vector, as input to a dynamic graph convolutional neural network, provides data for graph structure updates and feature propagation. Warehouse management decisions refer to the optimized solutions output by the model, which can include: warehouse layout strategies (involving parameters such as shelf adjustments and area functional divisions, e.g., placing high-frequency storage areas near exits); inventory scheduling strategies (including replenishment quantities and allocation routes, e.g., adjusting replenishment cycles based on sales forecasts); and item storage location strategies (including specific shelf numbers and shelf levels, e.g., placing heavy items on lower shelves). An adjacency matrix records the connection relationships between nodes in a graph. Matrix operations allow for rapid updates to the graph structure; when node relationships change, only the element values ​​at corresponding positions need to be modified. Graph convolution operations refer to the operation of updating the current node's features by aggregating the features of neighboring nodes. Each node's new feature is obtained by a weighted sum of its own features and neighboring features, with weights related to the distance / importance between nodes. Aggregating neighbor information allows nodes to perceive global relationships; for example, storage decisions for a certain item need to consider surrounding equipment and paths.

[0060] In this embodiment, the parameters of the warehouse layout strategy may include adjustments to area functions (e.g., changing area A to a high-frequency access area), optimization of shelf spacing, etc. The warehouse layout strategy is formulated based on adjusting the layout according to space utilization and path efficiency to reduce transportation distance. The parameters of the inventory scheduling strategy may include replenishment quantity, allocation priority, and safety stock threshold, etc. The inventory scheduling strategy is formulated based on combining sales forecasts and current inventory to balance supply and demand and avoid stockpiling or shortages. The parameters of the item storage location strategy may include shelf number, shelf level (e.g., placing heavy items on lower shelves), and adjacent item combinations (e.g., storing highly related items nearby). The item storage location strategy is formulated based on matching the optimal location according to item attributes (weight, access frequency) to improve efficiency.

[0061] In this embodiment, the dynamic graph convolutional neural network captures the dynamic relationships between warehouse elements through a real-time updated graph structure, uses graph convolution operations to propagate node features, learns the dependencies between elements, and finally outputs warehouse management decisions that meet the goals of space utilization and path efficiency.

[0062] For example, this embodiment can define node types including item nodes and device nodes. This embodiment can define edge rules including direct connections between items and storage areas, and connections between devices and reachable items / areas. This embodiment transforms complex storage relationships into a computable graph structure, facilitating model learning of dependencies between elements. This embodiment can use deployed sensors to collect real-time data on item location (two-dimensional coordinates), device status (available / faulty), and changes in transportation paths. When an item moves, this embodiment can delete the edge between that node and its original neighboring nodes and add an edge between that node and a node at the new location; when a device fails, this embodiment can set the edge weight of the edge associated with that device to 0.

[0063] This embodiment initializes the adjacency matrix based on the initial layout of the graph structure: the adjacency matrix records the connection relationships between nodes, with a value of 1 for a connection and 0 for no connection. For example, if item A is in region X, the corresponding position in the matrix is ​​1. When an item is moved, this embodiment can find the indices of the original and new positions and modify the values ​​of the corresponding rows and columns in the matrix. When a device malfunctions, this embodiment can set all elements of its corresponding row / column to 0.

[0064] The graph convolution operation specifically includes: for each node, this embodiment can collect the features of all nodes connected to that node and perform a weighted summation according to the edge weights. This embodiment can use a weight matrix to perform a linear transformation on the aggregated features, and then process them through an activation function. This embodiment can gradually extract higher-order features through multi-layer operations, such as from single item attributes to regional layout patterns.

[0065] This embodiment feeds the node features after graph convolution into the decision layer, which then outputs specific suggestions, such as "move item B to region Y," through a fully connected network. This embodiment can also filter infeasible solutions by incorporating warehouse physical constraints (such as shelf load-bearing capacity) and finally output executable instructions.

[0066] As can be seen from the above, traditional warehouse management methods typically rely on single-modal data and static rules, failing to effectively integrate heterogeneous information from different data sources and neglecting to fully consider the complex relationships between items, equipment, and paths within the warehouse, resulting in limited optimization effects. This application's embodiment, by introducing an improved attention mechanism, can effectively integrate information from different data sources such as inventory data, cargo attributes, warehouse layout, and transportation routes, automatically learn the relationships between various modalities, and dynamically adjust the weights between modalities based on real-time environmental feedback. This adaptive adjustment method of this application embodiment can automatically optimize decisions based on the latest data at every moment, avoiding the limitations of optimization under traditional static rules, and significantly improving the accuracy and intelligence level of warehouse management decisions.

[0067] This embodiment combines the advantages of dynamic graph convolution and graph neural networks. Unlike the static graph structure of traditional graph neural networks, this embodiment uses dynamic graph convolution, enabling the graph structure to update in real time as the warehouse environment changes. This ensures that the decision-making process can quickly respond to environmental changes. This dynamic graph convolution operation optimizes the relationships between nodes, precisely optimizing decisions such as warehouse layout, path scheduling, and inventory scheduling. It can respond to changes in warehouse demand in real time, automatically adjust decisions, avoid inventory surpluses or shortages, effectively reduce operating costs, and improve management efficiency.

[0068] In summary, this embodiment not only automatically generates optimized suggestions for warehouse layout, inventory scheduling, and item storage locations, significantly reducing warehouse space waste and improving item access efficiency, but also responds in real-time to changes in warehouse demand. Compared to traditional methods, this embodiment provides dynamic adjustment and optimization solutions, automatically adjusting decisions based on real-time changes in the warehouse environment and demand, ensuring the flexibility and accuracy of warehouse management, thereby further reducing operating costs and improving overall management efficiency. The method of this embodiment has high operability and practical value, and is particularly suitable for the management and operation of large-scale, complex warehouses, with broad application prospects.

[0069] In one embodiment of this application, a target storage feature is a modal feature; the storage environment sensor data includes storage environment temperature data; attention-weighted fusion of multiple target storage features based on the storage environment sensor data is performed to obtain a target feature vector, including: calculating an environmental impact factor based on the storage environment temperature data; calculating the weight corresponding to each modal feature based on the environmental impact factor using a weight calculation formula; and performing attention-weighted fusion of multiple modal features based on the weights corresponding to each modal feature to obtain a target feature vector; wherein, the weight calculation formula is:

[0070]

[0071] in, It is the weight of the i-th modal feature at time t. It is the preset adjustment parameter for the i-th modal feature, f It is the environmental influencing factor at time t. It is the temperature data of the storage environment at time t.

[0072] In this embodiment, the environmental impact factor is calculated based on the warehouse environment temperature data. Specifically, this includes: calculating the environmental impact factor using the environmental impact factor calculation formula based on the warehouse environment temperature data; wherein, the environmental impact factor calculation formula is:

[0073]

[0074] in, It is the adjustment coefficient. It is the preset target temperature.

[0075] In this embodiment, the environmental impact factor is an indicator that quantifies the importance of environmental data to warehousing characteristics. For example, for cold chain goods, the greater the temperature deviation, the higher the environmental impact factor. The environmental impact factor can convert environmental changes into calculable values, which can be used to adjust the weights of different warehousing characteristics. Considering that different environmental factors have varying degrees of influence on warehousing decisions, this embodiment uses environmental impact factor quantification to achieve accurate weighting of multimodal characteristics. The weight calculation formula is a mathematical formula used to calculate the weights of each warehousing characteristic based on the environmental impact factor.

[0076] In this embodiment, the weight of the target warehousing feature refers to the importance coefficient of each warehousing feature (such as inventory and goods attributes) during fusion, specifically including the real-time weight value of each modal feature. The weight calculation formula maps the influencing factors to weight values ​​between 0 and 1, ensuring that the weights are smoothly adjusted as the environment changes. For example, the weight of the inventory feature of temperature-sensitive items will approach 1 as the temperature deviation increases. The weight calculation formula in this embodiment avoids sudden weight changes and ensures decision stability. Attention-weighted fusion is the process of generating a comprehensive feature vector by combining the weights of each feature. Weighted fusion allows the model to capture key information. The weight calculation formula in this embodiment allows features that are more sensitive to the current environment to receive higher weights, such as increasing the weight of the attribute features of temperature-sensitive items when temperatures are high. Dynamic weight adjustment ensures that the fused feature vector better fits the needs of the current environment and improves the relevance of the decision.

[0077] For example, in order to respond to changes in the warehouse environment, this embodiment uses sensor data to control the dynamic updating of different modal weights, specifically including:

[0078] This embodiment can acquire multimodal data related to the warehouse environment in real time through environmental sensors. Specifically, this data can include temperature (T), humidity (H), and illumination (L), which can be represented as E={T, H, L}. The environmental impact factor calculation formula is for temperature; similarly, the calculation methods for humidity and illumination are similar to the environmental impact factor calculation formula.

[0079] In warehouse management, environmental changes can significantly impact certain environmentally sensitive items, especially temperature-sensitive ones. This embodiment dynamically adjusts the weight of each modality based on environmental changes. The weight calculation formula in this embodiment quantifies the influence of environmental factors on modal importance through environmental impact factors, mapping it to a smooth, bounded weight value. This ensures that weight adjustments respond to environmental changes while maintaining numerical stability and avoiding drastic weight fluctuations.

[0080] For each mode, environmental data can be used to calculate the environmental impact factor for that mode. Taking temperature-sensitive items as an example, the adjustment coefficient in the environmental impact factor calculation formula... Determines temperature deviation The scaling factor (or sensitivity) to the influence of modal weights on inventory data. Specifically, The larger the value, the greater the impact (increase or decrease) of the same temperature deviation on the modal weights of the inventory data, meaning the model is more sensitive to temperature changes; The smaller the value, the smaller the impact of the same temperature deviation on the modal weights of the inventory data; that is, the model's response to temperature changes is relatively mild. The result of the environmental impact factor calculation formula directly affects the modal weights. This increases the weight of inventory data modalities related to temperature-sensitive items in the warehouse under high or low temperature conditions.

[0081] The dynamic weight adjustment in this embodiment enables the model to focus on environmentally sensitive features in real time, significantly enhancing the adaptability of decisions to the current environment and avoiding the problem of lagging response to environmental changes in traditional static fusion. The sigmoid function ensures smooth weight changes, avoiding decision fluctuations caused by abrupt changes and guaranteeing the stability of warehousing operations. The adjustment coefficient can be flexibly adapted to the sensitivity of different items, improving the universality of the method. The target feature vector finally generated in this embodiment can accurately reflect the correlation between the environment and warehousing features, providing high-quality input for subsequent graph neural network decision-making, improving inventory accuracy, space utilization, and item access efficiency, and reducing the risk of decision errors caused by environmental factors.

[0082] In one embodiment of this application, warehouse management decisions are generated based on multiple target warehouse characteristics and through a decision generation model;

[0083] The decision generation model is trained in the following way:

[0084] Multiple training samples and their corresponding annotation results are obtained. Each training sample includes multiple sample storage features. The sample storage features are obtained by extracting features from the multimodal storage data samples based on the data type.

[0085] Based on each training sample and through the initial model, corresponding predictive warehouse management decisions are obtained. These predictive warehouse management decisions include: numerical predictive decisions and categorical predictive decisions.

[0086] Based on each predicted warehouse management decision and its corresponding labeling results, the first loss is determined;

[0087] Based on the labeling results of each numerical prediction decision and its corresponding numerical warehouse management decision, the second loss is determined;

[0088] Based on the labeling results of each type of predictive decision and its corresponding type of warehouse management decision, the third loss is determined;

[0089] The initial model is trained based on the first loss, the second loss, and the third loss to obtain the decision generation model.

[0090] In this embodiment, the predicted warehouse management decisions include: warehouse layout adjustment prediction results, inventory scheduling prediction results, and storage location recommendation prediction results. The warehouse management decision labeling results include: warehouse layout adjustment labeling results, inventory scheduling labeling results, and storage location recommendation labeling results. Based on each predicted warehouse management decision and its corresponding labeling results, the first loss is determined, including: based on each warehouse layout adjustment prediction result and its corresponding labeling results, the first sub-loss is determined.

[0091] The second sub-loss is determined based on the inventory scheduling forecast results and their corresponding inventory scheduling labeling results;

[0092] The third sub-loss is determined based on the prediction results of each storage location and the corresponding storage location recommendation annotation results.

[0093] The first loss is determined based on the first sub-loss, the second sub-loss, and the third sub-loss.

[0094] In this embodiment, the first loss can be obtained by the following formula:

[0095]

[0096] in, This is the multi-target loss value, also known as the first loss; , and All are weighting coefficients. This refers to the loss in warehouse space utilization, also known as the first sub-loss. The loss is the cost of optimizing the transportation route, also known as the second sub-loss. This is the inventory scheduling loss, also known as the third sub-loss.

[0097] In this embodiment, the second loss can be obtained by the following formula:

[0098]

[0099] in, This is the numerical prediction decision corresponding to the i-th training sample output by the model. This represents the numerical warehouse management decision labeling result corresponding to the i-th training sample. The number of training samples for numerical decision-making;

[0100] In this embodiment, the third loss can be obtained by the following formula:

[0101]

[0102] in, The model outputs a categorical prediction decision for the j-th training sample. This represents the categorical warehouse management decision labeling result corresponding to the j-th training sample, where M is the number of training samples for numerical decisions.

[0103] In this embodiment, the initial model is trained based on the first loss, the second loss, and the third loss to obtain a decision generation model. Specifically, this includes: obtaining the total loss based on the first loss, the second loss, and the third loss, and training the initial model based on the total loss.

[0104] In this embodiment, the total loss is obtained using the following formula:

[0105]

[0106] in, Indicates the total loss. These are the weighting coefficients. .

[0107] In this embodiment, warehouse management decisions refer to optimization schemes generated based on multi-dimensional warehouse characteristics, including warehouse layout adjustments (such as regional functional divisions), inventory scheduling (such as replenishment quantities), and storage location recommendations (such as shelf numbers). The decision generation model is a prediction model that is trained and optimized. Training depends on training samples (including sample warehouse characteristics such as inventory and product attributes) and labeled results. During training, the initial model outputs predicted warehouse management decisions, which can be divided into numerical prediction decisions (such as quantitative results like replenishment quantities and access times) and categorical prediction decisions (such as classification results like regional divisions and path types).

[0108] In this embodiment, the total loss is the overall optimization objective that needs to be minimized during the training of the decision generation model, encompassing the core optimization objectives of warehouse management decisions. The first loss is the prediction error for different decision categories, consisting of three sub-losses corresponding to warehouse layout adjustment, inventory scheduling, and storage location recommendation. Specifically, the first sub-loss refers to the matching degree between the layout adjustment prediction and the label, such as the area adjustment error rate; the second sub-loss refers to the numerical deviation of inventory scheduling, such as the replenishment quantity error; and the third sub-loss refers to the category deviation of storage location recommendation, such as the shelf number error rate. The second loss is mainly used to calculate the error of numerical decisions, such as the difference between the predicted replenishment quantity and the actual replenishment quantity; the third loss is mainly used to calculate the error of categorical decisions, such as the error rate of area classification. Space utilization loss measures the efficiency of warehouse space utilization, aiming to maximize space utilization. Path optimization loss measures the degree of optimization of material handling paths, aiming to minimize transportation costs and time. Inventory scheduling loss measures the rationality of inventory management, aiming to avoid inventory overstocking or shortages.

[0109] For example, when the core bottleneck of the warehouse is space shortage, this embodiment can be appropriately increased. (Space utilization weight) is used to prioritize optimizing spatial layout; when low transportation efficiency or excessively high route costs become the main problems, this embodiment can be appropriately increased. (Path optimization weight) is used to focus on optimizing the transportation path; when inventory management problems are prominent (such as frequent stockouts or high inventory backlog), this embodiment can be appropriately increased. (Inventory scheduling weights) are used to enhance the optimization of inventory control. During model training, this embodiment can also employ the gradient descent algorithm (Adam optimizer) to gradually adjust model parameters. The Adam optimizer can automatically adjust the learning rate, making the training process more stable and efficient. The update formula for the Adam optimizer is:

[0110]

[0111] in, For the first The updated training parameters during the iteration cycle For the first The updated training parameters within the iteration cycle, i.e., the [number]th iteration... The parameters to be trained before updating during the iteration cycle; For learning rate, For first-order moment estimation, For second-order moment estimation, It is a constant. This is used to prevent division by zero errors. The model learning rate can be set to 0.001, the number of graph neural network layers can be set to 3, the batch size can be set to 64, and the number of training epochs can be set to 100. During training, the model training is completed when the validation set loss no longer decreases over several consecutive training epochs.

[0112] After training and optimization, the core task of the model is to generate specific decisions regarding warehouse management. Based on features generated by multimodal data fusion and graph neural networks, the model can provide comprehensive decision-making suggestions for warehouse management, optimizing aspects such as warehouse layout, inventory scheduling, and item storage locations.

[0113] For example, during training, the graph neural network learns the complex relationships between various items, equipment, and transportation routes in the warehouse. The core task of the decision generation stage is to generate decisions about warehouse management based on the node features obtained after graph convolution, combined with various optimization objectives. These decisions mainly include the following three categories: Warehouse layout strategy: The model automatically generates warehouse layout adjustment schemes based on the output of the graph neural network (information such as item storage location and equipment distribution) through the decision generation layer. This scheme considers not only the effective use of space but also factors such as the frequency of goods storage and retrieval and the correlation between items. Output example: Goods A should be transferred from area X to area Y to improve space utilization and reduce the transportation distance of the goods.

[0114] Inventory scheduling strategy: Combining real-time inventory information and demand forecasts, the model can provide dynamic inventory scheduling suggestions. Through a multimodal cross-attention mechanism, the model can automatically adjust inventory allocation to ensure timely replenishment at the lowest cost. Output example: Demand for product B is expected to increase by 20% in the next 24 hours. It is recommended to immediately replenish 300 units and store them near the shipping area.

[0115] Item storage location optimization strategy: The model recommends the optimal storage location based on the characteristics of the items and the layout of the warehouse space. This process improves storage efficiency and reduces retrieval time by combining the physical space of the warehouse with the characteristics of the items. Output example: Item C should be stored in the bottom layer of Zone 5 to maximize space utilization and reduce retrieval time.

[0116] After generating warehouse management decisions, the model can output these decisions as clear and easily executable text instructions. These instructions can be directly used in warehouse management decision support systems to guide warehouse operations. The output decisions can include specific operational instructions, timelines, and spatial adjustment plans, ensuring that managers can execute them quickly. For example, layout adjustment suggestions include specific area numbers and adjustment operations, while inventory scheduling suggestions specify replenishment quantities and storage locations. The output can include not only specific instructions but also execution priorities, affected warehouse areas, and time requirements. For example: Product A should be transferred to Zone 3 within 24 hours, prioritizing areas with lower inventory levels.

[0117] For example, in the decision generation phase, the model continuously optimizes its output quality through the training process and dynamically adjusts its parameters. Specifically, the process of dynamically adjusting parameters involves updating the model's weights and learning rate based on feedback from the current training data in each training epoch to improve the accuracy and robustness of decision generation.

[0118] During model training, regression loss functions (i.e., the loss function corresponding to the second loss) and classification loss functions (i.e., the loss function corresponding to the third loss) can be used to improve model accuracy. When decision-making involves numerical data (such as inventory scheduling quantity, replenishment quantity, etc.), regression loss functions can be used to measure the difference between numerical predictions and actual values. When decision-making involves categories (such as item storage area, route selection, etc.), classification loss functions can be used to measure the accuracy of categorical predictions.

[0119] For example, in the implementation of this embodiment, the model's parameters can be dynamically adjusted to optimize model performance in real time based on feedback during training. First, the learning rate is dynamically adjusted according to the model's training effect. For instance, when the loss function of the validation set does not show a significant decrease over multiple epochs, the learning rate is automatically reduced to avoid overfitting during training and improve convergence speed. Specifically, when the validation set loss does not show significant improvement over several consecutive epochs, the learning rate will be reduced to a preset proportional threshold, thereby accelerating fine-tuning in the later stages of training. The batch size parameter can also be dynamically adjusted during training based on training progress. Initially, a smaller batch size is used to increase the model update frequency, and the batch size is gradually increased as training progresses to improve training efficiency and ensure model stability. In addition, the model's hyperparameters can also be adjusted based on performance feedback, such as the number of layers in the graph neural network, the number of nodes per layer, and the weights of the loss function. These hyperparameters can be adjusted through automated hyperparameter optimization methods to ensure balanced optimization of different objectives. Finally, the model training can employ an early stopping strategy. The training process automatically stops when the validation set loss does not significantly decrease over multiple consecutive epochs, avoiding unnecessary computation and ensuring the model's generalization ability. By dynamically adjusting the strategy, the model can adaptively optimize various parameters throughout the training process, thereby continuously improving the accuracy and robustness of decision generation, ensuring that the final generated warehouse management decisions are the most efficient and executable.

[0120] Through the above steps, this embodiment provides a warehouse management decision generation method based on a multimodal cross-attention mechanism and graph neural network. It can comprehensively consider various types of warehouse data and generate specific optimization suggestions for warehouse layout, inventory scheduling, and item storage location, significantly improving the efficiency and space utilization of warehouse management.

[0121] For example, this embodiment collects actual data from a large warehousing and logistics center, including inventory data, cargo attribute data, warehouse layout data, transportation route data, and environmental sensor data. The dataset includes detailed information on 10,000 items, with attributes such as volume, weight, and storage requirements accurately recorded for each item. It also includes warehouse spatial layout information and real-time transportation route data. This embodiment cleans, standardizes, and formats the raw data. Missing values ​​in the inventory data are filled with the mean, and cargo attribute data is standardized using Z-scores to ensure that all numerical data are on the same scale. Warehouse layout data and item storage locations are encoded and converted into a unified format for subsequent model processing.

[0122] This embodiment employs a hybrid model based on a multimodal cross-attention mechanism and a graph neural network. Each data modality is transformed into a low-dimensional embedding through an independent encoder and then fused using the cross-attention mechanism. The fused features are input into the graph neural network for global optimization to generate warehouse management decisions. This embodiment uses the Adam optimizer for model training, with a learning rate of 0.001, a batch size of 64, and 100 training epochs. The model loss function (i.e., the loss function corresponding to the first loss) includes space utilization loss, path optimization loss, and inventory scheduling loss, with the loss function weights set to... .

[0123] The model generates warehouse management decisions based on the training results, including warehouse layout adjustments, inventory scheduling schemes, and item storage location recommendations. For example, the model suggests moving high-demand item A from region X to region Y to improve space utilization and reduce item retrieval time.

[0124] This embodiment integrates multi-source data such as inventory and goods attributes through a multimodal cross-attention mechanism, and combines it with a dynamic graph neural network to model warehouse elements, thereby achieving intelligent generation of warehousing decisions. Its core is a dynamic adjustment mechanism, which can optimize feature weights in real time based on environmental sensor data and update the relationships between warehouse elements through dynamic graph convolution. Compared with traditional methods, the method in this embodiment can significantly improve space utilization, inventory scheduling accuracy, and item storage and retrieval efficiency, while also improving the decision generation response speed.

[0125] Corresponding to the intelligent warehousing decision generation method in the above embodiment, Figure 2 This is a structural block diagram of an intelligent warehousing decision generation system according to an embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 2 The intelligent warehousing decision generation system 20 includes: a multimodal feature extraction module 21, a feature attention fusion module 22, and a decision generation module 23.

[0126] The multimodal feature extraction module 21 is used to acquire multimodal warehousing data, which includes inventory data, cargo attribute data, warehouse layout data, and transportation route data. Features are extracted from the multimodal warehousing data according to the data type to obtain multiple warehousing features. The data types include numerical, categorical, and textual types.

[0127] The feature attention fusion module 22 is used to map multiple warehouse features to a unified feature space to obtain multiple target warehouse features, one target warehouse feature corresponds to one modal warehouse data; based on warehouse environment sensor data, attention weighted fusion of multiple target warehouse features is performed to obtain target feature vectors;

[0128] The decision generation module 23 is used to generate warehouse management decisions based on the target feature vector; the warehouse management decisions include warehouse layout strategy, inventory scheduling strategy and item storage location strategy.

[0129] In one embodiment of this application, the warehouse features include numerical feature vectors, category feature vectors, and text feature vectors; the multimodal feature extraction module 21 is specifically used to convert numerical warehouse data into numerical feature vectors; for category warehouse data, one-hot encoding is performed on the warehouse data to obtain a binary feature representation, and the binary feature representation is used as the category feature vector corresponding to the warehouse data; for text warehouse data, the warehouse data is converted into text feature vectors through a word embedding model.

[0130] In one embodiment of this application, a target storage feature is a modal feature; the storage environment sensor data includes storage environment temperature data; the feature attention fusion module 22 is specifically used to calculate an environmental impact factor based on the storage environment temperature data; based on the environmental impact factor, calculate the weights corresponding to each modal feature using a weight calculation formula; and perform attention weighted fusion on multiple modal features based on the weights corresponding to each modal feature to obtain a target feature vector.

[0131] The formula for calculating the weights is as follows:

[0132]

[0133] in, It is the weight of the i-th modal feature at time t. It is the preset adjustment parameter for the i-th modal feature, f It is the environmental influencing factor at time t. It is the temperature data of the storage environment at time t.

[0134] In one embodiment of this application, the feature attention fusion module 22 is further configured to calculate an environmental impact factor based on warehouse environment temperature data using an environmental impact factor calculation formula; wherein, the environmental impact factor calculation formula is:

[0135]

[0136] in, It is the adjustment coefficient. It is the preset target temperature.

[0137] In one embodiment of this application, warehouse management decisions are generated based on multiple target warehouse characteristics and through a decision generation model;

[0138] The decision generation model is trained in the following way:

[0139] Multiple training samples and their corresponding annotation results are obtained. Each training sample includes multiple sample storage features. The sample storage features are obtained by extracting features from the multimodal storage data samples based on the data type.

[0140] Based on each training sample and through the initial model, corresponding predictive warehouse management decisions are obtained. These predictive warehouse management decisions include: numerical predictive decisions and categorical predictive decisions.

[0141] Based on each predicted warehouse management decision and its corresponding labeling results, the first loss is determined;

[0142] Based on the labeling results of each numerical prediction decision and its corresponding numerical warehouse management decision, the second loss is determined;

[0143] Based on the labeling results of each type of predictive decision and its corresponding type of warehouse management decision, the third loss is determined;

[0144] The initial model is trained based on the first loss, the second loss, and the third loss to obtain the decision generation model.

[0145] In one embodiment of this application, the predicted warehouse management decision includes: warehouse layout adjustment prediction results, inventory scheduling prediction results, and storage location recommendation prediction results; the warehouse management decision annotation results include: warehouse layout adjustment annotation results, inventory scheduling annotation results, and storage location recommendation annotation results; and the determination of a first loss based on each predicted warehouse management decision and its corresponding warehouse management decision annotation results includes: determining a first sub-loss based on each warehouse layout adjustment prediction result and its corresponding warehouse layout adjustment annotation results.

[0146] The second sub-loss is determined based on the inventory scheduling forecast results and their corresponding inventory scheduling labeling results;

[0147] The third sub-loss is determined based on the prediction results of each storage location and the corresponding storage location recommendation annotation results.

[0148] The first loss is determined based on the first sub-loss, the second sub-loss, and the third sub-loss.

[0149] In one embodiment of this application, the decision generation model includes a dynamic graph convolutional neural network;

[0150] Warehouse management decisions are determined based on target feature vectors and dynamic graph convolutional neural networks, and in the following way:

[0151] The adjacency matrix corresponding to the initial graph structure in the dynamic graph convolutional neural network is updated based on the target feature vector to obtain the updated target adjacency matrix. The initial graph structure is a graph structure modeled based on the initial items, initial equipment, and initial transportation paths in the warehouse. The nodes in the initial graph structure are used to represent the initial items and initial equipment, and the edges in the initial graph structure are used to represent the initial transportation paths between nodes. The target adjacency matrix includes node features corresponding to multiple nodes.

[0152] Multiple graph convolution operations are performed on the target adjacency matrix. In each graph convolution operation, the updated node features are obtained through the graph convolution update formula of that operation.

[0153] Warehouse management decisions are generated based on the updated node characteristics;

[0154] Among them, the Graph convolution operations The graph convolution update formula is:

[0155]

[0156] in, For the nodes in the target adjacency matrix After the first The node features obtained after the layer graph convolution operation are updated. For nodes in the target adjacency matrix One of the neighboring nodes, For the nodes in the target adjacency matrix In the Layer node characteristics, For the nodes in the target adjacency matrix The set of neighboring nodes, For the first The weight matrix of the layer, For activation function, For the first Layer bias terms.

[0157] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned system embodiments, for example... Figure 2 The functions of the multimodal feature extraction module 21, the feature attention fusion module 22, and the decision generation module 23 are shown.

[0158] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0159] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0160] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store information about warehousing data.

[0161] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in the embodiments of the intelligent warehousing decision generation method provided in this application, or they can execute the implementation methods of the electronic device 300 described in the embodiments of this application, which will not be repeated here.

[0162] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0163] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD) card, flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0164] Those skilled in the art will recognize that the modules / units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0165] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0166] In the embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the division of modules / units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules, units, or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or modules / units, or it may be an electrical, mechanical, or other form of connection.

[0167] The modules / units described as separate components may or may not be physically separate. Similarly, the components shown as modules / units may or may not be physical modules / units; they may be located in one place or distributed across multiple network modules / units. Some or all of the modules / units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0168] Furthermore, the functional modules / units in the various embodiments of this application can be integrated into one processing module / unit, or each module / unit can exist physically separately, or two or more modules / units can be integrated into one module / unit. The integrated modules / units described above can be implemented in hardware or in the form of software functional modules / units.

[0169] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating intelligent warehousing decisions, characterized in that, include: Acquire multimodal warehousing data, which includes inventory data, cargo attribute data, warehouse layout data, and transportation route data; extract features from the multimodal warehousing data according to the data type to obtain multiple warehousing features; the data types include numerical, categorical, and textual. The multiple warehouse features are mapped to a unified feature space to obtain multiple target warehouse features, and each target warehouse feature corresponds to a modal warehouse data; attention-weighted fusion of the multiple target warehouse features is performed based on warehouse environment sensor data to obtain a target feature vector; Warehouse management decisions are generated based on the target feature vector; the warehouse management decisions include warehouse layout strategies, inventory scheduling strategies, and item storage location strategies. A target warehouse feature is a modal feature; The warehouse environment sensor data includes warehouse environment temperature data; The attention-weighted fusion of multiple target warehouse features based on warehouse environment sensor data yields a target feature vector, including: Based on the warehouse environment temperature data, the environmental impact factor is calculated using the environmental impact factor calculation formula. The formula for calculating the environmental impact factor is as follows: in, It is the adjustment coefficient. It is the preset target temperature; Based on the aforementioned environmental impact factors, the weights corresponding to each modal feature are calculated using the weighting formula. Based on the weights corresponding to each modal feature, attention-weighted fusion is performed on multiple modal features to obtain the target feature vector; The weight calculation formula is as follows: in, It is the weight of the i-th modal feature at time t. It is the preset adjustment parameter for the i-th modal feature, f It is the environmental influencing factor at time t. It is the temperature data of the storage environment at time t.

2. The intelligent warehousing decision generation method as described in claim 1, characterized in that, The process involves extracting features from the multimodal warehousing data according to its data type, resulting in multiple warehousing features, including: For numerical warehouse data, the warehouse data is converted into a numerical feature vector; For categorical warehouse data, one-hot encoding is performed on the warehouse data to obtain a binary feature representation, and the binary feature representation is used as the category feature vector corresponding to the warehouse data. For text-based warehouse data, a word embedding model is used to transform the warehouse data into a text feature vector.

3. The intelligent warehousing decision generation method as described in claim 1, characterized in that, The warehouse management decision is generated based on the multiple target warehouse characteristics and through a decision generation model; The decision generation model was trained in the following way: Multiple training samples and their corresponding annotation results are obtained. Each training sample includes multiple sample storage features. The sample storage features are obtained by extracting features from the multimodal storage data samples based on the data type. Based on each training sample and through the initial model, corresponding predictive warehouse management decisions are obtained. These predictive warehouse management decisions include: numerical predictive decisions and categorical predictive decisions. Based on each of the predicted warehouse management decisions and their corresponding warehouse management decision labeling results, the first loss is determined; Based on the numerical prediction decisions and their corresponding numerical warehouse management decision labeling results, the second loss is determined; Based on the labeling results of each of the aforementioned categorical predictive decisions and their corresponding categorical warehouse management decisions, the third loss is determined; The initial model is trained based on the first loss, the second loss, and the third loss to obtain the decision generation model.

4. The intelligent warehousing decision generation method according to claim 3, characterized in that, The predictive warehouse management decisions include: warehouse layout adjustment prediction results, inventory scheduling prediction results, and storage location recommendation prediction results; the warehouse management decision annotation results include: warehouse layout adjustment annotation results, inventory scheduling annotation results, and storage location recommendation annotation results. The determination of the first loss based on each of the predicted warehouse management decisions and their corresponding warehouse management decision labeling results includes: Based on the prediction results of each warehouse layout adjustment and their corresponding warehouse layout adjustment annotation results, the first sub-loss is determined; The second sub-loss is determined based on the inventory scheduling prediction results and their corresponding inventory scheduling labeling results; The third sub-loss is determined based on the prediction results of each storage location recommendation and the corresponding storage location recommendation annotation results. The first loss is determined based on the first sub-loss, the second sub-loss, and the third sub-loss.

5. The intelligent warehousing decision generation method as described in claim 3, characterized in that, The decision generation model includes a dynamic graph convolutional neural network; The warehouse management decision is determined based on the target feature vector and the dynamic graph convolutional neural network in the following way: The adjacency matrix corresponding to the initial graph structure in the dynamic graph convolutional neural network is updated based on the target feature vector to obtain the updated target adjacency matrix. The initial graph structure is a graph structure modeled based on the initial items, initial equipment, and initial transportation routes in the warehouse. The nodes in the initial graph structure are used to represent the initial items and the initial equipment, and the edges in the initial graph structure are used to represent the initial transportation routes between nodes. The target adjacency matrix includes node features corresponding to multiple nodes. Multiple graph convolution operations are performed on the target adjacency matrix, wherein each graph convolution operation obtains the updated node features through the graph convolution update formula for that operation. Warehouse management decisions are generated based on the updated node characteristics; Among them, the The graph convolution update formula for the nth graph convolution operation is: in, For the nodes in the target adjacency matrix After the first The node features obtained after the layer graph convolution operation are updated. For nodes in the target adjacency matrix One of the neighboring nodes, For the nodes in the target adjacency matrix In the Layer node characteristics, For the nodes in the target adjacency matrix The set of neighboring nodes, For the first The weight matrix of the layer, For activation function, For the first Layer bias terms.

6. An intelligent warehousing decision generation system, characterized in that, include: A multimodal feature extraction module is used to acquire multimodal warehousing data, which includes inventory data, cargo attribute data, warehouse layout data, and transportation route data; features are extracted from the multimodal warehousing data according to the data type to obtain multiple warehousing features; the data types include numerical, categorical, and textual. The feature attention fusion module is used to map the multiple warehouse features to a unified feature space to obtain multiple target warehouse features, one of which corresponds to one modal warehouse data; and to perform attention-weighted fusion of the multiple target warehouse features based on warehouse environment sensor data to obtain a target feature vector; A target warehouse feature is a modal feature; The warehouse environment sensor data includes warehouse environment temperature data; The feature attention fusion module is specifically used to calculate the environmental impact factor based on the warehouse environment temperature data using the environmental impact factor calculation formula. The formula for calculating the environmental impact factor is as follows: in, It is the adjustment coefficient. It is the preset target temperature; Based on the aforementioned environmental impact factors, the weights corresponding to each modal feature are calculated using the weighting formula. Based on the weights corresponding to each modal feature, attention-weighted fusion is performed on multiple modal features to obtain the target feature vector; The weight calculation formula is as follows: in, It is the weight of the i-th modal feature at time t. It is the preset adjustment parameter for the i-th modal feature, f It is the environmental influencing factor at time t. It is the warehouse environment temperature data at time t; The decision generation module is used to generate warehouse management decisions based on the target feature vector; the warehouse management decisions include warehouse layout strategies, inventory scheduling strategies, and item storage location strategies.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 5.

Images