Intelligent warehouse dynamic replenishment method and system based on deep learning

Abstract

The present application relates to the technical field of data processing, more particularly, the present application relates to a kind of intelligent warehousing dynamic replenishment method and system based on deep learning, in response to the replenishment request of goods in warehousing system, determine the replenishment path from storage area to picking area;The Euclidean distance of the replenishment path is obtained;The current active device quantity in the adjacent area of the replenishment path is collected;The topological information of the adjacent area of the replenishment path is obtained.The present application introduces dynamic response time delay index, quantifies the nonlinear logistics time delay risk caused by warehousing equipment density using graph neural network and exponential model, and accordingly constructs demand fluctuation sensitive coefficient dynamic adjustment replenishment threshold.The method realizes the deep decoupling and alignment of logical demand prediction and physical execution capability, effectively solves the replenishment lag and shortage problem caused by AGV transportation congestion in high-density warehousing scene.

Description

Technical Field

[0001] This invention relates to the field of data processing technology. More specifically, this invention relates to a method and system for intelligent dynamic replenishment in warehousing based on deep learning. Background Technology

[0002] Intelligent warehousing systems widely utilize automated guided vehicles (AGVs) to automate the scheduling and replenishment of goods between storage and picking areas, maintaining efficient operational flow. In this scenario, accurately triggering replenishment instructions based on inventory depletion to ensure timely arrival of goods at the picking area is crucial for maintaining supply chain continuity.

[0003] To optimize the replenishment process, existing technical solutions typically incorporate deep learning algorithms (such as the Transformer model) to analyze the historical turnover sequence of inventory units. These algorithms primarily utilize self-attention mechanisms to capture long-range dependencies in time-series data, thereby predicting future demand for goods and, based on this, combining it with a fixed inventory threshold to formulate a replenishment plan.

[0004] However, the aforementioned methods of demand forecasting using deep learning have significant drawbacks in real-world high-density warehousing scenarios. Existing forecasting models are often limited to time-series analysis at the data level, assuming ideally unobstructed logistics execution channels while ignoring the dynamic constraints of the physical execution environment. In actual operations, increased equipment density within the warehouse space leads to congestion phenomena such as path conflicts and obstacle avoidance waiting, causing the actual replenishment execution time to increase dramatically and non-linearly rather than simply dividing distance by speed. This approach, which separates logical demand forecasting from physical execution capabilities, prevents the system from recognizing the latency risks caused by physical congestion. Replenishment is often triggered only when demand is predicted, but excessive AGV travel time can result in stockouts in the picking area before goods arrive, failing to meet the dynamic and accurate requirements for replenishment timing in complex physical environments. Summary of the Invention

[0005] This invention provides a method and system for dynamic replenishment in intelligent warehousing based on deep learning. It aims to solve the problem that in related technologies, replenishment is often triggered only when demand is predicted, but due to the long time that AGVs take to travel, the picking area will be out of stock before the goods are delivered, which cannot meet the dynamic and accurate requirements for replenishment timing in complex physical environments.

[0006] In a first aspect, the present invention provides a method and system for intelligent warehouse dynamic replenishment based on deep learning, comprising: determining a replenishment path from the storage area to the picking area in response to a replenishment request for goods in the warehousing system; The process involves: obtaining the Euclidean distance of the replenishment path; collecting the number of currently active devices in the vicinity of the replenishment path; obtaining the topological information of the vicinity of the replenishment path, including the graph node connectivity and connectivity of each storage location; the connectivity being the degree of the graph node corresponding to each storage location in the graph structure; determining the replenishment impact area based on the topological information; calculating the dynamic response delay index, which is positively correlated with the Euclidean distance, exponentially positively correlated with the number of active devices, and negatively correlated with the replenishment impact area; calculating the demand fluctuation sensitivity coefficient of the goods based on the dynamic response delay index and the coefficient of variation of the historical demand data of the goods; calculating the dynamic replenishment threshold based on the predicted demand of the goods and the demand fluctuation sensitivity coefficient; and generating a replenishment instruction based on the dynamic replenishment threshold and the real-time inventory of the picking area; wherein the replenishment impact area is positively correlated with the connectivity of each storage location in the topological information. By introducing dynamic response latency indicators and demand fluctuation sensitivity coefficients, this invention combines the physical congestion status of the warehousing environment (number of active devices, topology) with cargo demand forecasting. Compared to existing technologies that rely solely on time series forecasting or fixed threshold replenishment, this invention can sense the nonlinear logistics latency caused by increased device density. When congestion or large demand fluctuations are predicted, it automatically raises the replenishment threshold, triggering replenishment instructions in advance. This effectively solves the stockout problem caused by excessive AGV transit time in high-density warehousing scenarios, achieving dynamic matching between replenishment timing and physical execution capacity.

[0007] Furthermore, the calculation formula for the dynamic response delay index is as follows: In the formula, Indicates the current time The dynamic response latency index This represents the Euclidean distance of the replenishment path. This indicates the average moving speed of the AGV. Indicates the current time Regional congestion factor, The base of the natural logarithm is used. An exponential function quantifies the non-linear impact of the number of active devices on movement speed. Compared to the simple calculation of distance divided by speed, this formula accurately depicts the sharp increase in time consumption caused by obstacle avoidance and waiting in congested environments, providing a quantitative basis for accurately assessing replenishment lead time.

[0008] Furthermore, the method for determining the area of ​​the replenishment impact zone includes: using a graph neural network model to extract features from the physical information of each storage location in the warehousing system, obtaining feature vectors for each storage location; searching for all neighboring nodes reachable in k steps from the graph nodes on the replenishment path in the graph structure, obtaining an initial set of neighboring nodes; calculating the cosine similarity between the feature vectors of each neighboring node in the initial set and the graph nodes on the replenishment path, and removing neighboring nodes with a cosine similarity less than a preset threshold; the area of ​​the replenishment impact zone is related to the physical area of ​​the storage location corresponding to the remaining neighboring nodes; wherein, the preset threshold ranges from 0.65 to 0.85; the graph neural network model is a graph convolutional network (GCN) model. By using a graph neural network to extract the topological features of the storage locations and combining cosine similarity to filter neighboring nodes, the area that actually causes congestion on the replenishment path is accurately delineated. Compared to simply defining the range by geometric distance, this method eliminates invalid nodes that are physically close but logically disconnected (such as partitions or different levels), improving the accuracy of congestion assessment.

[0009] Furthermore, this includes: weighting and summing the physical areas and connectivity of the remaining neighboring nodes' corresponding storage locations to obtain the area of ​​the replenishment-affected region; wherein, the weight coefficient of each node is positively correlated with the node's degree. By combining the physical area and connectivity weights to calculate the affected region area, the spatial occupancy of different storage locations and the attributes of transportation hubs are comprehensively considered, making the congestion assessment more consistent with the actual physical environment's traffic capacity limitations, thereby more realistically reflecting the region's capacity and congestion potential.

[0010] Furthermore, the formula for calculating the demand fluctuation sensitivity coefficient is as follows: In the formula, This indicates the current inventory unit at the moment. Demand fluctuation sensitivity coefficient This represents the standard deviation of the historical demand series for this inventory unit. This represents the mean of the historical demand series for this inventory unit. This represents a constant used to prevent the denominator from being zero. Indicates the current time The dynamic response latency index This represents the natural logarithm function. A demand fluctuation sensitivity coefficient was constructed, coupling the traditional inventory variation coefficient with the physical latency indicator. This enables the system to exhibit higher replenishment sensitivity when facing the dual risks of large demand fluctuations and physical traffic congestion, effectively preventing the risk of inventory breakdown under adverse operating conditions.

[0011] Furthermore, the formula for the dynamic replenishment threshold is: In the formula, This represents the dynamic replenishment threshold at the current time t. This indicates the baseline threshold for the inventory unit based on the safety stock. This represents the predicted demand for the next fixed period, as output by the deep learning model. This indicates the current inventory unit at the moment. The demand fluctuation sensitivity coefficient was determined. A dynamic replenishment threshold calculation mechanism was established, which superimposes dynamic increments based on deep learning predictions and sensitivity coefficients on top of the safety stock. Compared to a fixed threshold, this mechanism can significantly increase the trigger level during congestion, allowing sufficient AGV transportation time to ensure that goods arrive at the picking area on time.

[0012] In a second aspect, a deep learning-based intelligent warehouse dynamic replenishment system is also provided, comprising a processor and a memory, characterized in that the memory stores a computer program, and the processor executes the computer program to implement the deep learning-based intelligent warehouse dynamic replenishment method as described in any one of claims 1-9.

[0013] Beneficial effects: By introducing a dynamic response delay index, this method quantifies the nonlinear logistics delay risk caused by warehousing equipment density using graph neural networks and exponential models, and constructs a demand fluctuation sensitivity coefficient to dynamically adjust replenishment thresholds. This approach achieves deep decoupling and alignment between logical demand forecasting and physical execution capabilities, effectively solving the replenishment lag and stockout problems caused by AGV transportation congestion in high-density warehousing scenarios. Attached Figure Description

[0014] Figure 1 This is a schematic flowchart illustrating a replenishment method according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating a comparison of the effects of embodiments according to the present invention with those of the prior art. Detailed Implementation

[0015] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0016] like Figure 1 As shown, S101: Data acquisition and topological feature extraction based on graph neural networks.

[0017] In this embodiment, multidimensional initial data is first collected from the warehousing system. This multidimensional initial data includes the historical turnover sequence of each inventory unit (warehouse) and the physical coordinate matrix of the picking area and storage area. To eliminate differences between data of different dimensions, the historical turnover sequence is normalized. It should be noted that the historical turnover sequence of each inventory unit is a set of data obtained by statistically analyzing the daily outbound frequency or quantity of that inventory unit within a fixed historical period. The fixed historical period can be set to 7 days or 15 days; in this embodiment, a 10-day fixed historical period is preferred.

[0018] Furthermore, considering that the connections between storage locations in a warehouse space depend not only on straight-line distance but also on aisle layout and area function, this embodiment utilizes a Graph Neural Network (GNN) to extract the physical topological relationships between storage locations. Specifically, a Graph Convolutional Network (GCN) architecture is used to construct the model. Each storage location in the warehouse space is considered a node in a graph, and the traversable paths between storage locations are considered edges.

[0019] The input is the feature vector of the node, which includes the physical coordinates of the storage location. The system identifies the location type (storage area or picking area) and shelf level. Through multi-layer graph convolution operations, high-dimensional feature vectors for each location node are output. These high-dimensional feature vectors implicitly contain information about the connectivity and spatial density of the locations within the warehouse network, more accurately reflecting their actual position in the logistics network.

[0020] S102: Construct dynamic response latency metrics.

[0021] In this embodiment, replenishment efficiency depends not only on the physical path length but also on the equipment density of the current work area. Therefore, this embodiment constructs a dynamic response latency index. To quantify this physical constraint.

[0022] First, the Euclidean distance of the replenishment path is calculated based on the location coordinate matrix. Then, to accurately assess environmental congestion, the regional congestion factor needs to be calculated. The formula for calculating the regional congestion factor is: In the formula, Indicates the current time Regional congestion factor, Indicates the current time Number of active devices in the area surrounding the replenishment path. Indicates the current time The area affected by the replenishment.

[0023] Regarding the area affected by replenishment The acquisition method involves first determining the replenishment path from the storage area to the picking area after receiving a replenishment request from the warehousing system, and then obtaining the set of graph nodes traversed by the replenishment path based on the current replenishment task. The replenishment path is represented as a sequence of nodes. For any node on the path Search for it in the graph structure The set of all neighboring nodes reachable in one step. In this embodiment, The preferred value is or .

[0024] Then, the union of the neighborhoods of all nodes along the path is calculated to obtain the initial set of influencing nodes. This initial set of influencing nodes is then refined using the aforementioned high-dimensional feature vectors: calculating the path nodes... Cosine similarity with neighboring nodes is used to retain only those with similarity greater than a preset threshold. The nodes, where the preset threshold The value range is 0.65~0.85, preferably a preset threshold. The value is 0.75. This filtering ensures... It only includes areas that are truly relevant in terms of functional logic and spatial connectivity, excluding irrelevant storage locations that are physically adjacent but located at different height levels or isolated by obstacles.

[0025] The filtered set of affected nodes is mapped back to physical space, and the area of ​​its replenishment impact region is calculated using the following formula: In the formula, Indicates the current time The area affected by replenishment This represents the set of affected nodes after filtering. Let u represent the area occupied by the u-th node. This represents the weight coefficient of the u-th node.

[0026] It should be noted that the weight coefficient of a node is positively correlated with its degree, which is the number of edges directly connected to that node in the graph structure, used to measure the connection complexity of the node in the warehouse topology. This is because devices located at high-dimensional nodes (such as intersections) generate more path conflict entropy and space occupancy rate compared to devices located at low-dimensional nodes (such as between single-row shelves). Therefore, assigning higher weights to nodes with higher degrees increases the likelihood of increased complexity. This results in a higher congestion factor in complex regions when the number of AGVs is the same. It is significantly higher than the simple region, thus more accurately characterizing the limitations of the physical execution environment on replenishment latency.

[0027] Next, a dynamic response latency metric is constructed to quantify the impact of physical congestion on replenishment execution time. The formula for calculating the dynamic response latency metric is: In the formula, Indicates the current time The dynamic response latency index This represents the Euclidean distance of the replenishment path. This indicates the average moving speed of the AGV. Indicates the current time Regional congestion factor, is the base of the natural logarithm.

[0028] From this formula, we can see that the dynamic response delay index With regional congestion factor It shows an exponential positive correlation. This means that as the equipment density in the work area increases (i.e., ... When the speed increases, due to the increased frequency of obstacle avoidance waiting, yielding, and route replanning, the actual replenishment execution time will no longer be a linear distribution of distance and speed, but will instead exhibit a non-linear and sharp increase. This indicator accurately depicts the amplifying effect of the physical execution pressure of storage space on time costs in congested environments.

[0029] S103: Construct a demand fluctuation sensitivity coefficient.

[0030] In this embodiment, based on the obtained historical inventory turnover sequence, a serial analysis is performed in conjunction with the dynamic response latency indicator. The analysis concludes that replenishment planning must consider the risk of inventory depletion due to increased latency within the forecast window.

[0031] The formula for calculating the demand fluctuation sensitivity coefficient of each inventory unit is as follows: In the formula, This indicates the current inventory unit at the moment. Demand fluctuation sensitivity coefficient This represents the standard deviation of the historical demand series for this inventory unit. This represents the mean of the historical demand series for this inventory unit. This represents a constant to prevent the denominator from being zero, for example, a value of... , Indicates the current time The dynamic response latency index This represents the natural logarithm function.

[0032] As can be seen from the above formula, the first half of the formula... Essentially, it's the coefficient of variation, reflecting the instability of demand for goods; the latter part... This reflects the impact of physical time delay. The logical relationship indicates that... Follow It increases with the increase of demand volatility, and also increases with the increase of demand volatility. This coefficient reflects a scenario where the physical path for replenishment is severely congested ( (Large) and the demand for the goods themselves is unpredictable and fluctuates greatly. When the stock shortage is large, the system has an extremely low tolerance for stockouts, therefore the calculated sensitivity coefficient is... This will increase significantly, suggesting that replenishment logic must capture demand signals earlier to compensate for potential lags at the physical level.

[0033] S104: Determine the dynamic replenishment threshold; In this embodiment, the demand fluctuation sensitivity coefficient obtained in step S103 is integrated into the deep learning prediction model to dynamically correct the original fixed replenishment threshold.

[0034] First, use a deep learning model to predict the demand value for the next fixed period. In this embodiment, the deep learning model preferably employs the Transformer model architecture. This model utilizes a self-attention mechanism to capture long-range dependencies in historical turnover sequences. Specifically, the model's input is a historical turnover sequence with a fixed past period, and the output is the predicted demand for a future fixed period. In the model hyperparameter settings, the number of heads in the multi-head attention mechanism is preferably set to [value missing]. or The hidden layer dimension is preferably set to or This is to ensure both feature extraction capability and computational efficiency.

[0035] Next, a dynamic replenishment threshold is constructed based on the prediction results and sensitivity coefficient. The formula for calculating the dynamic replenishment threshold is: In the formula, This represents the dynamic replenishment threshold at the current time t. This indicates the baseline threshold for the safety stock of that inventory unit. For example, the baseline threshold for the safety stock is 20% of the storage capacity of that inventory unit. This represents the predicted demand for the next fixed period, as output by the deep learning model. This indicates the current inventory unit at the moment. Demand fluctuation sensitivity coefficient.

[0036] From the above formula, it can be seen that, and This shows a positive correlation, achieving deep decoupling and real-time alignment between replenishment triggering timing and warehouse physical execution capabilities. Specifically, when the physical environment deteriorates ( Elevation leads to (increase) or expected surge in demand ( When the price increases, the dynamic replenishment threshold... It will be significantly higher than the base threshold. This means that the system will trigger replenishment when inventory is still high, allowing more time for the AGV to navigate through the congested warehouse, so that when it physically arrives at the picking area, it meets the demand and avoids stockouts.

[0037] S105: Enables dynamic replenishment scheduling.

[0038] In this embodiment, the system determines the current real-time inventory level in the picking area. The dynamic replenishment threshold calculated in step S104 Perform real-time comparison. If If the system detects an issue, it determines that replenishment is needed, automatically generates a replenishment instruction, and dispatches the AGV to transfer goods from the storage area to the picking area. If the current inventory is sufficient, no replenishment is needed. In this way, the system can adjust the replenishment timing based on real-time demand forecasts and physical congestion, avoiding ineffective replenishment (leading to overflow in the picking area) and untimely replenishment (leading to stockouts in the picking area) that may occur under the fixed threshold mode, thus improving overall operational efficiency.

[0039] like Figure 2 As shown in the image, the content is as follows: The image compares two inventory curves. The solid green line represents the solution of this invention, while the dashed red line represents the fixed threshold solution of the prior art. During peak demand and severe congestion, the prior art repeatedly causes inventory to bottom out and drop to the zero axis (red shaded area), resulting in significant shortages. However, because this invention dynamically adjusts the trigger point, it successfully completes the allocation before the inventory is depleted, and the inventory curve always remains above a safe level.

[0040] The present invention also provides a deep learning-based intelligent warehouse dynamic replenishment system. The system includes a processor and a memory, the memory storing computer program instructions. When the processor executes the computer program instructions, it implements the deep learning-based intelligent warehouse dynamic replenishment method according to the first aspect of the present invention.

[0041] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and therefore will not be described in detail here.

[0042] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented using computer-readable / executable instructions stored or otherwise maintained on such a computer-readable medium.

[0043] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent warehouse dynamic replenishment method, characterized in that, include: In response to a replenishment request for goods within the warehousing system, determine a replenishment path from the storage area to the picking area; obtain the Euclidean distance of the replenishment path; and collect the number of currently active devices in the vicinity of the replenishment path. Obtain the topology information of the adjacent area of ​​the replenishment path. The topology information includes the graph node connection relationship and connectivity of each storage location. The connectivity is the degree of the graph node corresponding to each storage location in the graph structure. Based on the topology information, the area affected by replenishment in the adjacent region is determined; Calculate the dynamic response delay index, which is positively correlated with the Euclidean distance, exponentially positively correlated with the number of active devices, and negatively correlated with the area of ​​the replenishment impact zone; calculate the demand fluctuation sensitivity coefficient of the goods based on the dynamic response delay index and the coefficient of variation of the historical demand data of the goods. Based on the predicted demand for the goods and the demand fluctuation sensitivity coefficient, a dynamic replenishment threshold is calculated; a replenishment instruction is generated according to the dynamic replenishment threshold and the real-time inventory of the picking area; wherein, the area of ​​the replenishment-affected region is positively correlated with the connectivity of each storage location in the topology information.

2. The intelligent warehousing dynamic replenishment method based on deep learning according to claim 1, characterized in that, The formula for calculating the dynamic response delay index is: ； In the formula, Indicates the current time The dynamic response latency index This represents the Euclidean distance of the replenishment path. This indicates the average moving speed of the AGV. Indicates the current time Regional congestion factor, is the base of the natural logarithm.

3. The intelligent warehousing dynamic replenishment method based on deep learning according to claim 1, characterized in that, The method for determining the area affected by the replenishment includes: A graph neural network model is used to extract features from the physical information of each storage location in the warehousing system to obtain the feature vector of each storage location; and all neighboring nodes reachable by k steps from the nodes on the replenishment path in the graph structure are searched to obtain the initial set of neighboring nodes. Calculate the cosine similarity between the feature vectors of each neighbor node in the initial neighbor node set and the nodes in the graph of the replenishment path, and remove neighbor nodes whose cosine similarity is less than a preset threshold; while the area of ​​the replenishment influence region is related to the physical area of ​​the storage location corresponding to the remaining neighbor nodes.

4. The intelligent warehousing dynamic replenishment method based on deep learning according to claim 3, characterized in that, include: The area of ​​the replenishment-affected region is obtained by weighted summing of the physical area of ​​the corresponding storage location of the remaining neighboring nodes and their connectivity.

5. The intelligent warehousing dynamic replenishment method based on deep learning according to claim 4, characterized in that, The weight coefficient of each node is positively correlated with the degree of that node.

6. The intelligent warehousing dynamic replenishment method based on deep learning according to claim 1, characterized in that, The formula for calculating the demand fluctuation sensitivity coefficient is as follows: ； In the formula, This indicates the current inventory unit at the moment. Demand fluctuation sensitivity coefficient This represents the standard deviation of the historical demand series for this inventory unit. This represents the mean of the historical demand series for this inventory unit. This represents a constant used to prevent the denominator from being zero. Indicates the current time The dynamic response latency index This represents the natural logarithm function.

7. The intelligent warehousing dynamic replenishment method based on deep learning according to claim 1, characterized in that, The formula for the dynamic replenishment threshold is: In the formula, This represents the dynamic replenishment threshold at the current time t. This indicates the baseline threshold for the inventory unit based on the safety stock. This represents the predicted demand for the next fixed period, as output by the deep learning model. This indicates the current inventory unit at the moment. Demand fluctuation sensitivity coefficient.

8. The intelligent warehousing dynamic replenishment method based on deep learning according to claim 3, characterized in that, The preset threshold value ranges from 0.65 to 0.

85.

9. The intelligent warehousing dynamic replenishment method based on deep learning according to claim 3, characterized in that, The graph neural network model is a graph convolutional network (GCN) model.

10. A deep learning-based intelligent warehouse dynamic replenishment system, comprising a processor and a memory, characterized in that, The memory stores a computer program, and the processor executes the computer program to implement the deep learning-based intelligent warehouse dynamic replenishment method as described in any one of claims 1-9.

Images