A warehouse intelligent management system based on deep learning

Abstract

The present application relates to warehouse management technical field, especially a kind of warehouse intelligent management system based on deep learning, by the space distribution information based on warehouse goods location, function site and channel interconnection, combine historical warehouse record, AGV track and dispatch log, construct the space-time graph model containing node and edge.By deep space-time graph convolution modeling, output goods location value score and traffic congestion probability, identify the high-frequency SKU should be preferentially allocated goods location area.And through reinforcement learning model generation goods location allocation strategy, realize high-frequency goods fast in and out of warehouse while avoiding channel excessive occupation and deadlock.Finally, through path planning algorithm generation AGV dispatch instruction set and issue execution.Goods location planning and AGV scheduling dynamic collaborative optimization are realized, while improving the efficiency of high-frequency goods operation, effectively avoid path congestion and conflict, improve system throughput and operation reliability.

Description

Technical Field

[0001] This invention relates to the field of warehouse management technology, and more particularly to a warehouse intelligent management system based on deep learning. Background Technology

[0002] As warehousing scale continues to expand, Automated Guided Vehicles (AGVs) are widely used in inbound and outbound operations. However, most existing warehouses employ static location planning, fixing the storage locations of goods initially and lacking the ability to dynamically optimize based on the frequency and flow patterns of goods entering and leaving the warehouse. This static planning approach leads to the following problems: Static location planning often ignores the distance relationship between high-frequency goods and main aisles; high-frequency SKUs may be stored far from the main aisles or even scattered across different areas of the warehouse. When performing high-frequency picking tasks, AGVs need to make long detours, increasing the length of the operating path and energy consumption, thus reducing overall operational efficiency. This problem is particularly pronounced when orders are concentrated, directly leading to delays in outbound response and decreased transportation efficiency.

[0003] Secondly, in actual operation, to alleviate the problem of low retrieval efficiency for high-frequency goods, it is often necessary to adjust the storage locations, moving high-frequency goods to closer positions near the main aisle or operating area. However, storage location adjustments can lead to a large number of AGVs occupying certain aisles in a short period of time, causing path conflicts and congestion. If multiple AGVs run towards each other in the same narrow aisle, it can easily lead to a deadlock, which can only be resolved through manual intervention or global rescheduling, resulting in a decrease in system throughput and an increase in scheduling costs. Summary of the Invention

[0004] To address the above problems, this invention provides a deep learning-based intelligent warehouse management system.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A deep learning-based intelligent warehouse management system includes:

[0007] The data processing module is used to construct a spatiotemporal graph model containing nodes and edges based on the spatial distribution information of storage locations and functional locations in the warehouse and the channel connectivity. It maps historical inbound and outbound records, AGV trajectory data and task scheduling logs to the node and edge features of the spatiotemporal graph model in a time series to obtain a multidimensional spatiotemporal feature sequence.

[0008] The cargo location assessment module is used to perform temporal modeling and graph structure learning on the multidimensional spatiotemporal feature sequence, and output the cargo location value score corresponding to each node, the traffic congestion probability corresponding to each edge, and the unit traffic delay correction factor.

[0009] The cargo location allocation module is used to filter the entire set of cargo locations based on the cargo location value score and the current cargo location occupancy status to obtain a cargo migration candidate set, and to construct a reinforcement learning state vector by combining the traffic congestion probability and the traffic delay correction factor, and output a cargo location allocation strategy through the reinforcement learning model.

[0010] The job scheduling module is used to generate an AGV scheduling instruction set based on the cargo location allocation strategy using a path planning algorithm, and then send the scheduling instruction set to the corresponding AGV terminal to execute the job.

[0011] Furthermore, the construction of a spatiotemporal graph model containing nodes and edges based on the spatial distribution information of storage locations and functional locations in the warehouse and the connectivity of passageways includes the following steps:

[0012] Based on the location and function number in the warehouse management system, their coordinate information is extracted, a set of nodes in the graph model is constructed, and the connectivity between nodes is logically calculated according to the passable path to form an edge set and mark the path length and direction of passage.

[0013] Based on the SKU number, location number and timestamp contained in the historical inbound and outbound records, they are mapped to the node set, and the time-series inbound and outbound frequency of each node is counted as the dynamic feature of the node.

[0014] Based on the path sequence and time period recorded in the AGV trajectory data and task scheduling log, the corresponding edge set is mapped, and the passage frequency and average occupation time per unit time are used as the dynamic features of the edge.

[0015] The node set, edge set, and edge dynamic features are time-aligned and normalized to construct a spatiotemporal graph model.

[0016] Furthermore, the step of mapping historical inbound and outbound records, AGV trajectory data, and task scheduling logs to the node and edge features of the spatiotemporal graph model according to time series to obtain a multidimensional spatiotemporal feature sequence includes the following steps:

[0017] Based on the SKU code, operation time and storage location number contained in the historical inbound and outbound records, each record is classified by timestamp and mapped to the corresponding node. The outbound frequency, inbound frequency and operation heat of each node within a fixed time window are calculated to generate the time series feature vector of the node.

[0018] Based on the starting position, path trajectory and time period recorded in the AGV trajectory data, the channel segments traversed by the AGV are mapped to the edge set in the graph. Based on the execution time and task density in the task scheduling log, the passage frequency and average occupation time of each edge in each time slice are statistically analyzed to generate the time series feature vector of the edge.

[0019] The time-series feature vectors of nodes and edges are time-aligned to obtain a multidimensional spatiotemporal feature sequence.

[0020] Furthermore, the location assessment module is used to perform the following steps:

[0021] Based on the node-based time series feature vector, feature segments of adjacent time steps are extracted according to a fixed-length time window. One-dimensional convolution operation and normalization are performed on the segments to obtain the time convolution representation of each node.

[0022] The temporal convolutional representation of the node is multiplied by the node adjacency matrix, and after weighted summation, nonlinear activation processing is performed to obtain a node graph convolutional representation containing neighbor relationships.

[0023] Based on the time series feature vectors of the edges, the passage frequency and occupation duration of each edge within the time window are processed by sliding convolution and standardization to obtain the time convolution representation of the edge.

[0024] The temporal convolutional representation of the edge is concatenated with the node graph convolutional representation corresponding to the start and end node indices of the edge, and then weighted and summed using a weight matrix to obtain the graph convolutional representation of the edge.

[0025] Input the node graph convolution representation into the first regression output unit to output the cargo location value score corresponding to each node; input the edge graph convolution representation into the second regression output unit to output the traffic congestion probability and unit traffic delay correction factor corresponding to each edge.

[0026] Furthermore, the first return-to-output unit is trained through the following steps:

[0027] Based on historical inbound and outbound records and AGV actual operation logs, the outbound frequency, inbound frequency and operation duration of each storage location within a fixed time window are extracted. Based on the average travel time of the storage location in the dense area of ​​the AGV trajectory, a real value rating label is generated, and a first training sample set containing the training node graph convolution representation and the real value rating label is constructed.

[0028] The training node graph convolution representation in the first training sample set is input into the first regression output unit to perform matrix multiplication and nonlinear transformation, and the predicted cargo location value score is output.

[0029] The mean squared error is calculated by comparing the predicted cargo location value score with the real value score labels in the training samples, and a loss function is formed by combining it with regularization constraints. The parameters are then updated iteratively through backpropagation.

[0030] Furthermore, the second return-to-output unit is trained through the following steps:

[0031] Based on AGV operation logs and task scheduling records, the passage frequency, average occupation time and waiting delay of each edge within a fixed time window are statistically analyzed. The actual passage congestion probability and unit passage delay correction factor are calculated as real passage parameter labels. A second training sample set containing edge convolutional representation and real passage parameter labels is constructed.

[0032] The training graph of the edges is convolved and input into the second regression output unit to perform numerical operations and nonlinear transformations, and outputs the predicted traffic congestion probability and the unit traffic delay correction factor.

[0033] The mean square error is calculated based on the predicted traffic congestion probability and the unit traffic delay correction factor, along with the actual traffic parameter labels. The parameters are then updated iteratively through backpropagation.

[0034] Furthermore, the step of filtering the entire set of cargo locations to obtain a feasible cargo migration candidate set based on the cargo location value score and the current cargo location occupancy status includes the following steps:

[0035] Based on the location value score and current occupied capacity of each location, all locations are traversed. Locations with a value score higher than the preset threshold and not full capacity are selected as candidates for shelving, while locations with a value score lower than the preset threshold and inventory waiting to be moved are selected as candidates for moving out.

[0036] Based on the SKU compatibility rules, combinations that do not meet compatibility or traffic constraints are eliminated from the candidates for shelf placement and candidates for relocation, resulting in a set of candidates for cargo relocation.

[0037] Furthermore, the reinforcement learning model is constructed through the following steps:

[0038] The cargo migration candidate set, traffic congestion probability and unit traffic delay correction factor are normalized and encoded into a state vector, and the state space of reinforcement learning is constructed through the state vector.

[0039] Each feasible migration or shelving operation in the candidate goods migration set is set as an independent action and assigned a unique action code. The action space of reinforcement learning is constructed through the set of action codes.

[0040] In the simulation execution, the corresponding travel time and congestion delay are extracted from the action encoding and the state vector to construct a reward function;

[0041] The state space, action space, and reward function are used to construct the training input of the reinforcement learning model. The policy network parameters are iteratively optimized to output the cargo location allocation policy, which is a set of action codes.

[0042] Furthermore, the job scheduling module is used to perform the following steps:

[0043] Based on the action codes output by the cargo location allocation strategy, the corresponding cargo migration tasks are parsed and generated based on the start and end nodes of each task and the channel connectivity.

[0044] The task path graph is input into the path planning algorithm, and path search is performed based on path length and connectivity between nodes. In the case of multiple tasks in parallel, conflict detection is performed to obtain a path scheme that meets the time constraints.

[0045] The AGV scheduling instruction set is generated according to the route scheme, and the AGV scheduling instruction set is sent to the corresponding AGV terminal to drive the AGV to complete the cargo transfer operation.

[0046] Furthermore, the path planning algorithm is a heuristic search algorithm.

[0047] The beneficial effects of this invention are as follows: This invention constructs a spatiotemporal graph model containing nodes and edges based on the spatial distribution information of storage locations and functional locations, the connectivity of channels, historical inbound / outbound records, AGV trajectory data, and task scheduling logs. It extracts multi-dimensional spatiotemporal feature sequences, thereby completely replicating the flow patterns of goods and the operating modes of AGVs. By performing deep spatiotemporal graph convolutional modeling on the multi-dimensional spatiotemporal features, it outputs storage location value scores and traffic congestion probabilities, effectively identifying storage location areas that should be prioritized for high-frequency SKUs. This avoids the problem of high-frequency goods being far from the main channel under static planning, enabling AGVs to complete high-frequency inbound / outbound operations with shorter paths. Based on the storage location value score, the traffic congestion probability and unit traffic delay correction factor are incorporated into the reinforcement learning state vector. The reinforcement learning model outputs a storage location allocation strategy, ensuring rapid inbound / outbound of high-frequency goods while avoiding excessive channel occupancy and deadlocks due to concentrated migration. A path planning algorithm is invoked to generate an AGV scheduling instruction set, which is then distributed to the corresponding terminals, ensuring that multiple AGVs do not conflict with each other when executing tasks in parallel. It achieves dynamic collaborative optimization of warehouse location planning and AGV scheduling, which improves the efficiency of high-frequency goods entry and exit while effectively avoiding path congestion and conflicts, and improves system throughput and operational reliability. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the structure of a deep learning-based intelligent warehouse management system according to the present invention.

[0049] Figure 2 This is a flowchart of the execution steps of the cargo location assessment module in this invention. Detailed Implementation

[0050] Please see Figures 1-2 As shown, this invention relates to a deep learning-based intelligent warehouse management system, comprising:

[0051] The data processing module is used to construct a spatiotemporal graph model containing nodes and edges based on the spatial distribution information of storage locations and functional locations in the warehouse and the channel connectivity. It maps historical inbound and outbound records, AGV trajectory data and task scheduling logs to the node and edge features of the spatiotemporal graph model in a time series to obtain a multidimensional spatiotemporal feature sequence.

[0052] The cargo location assessment module is used to perform temporal modeling and graph structure learning on the multidimensional spatiotemporal feature sequence, and output the cargo location value score corresponding to each node, the traffic congestion probability corresponding to each edge, and the unit traffic delay correction factor.

[0053] The cargo location allocation module is used to filter the entire set of cargo locations based on the cargo location value score and the current cargo location occupancy status to obtain a cargo migration candidate set, and to construct a reinforcement learning state vector by combining the traffic congestion probability and the traffic delay correction factor, and output a cargo location allocation strategy through the reinforcement learning model.

[0054] The job scheduling module is used to generate an AGV scheduling instruction set based on the cargo location allocation strategy using a path planning algorithm, and then send the scheduling instruction set to the corresponding AGV terminal to execute the job.

[0055] In some embodiments, a graph structure is first constructed based on the spatial distribution coordinates of warehouse locations and functional locations and the connectivity of passageways, forming a complete spatial topology between location nodes and passageway edges. Simultaneously, historical inbound / outbound records, AGV trajectory data, and task scheduling logs are introduced. Through time-series mapping, nodes and edges acquire temporal dynamic characteristics, thus forming a multi-dimensional spatiotemporal feature sequence that reflects the frequency of goods flow and AGV traffic load. Compared with existing technologies, this modeling method not only replicates the static spatial layout but also depicts dynamic flow patterns, enabling the system to simultaneously perceive the demand for goods and traffic congestion trends. The aforementioned feature sequence is learned and modeled using a deep spatiotemporal graph convolutional network. After fusion of temporal and graph convolutions, node features output a location value score to reflect the importance of each location in high-frequency inbound / outbound operations; edge features are processed through convolution and normalization to generate traffic congestion probability and unit traffic delay correction factors, quantifying the potential conflict risks and delay costs of different passageways. This process differs from traditional methods that rely on human experience or static weighted scoring. Instead, it utilizes deep learning to automatically extract spatiotemporal features, making the assessment of storage location and aisle value more dynamic and accurate. Furthermore, based on storage location value scores, the current occupancy status, congestion probability, and unit passage delay correction factor are combined into a reinforcement learning state vector. A storage location allocation strategy is then generated through a reinforcement learning model. This strategy not only ensures that high-frequency SKUs are prioritized for allocation to near-aisle, high-value areas, improving inbound and outbound efficiency, but also avoids aisle congestion and deadlock problems caused by concentrated storage location adjustments in a particular area. Compared to existing migration strategies that rely on heuristic rules, this embodiment achieves Pareto optimality in long-term reward through the training of the strategy network. In the job scheduling stage, the system generates migration and inbound / outbound tasks based on the storage location allocation strategy and calls a path planning algorithm to calculate the optimal travel path, generating an AGV scheduling instruction set containing time constraints. After this instruction set is issued to the AGV terminal, it enables conflict-free operation of multiple AGVs in parallel, avoiding congestion and manual intervention caused by path intersections or inconsistent timing in traditional scheduling. This embodiment innovatively combines spatiotemporal graph learning, depth graph convolution, and reinforcement learning to achieve dynamic and interconnected optimization of warehouse location value assessment, allocation, and scheduling. Its innovation, distinct from existing technologies, lies in not only considering the matching relationship between high-frequency SKUs and channel distances but also incorporating traffic delay and congestion risk into the optimization objectives. This allows the warehouse system to avoid traffic deadlocks while ensuring efficient inbound and outbound operations, significantly improving the intelligence and reliability of warehouse operations.

[0056] Furthermore, the construction of a spatiotemporal graph model containing nodes and edges based on the spatial distribution information of storage locations and functional locations in the warehouse and the connectivity of passageways includes the following steps:

[0057] Based on the location and function number in the warehouse management system, their coordinate information is extracted, a set of nodes in the graph model is constructed, and the connectivity between nodes is logically calculated according to the passable path to form an edge set and mark the path length and direction of passage.

[0058] Based on the SKU number, location number and timestamp contained in the historical inbound and outbound records, they are mapped to the node set, and the time-series inbound and outbound frequency of each node is counted as the dynamic feature of the node.

[0059] Based on the path sequence and time period recorded in the AGV trajectory data and task scheduling log, the corresponding edge set is mapped, and the passage frequency and average occupation time per unit time are used as the dynamic features of the edge.

[0060] The node set, edge set, and edge dynamic features are time-aligned and normalized to construct a spatiotemporal graph model.

[0061] In some embodiments, the system first extracts the three-dimensional coordinate data of the storage location numbers and function location numbers maintained in the warehouse management system in physical space. Each storage location or function location is set as a node, and its corresponding spatial coordinates are vectors, thereby constructing a node set for a graph neural network model. Since there are multiple functional areas in the warehouse, such as storage areas, picking areas, temporary storage areas, and entrances / exits, a semantic type label field needs to be added to the node attributes to distinguish the function type. At the same time, a regional hierarchical encoding is introduced to support scene modeling of multi-floor and multi-channel structures. Based on this, according to the warehouse navigation map or AGV map, the system establishes the connectivity between nodes through an adjacency matrix. If there is a passable path between two nodes, an edge is introduced into the graph structure, and each edge is assigned a path weight, such as Euclidean distance, actual path length, or passage cost. At the same time, the direction attribute (one-way or two-way) is labeled, ultimately forming an initial directed weighted graph G=(V,E) containing a set of nodes V and a set of edges E. Historical inbound and outbound records and AGV scheduling trajectory data are introduced as the basis for generating dynamic features. Inbound and outbound records typically include SKU number, operation type (inbound / outbound), operation time, and corresponding storage location number. Based on these records, the system maps SKU inbound and outbound behavior to corresponding nodes in the graph and statistically analyzes dynamic characteristics such as inbound / outbound frequency and SKU type changes for each node within a set time window (e.g., every 5 minutes). This statistical analysis forms a dynamic feature matrix of nodes that changes over time, characterizing the task density and logistics level of each storage location during operation. For edge dynamic feature modeling, the path trajectories and task allocation logs in the AGV scheduling system are used to identify indicators such as the frequency of historical tasks on each channel segment (i.e., edges in the graph), average travel time, and maximum simultaneous travel count, which serve as dynamic attributes of the edges. For example, if a path connecting an entrance / exit to a high-frequency cargo area is repeatedly used by multiple AGVs within a unit of time, it may lead to path congestion or waiting. This information is crucial for identifying potential channel bottlenecks and adjusting path weights. To ensure that all dynamic features can be used as input to spatiotemporal graph neural networks (such as ST-GCN and TGAT), the system needs to align and normalize the dynamic attributes of nodes and edges along the time dimension, ensuring the data integrity and comparability at each time step. The final spatiotemporal graph model not only includes static structural information (spatial topology of nodes and edges) and static attributes of nodes / edges, but also incorporates temporal dynamic features reflecting actual task flows and traffic flows. This provides a unified modeling foundation for subsequent path conflict prediction, heat migration simulation, and AGV scheduling optimization.

[0062] Furthermore, the step of mapping historical inbound and outbound records, AGV trajectory data, and task scheduling logs to the node and edge features of the spatiotemporal graph model according to time series to obtain a multidimensional spatiotemporal feature sequence includes the following steps:

[0063] Based on the SKU code, operation time and storage location number contained in the historical inbound and outbound records, each record is classified by timestamp and mapped to the corresponding node. The outbound frequency, inbound frequency and operation heat of each node within a fixed time window are calculated to generate the time series feature vector of the node.

[0064] Based on the starting position, path trajectory and time period recorded in the AGV trajectory data, the channel segments traversed by the AGV are mapped to the edge set in the graph. Based on the execution time and task density in the task scheduling log, the passage frequency and average occupation time of each edge in each time slice are statistically analyzed to generate the time series feature vector of the edge.

[0065] The time-series feature vectors of nodes and edges are time-aligned to obtain a multidimensional spatiotemporal feature sequence.

[0066] In some embodiments, the system first constructs node features based on historical inbound and outbound records. Each record contains an SKU code, an operation timestamp, and a storage location number. The system categorizes this data by timestamp, dividing inbound and outbound events into multiple discrete time slices using a sliding time window (e.g., at 5-minute intervals). Within each time slice, the system counts the outbound frequency (i.e., the number of times an SKU is extracted from that node) and the inbound frequency (i.e., the number of times an SKU is stored in that node) for each storage location node. Furthermore, it combines the inbound / outbound alternation frequency with SKU category richness to calculate a comprehensive operation popularity index. Finally, each node generates a dynamic feature in vector form for each time slice, representing the outbound frequency, inbound frequency, and popularity index, respectively. As the time slice progresses, this vector forms the node's time-series feature vector in the time dimension. The construction of edge features relies on the fusion processing of AGV trajectory data and task scheduling logs. Each AGV trajectory data typically includes the starting position, the node sequence of the travel path, and the corresponding task time period. The system maps each travel path segment in the trajectory to an edge in the graph model and counts the travel frequency of that edge in each time slice, i.e., how many AGV tasks use it per unit time. Simultaneously, by combining the task start and end times and the number of AGVs executed recorded in the task scheduling log, the system further estimates the average occupancy time and task concurrency of that edge to reflect the traffic pressure on that edge. A feature vector is generated for each edge within each time slice. To enable the feature sequences of nodes and edges to be used for spatiotemporal graph modeling and downstream prediction tasks, such as congestion prediction or path optimization, the system performs alignment processing on the feature vectors of all nodes and edges in the time dimension. This processing ensures that all nodes and edges have synchronized feature value sequences in the same time slice set, avoiding temporal information mismatch caused by inconsistent time granularity. The alignment method can adopt a time window mapping strategy based on unified timestamp alignment, combined with a missing data filling mechanism to handle sparse data segments. The resulting multidimensional spatiotemporal feature sequence will be input into the subsequent graph temporal modeling network in the form of a three-dimensional tensor. The first dimension represents the time step, the second the node or edge index, and the third the corresponding feature dimension. This implementation scheme effectively captures the activity level of storage locations and path congestion characteristics while preserving the temporal dynamics of operations and scheduling, providing high-dimensional, multi-granularity spatiotemporal data support for graph neural network-based prediction and scheduling optimization algorithms. This implementation is particularly suitable for complex warehousing scenarios with high throughput and multiple AGV scheduling, helping to dynamically identify bottleneck paths and deduce the impact of storage location migration strategies on path efficiency.

[0067] Furthermore, the location assessment module is used to perform the following steps:

[0068] Based on the node-based time series feature vector, feature segments of adjacent time steps are extracted according to a fixed-length time window. One-dimensional convolution operation and normalization are performed on the segments to obtain the time convolution representation of each node.

[0069] The temporal convolutional representation of the node is multiplied by the node adjacency matrix, and after weighted summation, nonlinear activation processing is performed to obtain a node graph convolutional representation containing neighbor relationships.

[0070] Based on the time series feature vectors of the edges, the passage frequency and occupation duration of each edge within the time window are processed by sliding convolution and standardization to obtain the time convolution representation of the edge.

[0071] The temporal convolutional representation of the edge is concatenated with the node graph convolutional representation corresponding to the start and end node indices of the edge, and then weighted and summed using a weight matrix to obtain the graph convolutional representation of the edge.

[0072] Input the node graph convolution representation into the first regression output unit to output the cargo location value score corresponding to each node; input the edge graph convolution representation into the second regression output unit to output the traffic congestion probability and unit traffic delay correction factor corresponding to each edge.

[0073] In some embodiments, based on the temporal series feature vectors of nodes, a fixed-length sliding time window is used to extract segments of cargo location operation data within a continuous time period. A one-dimensional convolutional network (1D-CNN) is then used to perform feature compression and local temporal pattern capture, while batch normalization is introduced to enhance training stability, resulting in a convolutional representation with temporal locality features. Subsequently, the temporal convolutional representation of the nodes is subjected to graph convolution with the node's adjacency matrix. Specifically, a weighted aggregation method based on the adjacency matrix is ​​used to fuse the feature information of adjacent nodes, and a non-linear activation function (such as ReLU or LeakyReLU) is introduced after aggregation to enhance the expressive power of the features. This process not only considers the historical entry and exit dynamics of a single cargo location but also introduces the entry and exit frequency patterns of its spatial neighbor nodes through structural propagation, thereby achieving contextual modeling of the cargo location's usage value. In the edge feature processing part, the system also uses the temporal series feature vectors of the edges to extract the changing patterns of passage frequency and passage duration using sliding convolution, and unifies the dimensions through z-score or min-max normalization to obtain the temporal convolutional representation of the edges. This representation is used to characterize the load dynamics of different channel segments within each time slice, helping to assess the congestion trend and path traffic stability in local areas. Further, the temporal convolutional representation of each edge is concatenated with the graph convolutional representations of its two endpoints, and weighted fusion is performed using a shared weight matrix to generate an edge-graph convolutional representation that includes spatial context information and historical traffic status. Finally, the node graph convolutional representation and the edge graph convolutional representation are input into two independent regression output units. The node branch outputs the value score of each cargo location, and the edge branch outputs the congestion probability and unit distance travel delay correction factor for each channel.

[0074] Furthermore, the first return-to-output unit is trained through the following steps:

[0075] Based on historical inbound and outbound records and AGV actual operation logs, the outbound frequency, inbound frequency and operation duration of each storage location within a fixed time window are extracted. Based on the average travel time of the storage location in the dense area of ​​the AGV trajectory, a real value rating label is generated, and a first training sample set containing the training node graph convolution representation and the real value rating label is constructed.

[0076] The training node graph convolution representation in the first training sample set is input into the first regression output unit to perform matrix multiplication and nonlinear transformation, and the predicted cargo location value score is output.

[0077] The mean squared error is calculated by comparing the predicted cargo location value score with the real value score labels in the training samples, and a loss function is formed by combining it with regularization constraints. The parameters are then updated iteratively through backpropagation.

[0078] Specifically, firstly, feature values ​​such as outbound frequency, inbound frequency, and operation time for each storage location within a fixed time window are extracted from the warehouse's historical inbound and outbound records. These values ​​serve as reference indicators reflecting the frequency of storage location usage and operational complexity. Simultaneously, combined with trajectory information recorded in the AGV operation logs, the average travel time of AGVs on the surrounding paths of each storage location is calculated to further evaluate its accessibility and scheduling efficiency in path planning. Based on this multi-source information, actual storage location value rating labels are generated through normalization and weighted fusion, and combined with the graph convolutional feature vectors of the corresponding nodes to construct a training sample set. During model training, the graph convolutional representation of each sample is fed as input into the first regression output unit, which contains several fully connected layers and nonlinear activation functions to map the node feature vectors to the prediction space of the target storage location value rating. The mean squared error is used as the basic loss function between the predicted rating and the true rating label output by the model. An L2 regularization term is introduced to suppress model overfitting, and the network parameters are continuously updated iteratively through gradient descent optimization and backpropagation mechanisms. In practical implementation, to improve the model's ability to distinguish between sparse and frequently used regions, a sample weighting mechanism can be introduced, assigning higher weights to high-frequency cargo location samples, thereby making the model focus more on optimizing key path nodes. After training, the first regression output unit can output a real-time value score for any cargo location node.

[0079] Furthermore, the second return-to-output unit is trained through the following steps:

[0080] Based on AGV operation logs and task scheduling records, the passage frequency, average occupation time and waiting delay of each edge within a fixed time window are statistically analyzed. The actual passage congestion probability and unit passage delay correction factor are calculated as real passage parameter labels. A second training sample set containing edge convolutional representation and real passage parameter labels is constructed.

[0081] The training graph of the edges is convolved and input into the second regression output unit to perform numerical operations and nonlinear transformations, and outputs the predicted traffic congestion probability and the unit traffic delay correction factor.

[0082] The mean square error is calculated based on the predicted traffic congestion probability and the unit traffic delay correction factor, along with the actual traffic parameter labels. The parameters are then updated iteratively through backpropagation.

[0083] Specifically, firstly, time-series data at the path segment level is extracted from AGV operation logs and task scheduling records, including metrics such as passage frequency, average occupancy time, and task waiting latency. Within a fixed-length time window, the frequency of actual AGV passage, the average occupancy time interval, and the waiting latency for the path to become passable in task scheduling are statistically analyzed for each edge (i.e., channel segment). This yields labeled data representing the actual path load status, including the probability of passage congestion (e.g., the proportion exceeding the concurrency threshold per unit time) and a unit passage latency correction factor (e.g., the latency deviation coefficient compared to an idle path). Subsequently, the graph convolutional representation of the edges is used as model input and fed into the second regression output unit for forward propagation. This regression unit consists of several linear layers and nonlinear activation functions stacked together, used for deep representation learning and nonlinear mapping processing of the input features, ultimately outputting predicted values, including two continuous values: the probability of passage congestion and the unit passage latency correction factor. During training, the system uses the predicted values ​​and the aforementioned statistically obtained true traffic parameter labels as supervision signals. It measures the prediction error using the mean squared error function as the loss term and incorporates an L2 regularization term to constrain the complexity of the model parameters, preventing overfitting. Through gradient backpropagation, the system updates the model parameters layer by layer according to the loss function, iterating until the loss converges. After training, the second regression output unit can output the traffic congestion probability and unit traffic delay correction factor for a path segment based on the side graph convolution features within any time period. This provides a dynamic scheduling basis for subsequent path selection, thereby avoiding potentially congested paths and improving the overall AGV scheduling efficiency while considering path length.

[0084] Furthermore, the step of filtering the entire set of cargo locations to obtain a feasible cargo migration candidate set based on the cargo location value score and the current cargo location occupancy status includes the following steps:

[0085] Based on the location value score and current occupied capacity of each location, all locations are traversed. Locations with a value score higher than the preset threshold and not full capacity are selected as candidates for shelving, while locations with a value score lower than the preset threshold and inventory waiting to be moved are selected as candidates for moving out.

[0086] Based on the SKU compatibility rules, combinations that do not meet compatibility or traffic constraints are eliminated from the candidates for shelf placement and candidates for relocation, resulting in a set of candidates for cargo relocation.

[0087] Specifically, the system first collects real-time occupancy information for each storage location in the current warehouse and combines this with a storage location value rating matrix constructed based on historical inbound / outbound frequencies and AGV accessibility assessments. Each storage location is modeled as a node, with attributes including its value rating and remaining capacity. Based on this, the system iterates through the entire set of storage locations one by one. If a storage location's value rating exceeds a set high-value threshold and its current remaining capacity is greater than zero, it is marked as a candidate storage location for shelving. Conversely, if the value rating is below a set low-value threshold and the location still has unused inventory SKUs, it is marked as a candidate storage location for relocation. This dual-threshold strategy helps quickly identify storage location combinations with significant migration potential without introducing large-scale global sorting overhead, avoiding ineffective handling. To further improve the practical feasibility of the candidate set, an SKU compatibility rule is introduced as a filtering condition. This rule constructs an incompatibility mapping table based on indicators such as storage constraints, picking method conflicts, and physical attribute differences between different SKUs. Simultaneously, based on the AGV path connectivity graph, the system verifies the reachability of the passageways between available shelving locations and available relocation locations. If a path has traffic constraints (such as high probability of congestion, path delay exceeding a preset threshold, or obstacle closure), the relocation pair is excluded. Finally, through this layer-by-layer screening process, the system obtains a set of candidate goods relocation locations that meet the relocation conditions. Each item in this set is a binary pair consisting of an available relocation location and its corresponding compatible available shelving location, possessing both feasibility and economic efficiency in performing the relocation action. This set can be used by the subsequent scheduling strategy module when making warehouse layout optimization, path planning, or wave merging decisions.

[0088] Furthermore, the reinforcement learning model is constructed through the following steps:

[0089] The cargo migration candidate set, traffic congestion probability and unit traffic delay correction factor are normalized and encoded into a state vector, and the state space of reinforcement learning is constructed through the state vector.

[0090] Each feasible migration or shelving operation in the candidate goods migration set is set as an independent action and assigned a unique action code. The action space of reinforcement learning is constructed through the set of action codes.

[0091] In the simulation execution, the corresponding travel time and congestion delay are extracted from the action encoding and the state vector to construct a reward function;

[0092] The state space, action space, and reward function are used to construct the training input of the reinforcement learning model. The policy network parameters are iteratively optimized to output the cargo location allocation policy, which is a set of action codes.

[0093] In some embodiments, firstly, the candidate set of goods migration obtained in the previous stage, along with its corresponding traffic congestion probability and unit traffic delay correction factor, is normalized to unify its data scale and prevent numerical instability caused by different dimensions in the subsequent learning process. The normalized data is then concatenated through feature splicing to form a state vector, which includes the ID encoding of the outgoing and incoming storage locations, path connectivity indicators, traffic resistance factors, and predicted value differences, constituting the state space for the reinforcement learning model. This state vector not only expresses the current state of the warehousing environment but also reflects the potential impact of potential migration actions on system structure optimization. Simultaneously, each pair of outgoing-incoming storage locations in the candidate set of goods migration is mapped to a unique action, and an action encoding set is constructed. Each action encoding includes the source storage location ID, the target storage location ID, and the SKU attribute ID, forming a discrete reinforcement learning action space. During training, each action code is input into the simulation environment to simulate the migration action. The shortest path for the AGV under this action is calculated using a path planning algorithm. Combined with the traffic congestion probability and unit delay correction factor, the actual travel time and waiting delay of this operation are accumulated as feedback for the action execution. Based on the above feedback signal, a composite reward function is constructed to measure the impact of a single cargo migration operation on the overall warehousing efficiency. This reward function can be defined as a weighted negative value of travel time and waiting delay, supplemented by the degree of improvement in the total value of the system's cargo locations after migration as an adjustment term, to achieve the dual optimization goals of operational efficiency and space utilization. To prevent the strategy from getting trapped in local optima, an entropy regularization term is introduced during training to maintain the exploratory nature of the strategy output. Finally, the state space, action space, and reward function are used as inputs to the reinforcement learning model. The policy gradient method or the proximal policy optimization (PPO) algorithm is used to train the policy network, iteratively updating the network parameters. The policy network structure adopts a multi-layer feedforward neural network, combining the state vector to output the action probability distribution, and learns the policy function of cargo migration operation by maximizing the expected cumulative reward. After training, the system outputs a set of corresponding action codes based on the current state vector input, which serves as the cargo allocation strategy for the current scheduling cycle, enabling autonomous intelligent decision-making in complex dynamic environments.

[0094] Furthermore, the job scheduling module is used to perform the following steps:

[0095] Based on the action codes output by the cargo location allocation strategy, the corresponding cargo migration tasks are parsed and generated based on the start and end nodes of each task and the channel connectivity.

[0096] The task path graph is input into the path planning algorithm, and path search is performed based on path length and connectivity between nodes. In the case of multiple tasks in parallel, conflict detection is performed to obtain a path scheme that meets the time constraints.

[0097] The AGV scheduling instruction set is generated according to the route scheme, and the AGV scheduling instruction set is sent to the corresponding AGV terminal to drive the AGV to complete the cargo transfer operation.

[0098] Specifically, the job scheduling module acts as a bridge between the location allocation strategy and the specific execution instructions. Its core function is to transform action codes into executable scheduling paths for the AGV, and to avoid path conflicts under multi-task concurrency. The system first parses the action codes output from the location allocation strategy. Each action code contains a source location ID, a target location ID, and a corresponding SKU number. By querying the location layout map and cargo storage records in the warehouse environment, the current location of the cargo and the target storage point are determined, thus generating migration tasks. Further, the starting and ending nodes of each migration task are mapped, and a task path graph is constructed by combining the topology of the warehouse's internal passageways (including main roads, branch intersections, and crossroads). The task path graph is a weighted directed graph, where nodes represent passage points, and edge weights are determined by a combination of path length and passage status. Subsequently, the task path graph is input into a path planning algorithm for solution. In this embodiment, an improved A* heuristic path search algorithm is preferred. By introducing a node passage priority function and path congestion weight to dynamically adjust the heuristic factor, the global optimality and real-time feasibility of path selection are enhanced. During path search, the system evaluates the total cost function of each feasible path. This function considers not only geometric distance but also incorporates the congestion probability predicted by the second regression unit and the unit travel delay correction factor, enabling dynamic avoidance of potential delays in the real scheduling environment. In multi-task concurrent scenarios, the path search process also needs to embed conflict detection and temporal constraint modeling. Based on the spatiotemporal scheduling graph, the system pre-allocates the access time windows of nodes in the path and uses a window sliding-based conflict detection mechanism to determine whether different AGVs have resource competition in similar time periods at the same node or edge. Once a path conflict is detected, the system will automatically trigger local path replanning, prioritizing the adjustment of the path start time or selecting alternative channels to ensure conflict-free execution of work paths among AGVs. Finally, the optimized path scheme is translated into an AGV scheduling instruction set. This instruction set includes AGV number, start and end node sequence, travel time of each path segment, node stopping actions, and transport commands for the executed SKU. All scheduling instructions, after being encapsulated by a protocol, are sent to the corresponding AGV control terminals via wireless communication modules, enabling AGVs to complete cargo transfer operations according to a unified rhythm and path strategy. This solution effectively solves the problems of path congestion and scheduling conflicts in traditional operation scheduling, improving the overall traffic efficiency of the warehousing system and the collaborative operation capability of AGVs.

[0099] Furthermore, the path planning algorithm is a heuristic search algorithm.

[0100] It should be noted that the path planning algorithm employs a heuristic search framework to address the multi-AGV scheduling path calculation problem in complex warehouse environments. This heuristic search algorithm is based on an improved A* algorithm, combining warehouse topology maps, traffic state prediction results, and task scheduling constraints to achieve efficient path generation. Specifically, the system first constructs a directed graph structure based on the task path graph. Nodes in the graph represent key locations within the warehouse (such as intersections, loading points, entrances / exits, etc.), and edges represent passable paths. The weight of each edge comprehensively considers dynamic factors such as path length, traffic congestion probability, and a unit traffic delay correction factor. During the path search process, the heuristic search algorithm evaluates its expandable neighboring nodes centered on the current node and uses a cost function f(n) = g(n) + h(n) to guide the search process. Here, g(n) represents the cumulative true cost from the starting point to the current node, and h(n) is the estimated optimal cost from the current node to the target node. The heuristic function h(n) uses a path distance estimate based on Manhattan distance or Euclidean distance and is dynamically adjusted using a unit delay correction factor to adapt to the current traffic conditions. In areas with severe path congestion or numerous branch nodes, the weight of the delay factor in the path heuristic function is increased to avoid hotspot areas and minimize latency. Furthermore, the algorithm introduces a closed node set and an open node priority queue mechanism, prioritizing the expansion of nodes with the smallest f(n) value to ensure rapid convergence of the search process towards the target direction. To enhance the coordination of multiple AGV paths, this embodiment also integrates a path adjustment module based on spatiotemporal conflict detection. Each time a path is expanded, the system detects whether the new path node conflicts with other AGVs assigned tasks in both spatial and temporal dimensions. If a conflict exists, path reordering or partial rollback is performed to ensure the stability of path scheduling in a multi-task environment.

[0101] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1.A warehouse intelligent management system based on deep learning, characterized in that, Comprise: A data processing module for constructing a space-time graph model containing nodes and edges based on the spatial distribution information of storage locations and functional locations in the warehouse and the channel connectivity relationship, mapping historical warehouse entry and exit records, AGV trajectory data and task scheduling logs into node and edge features of the space-time graph model according to time sequence, and obtaining a multi-dimensional space-time feature sequence; A storage location evaluation module for time series modeling and graph structure learning on the multi-dimensional space-time feature sequence, outputting storage location value scores corresponding to each node, passing congestion probabilities corresponding to each edge, and unit passing time delay correction factors; A storage location allocation module for filtering the entire set of storage locations based on the storage location value scores and the current storage location occupancy state to obtain a candidate set of goods migration, and constructing a reinforcement learning state vector combining the passing congestion probability and the passing time delay correction factor, and outputting a storage location allocation strategy through a reinforcement learning model; A job scheduling module for generating an AGV scheduling instruction set through a path planning algorithm according to the storage location allocation strategy, and issuing the scheduling instruction set to the corresponding AGV terminal to execute the job; The space-time graph model containing nodes and edges based on the spatial distribution information of storage locations and functional locations in the warehouse and the channel connectivity relationship comprises the following steps: Based on the storage location and functional location numbers in the warehouse management system, the coordinate information is extracted to construct the node set in the graph model, and the connectivity relationship between nodes is calculated according to the passable path logic to form the edge set and label the path length and passing direction; According to the SKU number, storage location number and timestamp contained in the historical warehouse entry and exit records, the time sequence entry and exit frequency of each node is counted as the node dynamic feature by corresponding mapping into the node set; According to the path sequence and time period recorded in the AGV trajectory data and task scheduling log, the passing frequency and average occupancy time per unit time are counted as the edge dynamic feature by corresponding mapping into the edge set; The node set, edge set and edge dynamic feature are time-aligned and normalized to construct a space-time graph model; The storage location evaluation module is used to perform the following steps: Based on the time sequence feature vector of the node, the feature segments of adjacent time steps are extracted according to the fixed length of the time window, one-dimensional convolution operation and normalization processing are performed on the segments to obtain the time convolution representation of each node; The matrix product operation is performed on the time convolution representation of the node and the node adjacency matrix, and the nonlinear activation processing is performed after weighted summation to obtain the node graph convolution representation containing neighbor relationship; Based on the time sequence feature vector of the edge, the passing frequency and occupancy time of each edge within the time window are processed by sliding convolution and standardization to obtain the time convolution representation of the edge; The feature splicing is performed on the time convolution representation of the edge and the node graph convolution representation corresponding to the start and end node index connected by the edge, and the weighted summation is performed through the weight matrix to obtain the graph convolution representation of the edge; The node graph convolution representation is input into the first regression output unit to output the storage location value scores corresponding to each node; the graph convolution representation of the edge is input into the second regression output unit to output the passing congestion probability corresponding to each edge and the unit passing time delay correction factor. 2.The warehouse intelligent management system based on deep learning according to claim 1, wherein, The method comprises the following steps: The SKU code, operation time and storage location number in the historical warehouse entry and exit records are classified according to the timestamp and mapped to the corresponding node, the frequency of warehouse-out, the frequency of warehouse-in and the operation heat of each node in a fixed time window are calculated, and the time sequence feature vector of the node is generated; The starting position, path trajectory and time period recorded in the AGV trajectory data are mapped to the edge set in the graph, the passing frequency and average occupancy time of each edge in each time slice are calculated according to the execution time and task density in the task scheduling log, and the time sequence feature vector of the edge is generated; The time sequence feature vector of the node and the time sequence feature vector of the edge are time-aligned to obtain a multi-dimensional space-time feature sequence. 3.The warehouse intelligent management system based on deep learning according to claim 1, characterized in that, The first regression output unit is trained by the following steps: Based on the historical warehouse entry and exit records and the AGV actual operation log, the warehouse-out frequency, warehouse-in frequency and operation time of each storage location in a fixed time window are extracted, and the real value score label is generated according to the average passing time of the storage location in the AGV trajectory dense area, and a first training sample set containing the training node graph convolution representation and the real value score label is constructed; The training node graph convolution representation in the first training sample set is input into the first regression output unit to perform matrix multiplication and nonlinear transformation, and the predicted storage location value score is output; The predicted storage location value score and the real value score label in the training sample are calculated by mean square error, and a loss function is formed by combining regularization constraints, and the parameters are updated by back propagation iteration. 4.The warehouse intelligent management system based on deep learning of claim 1, wherein, The second regression output unit is trained by the following steps: Based on the AGV operation log and the task scheduling record, the passing frequency, average occupancy time and waiting time delay of each edge in a fixed time window are calculated, and the actual passing congestion probability and unit passing time delay correction factor are calculated as real passing parameter labels, and a second training sample set containing the training graph convolution representation of the edge and the real passing parameter labels is constructed; The training graph convolution representation of the edge is input into the second regression output unit to perform numerical operation and nonlinear transformation, and the predicted passing congestion probability and unit passing time delay correction factor are output; The predicted passing congestion probability and unit passing time delay correction factor and the real passing parameter label are calculated by mean square error, and the parameters are updated by back propagation iteration. 5.The warehouse intelligent management system based on deep learning of claim 1, wherein, The feasible goods migration candidate set is obtained by screening the whole set of storage locations based on the storage location value score and the current storage location occupancy state, which comprises the following steps: All storage locations are traversed according to the storage location value score and the current occupancy capacity of each storage location, the storage locations with a storage location value score higher than a preset threshold and a capacity not full are selected as the up-shelf candidates, and the storage locations with a storage location value score lower than a preset threshold and a stock to be migrated are selected as the migration-out candidates; The combinations of the up-shelf candidates and the migration-out candidates that do not meet the compatibility or traffic constraints are removed according to the SKU compatibility rules, and a goods migration candidate set is obtained. 6.The warehouse intelligent management system based on deep learning according to claim 1, wherein, The reinforcement learning model is built by the following steps: The candidate set of goods migration, the probability of passing congestion and the unit passing time delay correction factor are normalized and coded into a state vector, and a state space of reinforcement learning is built through the state vector; Each feasible migration or shelving operation in the candidate set of goods migration is set as an independent action, and a unique action code is assigned, and an action space of reinforcement learning is built through the set of action codes; The driving time and congestion delay corresponding to the action code and the state vector are extracted in simulation execution to build a reward function; The state space, action space and reward function are used to build the training input of the reinforcement learning model, the policy network parameters are iteratively optimized, and a storage allocation strategy is output, wherein the storage allocation strategy is a set of action codes. 7.The warehouse intelligent management system based on deep learning according to claim 6, characterized in that, The job scheduling module is used to perform the following steps: According to the action code output by the storage allocation strategy, the corresponding goods migration task is parsed, and a task path graph is generated based on the start and end nodes and the channel connectivity of each task; The task path graph is input into a path planning algorithm, path search is performed according to the path length and node connectivity, conflict detection is performed in the case of multiple task parallelism, and a path scheme satisfying the timing constraint is obtained; According to the path scheme, an AGV scheduling instruction set is generated, and the AGV scheduling instruction set is sent to the corresponding AGV terminal to drive the AGV to complete the goods migration job. 8.The warehouse intelligent management system based on deep learning according to claim 7, characterized in that, The path planning algorithm is a heuristic search algorithm.

Images