A deep learning warehouse location intelligent allocation method and system

Abstract

The application relates to the technical field of warehouse logistics automation, and particularly discloses a deep learning warehouse location intelligent allocation method and system, which comprises a data perception fusion module, a space-time feature extraction module, a multi-target decision module and a dynamic execution feedback module. Through the fusion of multi-dimensional space-time features and real-time operation data, the graph neural network and the time sequence convolution network are used to model the commodity correlation and the location accessibility, the multi-target evolutionary algorithm is used to optimize the picking efficiency, the warehouse area load balancing and the real-time adaptability, and finally the dynamic optimization and the self-adaptive adjustment of the location allocation are realized through the closed-loop feedback, so that the overall efficiency of the warehouse operation is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of warehouse logistics automation technology, specifically relating to a deep learning-based intelligent allocation method and system for warehouse locations. Background Technology

[0002] In the field of modern logistics and supply chain management, the level of intelligence and automation of warehousing systems is a key indicator for measuring enterprise operational efficiency. Warehouse location allocation, as a core aspect of warehouse operations, directly impacts order picking efficiency, inventory turnover rate, and overall operating costs.

[0003] Among them, deep learning-based intelligent warehouse location allocation aims to optimize storage strategies through data-driven methods. This method analyzes multi-dimensional information such as historical order data, product correlation, and warehouse layout to achieve dynamic optimization and adaptive adjustment of location allocation, thereby improving the overall efficiency of warehousing operations.

[0004] Existing technologies primarily rely on warehouse location allocation methods based on fixed partitions or simple rules. These methods fail to effectively perceive and model the complex and dynamically changing spatiotemporal dependencies within the warehouse. For example, when best-selling items are allocated to remote locations, it significantly extends the picking path for subsequent related orders. Furthermore, existing allocation strategies generally ignore the real-time locations and congestion levels of pickers and handling equipment, causing system efficiency to gradually decline as operations progress. Traditional methods have significant limitations in addressing multi-objective conflicts and adapting to dynamic environments, necessitating an intelligent warehouse location allocation solution that deeply integrates spatiotemporal characteristics with real-time operational status. Summary of the Invention

[0005] The technical problem this invention aims to solve is to overcome the shortcomings of existing warehouse location allocation methods, such as their inability to effectively model complex spatiotemporal dependencies, difficulty in adapting to dynamic operating environments, and insufficient ability to coordinate multi-objective conflicts. This invention provides a deep learning-based intelligent warehouse location allocation method and system. This system aims to achieve dynamic optimization and adaptive adjustment of location allocation by deeply integrating multi-dimensional spatiotemporal features with real-time operational status, thereby improving the overall efficiency of warehousing operations.

[0006] The technical solution of this invention is to construct a deep learning-based intelligent warehouse location allocation system. This system includes a data perception and fusion module, a spatiotemporal feature extraction module, a multi-objective decision-making module, and a dynamic execution feedback module. The data perception and fusion module is used to collect and integrate various data from warehouse IoT devices in real time, including historical order data, product master data, warehouse 3D layout data, real-time inventory data, real-time location and movement trajectory data of pickers and handling equipment, and congestion status data of work aisles. This module unifies heterogeneous data sources into a standardized data stream with consistent spatiotemporal reference through preset data cleaning rules and timestamp alignment algorithms.

[0007] The spatiotemporal feature extraction module is connected to the data-aware fusion module, used to extract and construct multidimensional feature representations from standardized data streams. This module includes an order product association submodule and a warehouse location spatiotemporal accessibility submodule. The order product association submodule uses a graph neural network to model the co-occurrence relationships of products in historical orders, learning the association strength weight between each product and other products. This weight is a floating-point number between 0 and 1. The warehouse location spatiotemporal accessibility submodule constructs a spatiotemporal graph structure with warehouse locations as nodes and paths as edges, and uses a temporal convolutional network to analyze the accessibility score of each warehouse location in different time slices. This score integrates information from three dimensions: path distance, expected travel time, and equipment load factor.

[0008] The multi-objective decision-making module receives association strength weights and storage location spatiotemporal accessibility scores from the spatiotemporal feature extraction module and performs multi-objective optimization calculations. This module first defines three core optimization objectives: the first objective is to maximize order picking efficiency by minimizing the overall picking path length between related products; the second objective is to balance warehouse load by controlling the inventory turnover variance within each shelf partition to not exceed a preset threshold; and the third objective is to dynamically adapt to real-time operational conditions by incorporating the picker's current location and equipment congestion status as decision variables into the optimization model. The multi-objective decision-making module employs a decomposition-based multi-objective evolutionary algorithm to optimize these objectives in parallel, ultimately outputting a Pareto-optimal candidate set of storage location allocation schemes.

[0009] The dynamic execution feedback module connects to the multi-objective decision-making module to select the final execution plan from the Pareto optimal candidate set and monitor the execution effect. This module includes a plan selection submodule and an online learning submodule. The plan selection submodule assigns different weights to the three optimization objectives based on the current operational strategy preferences, calculates the weighted total score of each candidate plan, and selects the plan with the highest total score as the final warehouse location allocation instruction to be issued to the warehouse management system. The online learning submodule continuously collects actual operational indicators after the plan is executed, including actual picking path length, task completion time, and equipment idle rate, and compares these actual performance data with the expected objectives to generate feedback signals for adjusting the model parameters in the spatiotemporal feature extraction module and the objective weights in the multi-objective decision-making module, thereby forming a closed-loop optimization system.

[0010] In a preferred embodiment of the present invention, the specific operation flow of the order product association submodule is as follows: First, the historical order sequence is constructed into a product co-occurrence graph, where nodes represent products and edges represent the frequency of a product appearing in the same order. Next, a graph attention network is used to encode the node features, and a multi-head attention mechanism is used to learn the representation vector of each product node in its neighbor context. Finally, by calculating the cosine similarity between product representation vectors and applying a sigmoid function mapping, association strength weights in the range of 0 to 1 are obtained.

[0011] Furthermore, the construction method of the warehouse location spatiotemporal accessibility submodule is as follows: The physical space of the warehouse is discretized into a grid composed of warehouse location units, with each grid unit corresponding to one warehouse location. Each time slice is defined as 15 minutes long, and a total of 96 time slices are planned to cover a 24-hour operation cycle. For each warehouse location in each time slice, its spatiotemporal accessibility score is calculated. This score is a weighted sum of three indicators: the reciprocal of the basic path distance, the reciprocal of the expected travel time, and the complement of the equipment load factor. The weighting coefficients are optimized and determined using a grid search method on historical data.

[0012] Furthermore, the execution steps of the decomposition-based multi-objective evolutionary algorithm in the multi-objective decision-making module are as follows: Initialize a population of 100 individuals, each representing a complete pool allocation scheme. Generate offspring individuals using simulated binary crossover and polynomial mutation operators. Transform the original multi-objective problem into multiple single-objective subproblems using the Chebyshev decomposition method, and assign a uniformly distributed weight vector to each subproblem. In each generation of evolution, calculate the fitness value of each individual on each subproblem, and retain superior individuals through a tournament selection mechanism. The evolutionary process continues until the maximum number of iterations (500) is reached; at this point, the non-dominated solution set in the population is the Pareto optimal candidate set.

[0013] Furthermore, the parameter update mechanism of the online learning submodule is as follows: The loss function is defined as the mean squared error between the expected and actual operating indicators. This loss function is minimized using the stochastic gradient descent algorithm, with a learning rate of 0.001 and a batch size of 32. The parameter update period is set to once after processing every 100 new orders to ensure the model can adapt to changes in the operating mode in a timely manner.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention effectively solves the adaptability problem of traditional methods in complex dynamic environments by constructing an intelligent warehouse location allocation system that deeply integrates spatiotemporal features and real-time operational status. The system utilizes graph neural networks and temporal convolutional networks to accurately model the correlation between commodities and the spatiotemporal characteristics of warehouse locations, providing a solid feature foundation for decision-making.

[0015] 2. The multi-objective evolutionary algorithm can simultaneously optimize multiple conflicting objectives such as picking efficiency, warehouse load balancing, and real-time adaptability, and provides diverse decision options through the Pareto optimal solution set. Closed-loop feedback and online learning mechanisms ensure that the system can continuously learn from actual operations and self-optimize, thereby maintaining a high level of allocation performance in the long term.

[0016] 3. The overall solution improves the intelligence level and operational efficiency of warehousing operations, and has good practical value and promotion prospects. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical architecture of the deep learning-based intelligent warehouse location allocation system proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of the order product association submodule and the warehouse location spatiotemporal reachability submodule in this invention; Figure 3 This is a logical flow diagram of the multi-objective decision-making module in this invention; Figure 4 This is a schematic diagram of the closed-loop optimization framework of the dynamic execution feedback module in this invention; Figure 5 This is a schematic diagram of the data flow and interaction relationships between the various modules of the system in this invention. Detailed Implementation

[0018] Example 1: This invention provides a deep learning-based intelligent warehouse location allocation system. Please refer to the appendix. Figure 1 The system's overall architecture consists of a data perception and fusion module, a spatiotemporal feature extraction module, a multi-objective decision-making module, and a dynamic execution feedback module, all connected sequentially. These modules work together to form a complete closed loop from data acquisition, feature analysis, intelligent decision-making to execution feedback. By deeply integrating multi-dimensional spatiotemporal features with real-time operational status, the system achieves dynamic optimization and adaptive adjustment of storage location allocation.

[0019] The data perception and fusion module is the system's fundamental data entry point. This module is responsible for real-time collection and integration of various data from IoT devices deployed within the warehouse. This data specifically includes historical order data, product master data, warehouse 3D layout data, real-time inventory data, real-time location and movement trajectory data of pickers and handling equipment, and congestion status data of work aisles. Historical order data records the product composition, order time, and completion status of all past orders. Product master data includes the physical dimensions, weight, storage temperature requirements, and classification information of the products. Warehouse 3D layout data accurately describes the coordinates and geometric relationships of shelves, storage locations, aisles, workstations, and other facilities in 3D space. Real-time inventory data dynamically reflects the types and quantities of products currently stored in each storage location.

[0020] Real-time location and movement trajectory data of pickers and handling equipment are collected through ultra-wideband positioning base stations deployed on the warehouse ceiling and tags carried by the equipment. Congestion status data of the work aisles is sensed by LiDAR sensors and vision sensors installed at key aisle nodes to quantify the smoothness of passage. The data perception and fusion module has a set of strict data cleaning rules pre-defined, defining methods for identifying and handling outliers for each data type. For example, equipment location data that clearly exceeds the physical boundaries of the warehouse is removed, and records with negative inventory quantities are marked and verified. Simultaneously, the module uses a timestamp alignment algorithm to unify event data from all heterogeneous data sources onto a global timeline with millisecond-level precision, ensuring a consistent spatiotemporal reference for subsequent analysis. Finally, the module outputs a standardized data stream with a unified format, high timeliness, and high consistency, providing reliable data input for subsequent modules.

[0021] The spatiotemporal feature extraction module is closely connected to the data perception and fusion module. The core function of this module is to extract and construct a multi-dimensional feature representation from standardized data streams that profoundly reflects the inherent patterns of warehouse operations. Please refer to the appendix. Figure 2 The spatiotemporal feature extraction module contains two core sub-modules: the order product association sub-module and the warehouse location spatiotemporal reachability sub-module.

[0022] The order product association submodule is specifically responsible for uncovering the inherent relationships between products. This submodule's operation begins with a deep analysis of historical order data. First, it constructs a product co-occurrence graph from the accumulated large-scale historical order sequences. In this graph, each node uniquely represents a product, and the edges between nodes represent the frequency with which two corresponding products have appeared in the same order. Higher frequencies indicate a greater likelihood that these two products are simultaneously in demand. Next, this submodule employs a graph attention network to perform deep learning on this product co-occurrence graph. Through its unique multi-head attention mechanism, the graph attention network can adaptively learn the representation vector of each product node within its first-order and even multi-order neighbors.

[0023] This process doesn't simply average neighbor information; instead, it assigns different importance weights to different neighbor nodes, thus capturing the association patterns between products more precisely. Finally, for any two products, this submodule quantifies the strength of their association by calculating the cosine similarity between their representation vectors after encoding through a graph attention network. The range of the cosine similarity is mapped through a sigmoid function and normalized to a floating-point number between 0 and 1, serving as the final association strength weight output. The closer this weight value is to 1, the stronger the association between the two products, and they should be placed as close as possible in warehouse location allocation to improve picking efficiency.

[0024] The warehouse location spatiotemporal accessibility submodule focuses on quantifying the accessibility of each warehouse location in the spatiotemporal dimension. This submodule is built upon the discretization of the warehouse's physical space. It discretizes the entire warehouse area into a fine-grained grid composed of numerous warehouse location units, with each cell uniquely corresponding to a physical warehouse location. To characterize changes over time, this submodule defines the concept of time slices, each lasting 15 minutes, with a complete 24-hour operating cycle divided into 96 consecutive time slices. For each warehouse location within each specific time slice, this submodule calculates a comprehensive spatiotemporal accessibility score. This score is a composite indicator that comprehensively reflects three dimensions of information: the basic path distance from the core warehouse operating area to the warehouse location, the expected travel time within the corresponding time slice, and the load factor of the handling equipment serving this area within that time slice. Specifically, the basic path distance is a fixed geometric distance; the expected travel time is predicted based on historical data and real-time congestion status; and the equipment load factor reflects the equipment's workload. The accessibility score is essentially a weighted sum of the three indicators mentioned above after normalization. The weighting coefficients are a set of fixed values ​​determined through repeated trials and optimizations on a large amount of historical operational data using a grid search method, aiming to make the score best predictive of actual operation time costs.

[0025] The multi-objective decision-making module is the core of the system's intelligent decision-making. This module receives two key inputs from the spatiotemporal feature extraction module: first, the inter-product association strength weights provided by the order product association submodule; and second, the spatiotemporal accessibility score for each storage location in each future time slice provided by the storage location spatiotemporal accessibility submodule. Based on these refined feature inputs, the multi-objective decision-making module performs complex multi-objective optimization calculations to generate a high-quality candidate set of storage location allocation schemes. Please refer to the appendix. Figure 3This module first clearly defines three interrelated yet potentially conflicting core optimization objectives. The first objective is to maximize order picking efficiency. This is achieved by minimizing the overall picking path length between related items. Specifically, for a newly received item, the system seeks storage locations that minimize the total expected picking path length between it and highly related items. The second objective is to balance warehouse load. This is achieved by controlling the inventory turnover variance within each shelf zone to not exceed a preset threshold. This avoids overcrowding in some areas while other areas remain idle, promoting the long-term stable utilization of warehouse resources. The third objective is to dynamically adapt to real-time operational conditions. This is achieved by directly incorporating the picker's current location and the real-time equipment congestion status as key decision variables into the optimization model, enabling the allocation scheme to respond to rapidly changing on-site conditions.

[0026] To achieve the synergistic optimization of the three objectives mentioned above, the multi-objective decision-making module employs a decomposition-based multi-objective evolutionary algorithm. The algorithm's execution involves a series of rigorous steps. First, the algorithm initializes a population of 100 individuals, each encoding a complete warehouse allocation scheme covering all items to be allocated. These initial schemes are typically randomly generated to maintain population diversity. Next, the algorithm enters an iterative evolutionary process. In each generation, the algorithm uses simulated binary crossover and polynomial mutation operators to operate on parent individuals to generate new offspring. The crossover operator combines the superior genes of the parent generation, while the mutation operator introduces new exploration directions, preventing the algorithm from prematurely getting trapped in local optima. To handle multiple objectives, the algorithm employs the Chebyshev decomposition method. This method decomposes the original three-objective optimization problem into several single-objective optimization subproblems. Each subproblem is defined by a uniformly distributed weight vector, representing its preference for the three original objectives. In each generation of evolution, the algorithm calculates the fitness value of each individual on each single-objective subproblem. Then, through a tournament selection mechanism, individuals with higher fitness are selected from the merged parent and offspring populations and retained for the next generation. This evolutionary process continues until a preset maximum number of iterations of 500 is reached. At this point, the solutions in the final population that are not dominated by other individuals in any objective constitute the Pareto optimal reservoir allocation candidate set. This candidate set provides multiple allocation strategies that achieve the best balance among different objectives.

[0027] The dynamic execution feedback module connects to the multi-objective decision-making module, responsible for translating optimization results into actual operational instructions and ensuring continuous system improvement. Please refer to the appendix. Figure 4This module comprises two key sub-modules: a solution selection sub-module and an online learning sub-module. The core task of the solution selection sub-module is to select a final warehouse location allocation solution from the Pareto optimal candidate set output by the multi-objective decision-making module. This selection process is not arbitrary but based on the specific operational strategy preferences of the current warehouse management. This sub-module allows managers to assign different priority weights to the three aforementioned optimization objectives. For example, during peak promotional periods, picking efficiency may be given higher weight; during off-peak periods, the balance of warehouse load may be more important. Based on the set weights, the solution selection sub-module calculates the weighted total score for each solution in the candidate set and selects the solution with the highest total score as the final decision. Subsequently, this decision is formatted into a standard warehouse location allocation instruction and issued to the warehouse management system via an application programming interface (API) to drive automated equipment or guide manual shelving operations.

[0028] The online learning submodule forms the feedback loop for the system's closed-loop optimization. This submodule continuously collects actual operational metrics data generated after the plan is executed. This data includes, but is not limited to: the actual path length walked by pickers during task execution, the actual task completion time from task assignment to completion, and the actual idle rate of handling equipment during task intervals. After collecting this actual data, the online learning submodule performs a detailed comparison with the target values ​​expected by the multi-objective decision-making module during decision-making. This comparison result is quantified as a feedback signal. Specifically, this submodule defines a loss function, typically using the mean squared error between the expected and actual operational metrics. The system employs a stochastic gradient descent algorithm to minimize this loss function.

[0029] The key parameters of the algorithm are set as follows: the learning rate is set to 0.001 to ensure the stability of parameter updates; the batch size is set to 32 to strike a balance between computational efficiency and gradient estimation accuracy. Parameter updates are not triggered in real-time, but rather after processing 100 new orders. This periodic update strategy ensures that the model can learn from the latest operational experience in a timely manner while avoiding system instability caused by overly frequent updates. In this way, the feedback signals generated by the online learning submodule are used to fine-tune the model parameters in the spatiotemporal feature extraction module and adjust the target weights used for scheme selection in the multi-objective decision-making module, enabling the entire system to continuously adapt to changes in operational models and form a self-improving intelligent closed loop. Please refer to the appendix. Figure 5 The diagram clearly illustrates the complete data flow and inter-module interactions, starting from data perception and fusion, through feature extraction, multi-objective decision-making, scheme execution, and data feedback, revealing the inherent logic of the system operating as an organic whole.

[0030] In the graph neural network processing of the order product association submodule, the update process of the node representation vector follows specific information aggregation rules. For each product node in the graph, its new representation vector is determined by the weighted sum of its own features and the features of its neighboring nodes, with the weights calculated through an attention mechanism.

[0031] Before outputting standardized data streams, the data-aware fusion module performs a rigorous data quality verification process. This process includes integrity verification, logical consistency verification, and range reasonableness verification. Integrity verification ensures that key fields are not missing; for example, orders must include a product list, and storage locations must be associated with coordinate information. Logical consistency verification ensures that data relationships are correct; for example, the quantity shipped cannot exceed the current inventory quantity. Range reasonableness verification ensures that numerical data is within a preset reasonable range; for example, product weights are not negative, and storage location coordinates are within warehouse boundaries. Any data record that fails verification is marked as an anomaly and triggers corresponding processing rules, which may include direct discarding, requesting a resend from the data source, or filling with historical averages. The specific strategy depends on the importance of the data field and the nature of the anomaly. The processed data, along with the verified data, enters the timestamp alignment stage.

[0032] Timestamp alignment algorithms are a key technology for ensuring spatiotemporal consistency in the data-aware fusion module. This algorithm first establishes a global clock with millisecond precision for the system. All connected IoT devices must include their local clock timestamps when reporting data. Since device clocks may drift, the alignment algorithm periodically synchronizes and calibrates with the global clock, calculating and compensating for clock deviations. For each batch of data to be processed, the algorithm maps it to the global timeline based on its accompanying timestamp and known clock deviation. Events very close on the global timeline are grouped into the same processing time window and treated as occurring simultaneously, laying the foundation for subsequent correlation analysis. After alignment, all data points within the data stream have a unified and comparable time reference.

[0033] The graph attention network in the order product association submodule employs a multi-head attention mechanism to enhance its learning ability. Specifically, this mechanism runs multiple independent attention computation processes in parallel, i.e., multiple heads. Each head learns the relationships between nodes in a different representation subspace. Afterwards, the outputs of all heads are concatenated or averaged to form the final, comprehensive representation vector of the nodes. This multi-head design allows the model to simultaneously focus on different aspects of information from neighboring nodes, such as simultaneously capturing functional and seasonal associations of products, thereby learning richer and more robust product association features. Ultimately, the association strength weights between products are calculated based on this comprehensive representation vector.

[0034] In the spatial discretization of the warehouse location spatiotemporal accessibility submodule, the size of the grid cell is a key parameter. This size is typically determined based on the actual warehouse layout, shelf dimensions, and warehouse location density to ensure that each location is accurately covered by one or more grid cells. For large locations spanning multiple grid cells, the system uses the grid cell containing its core point or center of gravity as its representative. The time-slicing strategy, with each slice lasting 15 minutes, is based on historical analysis of warehouse operation rhythms. This duration captures the macro-level trends in operational intensity throughout the day without incurring excessive computational overhead due to overly fine segmentation. The predictive model for expected passage time comprehensively considers factors such as the average passage time for the same historical time period, the currently perceived channel congestion status, and whether there are planned operational activities.

[0035] The decomposition-based multi-objective evolutionary algorithm in the multi-objective decision-making module uses Chebyshev decomposition as its core idea. This method associates subproblems corresponding to neighboring weight vectors, allowing optimization of one subproblem to contribute to the optimization of its neighboring subproblems. Each weight vector is a three-dimensional vector, corresponding to the weights of the three objectives: selection efficiency, load balancing, and real-time adaptability. These weight vectors are uniformly distributed in the three-dimensional objective space to ensure that the resulting Pareto front effectively covers the entire objective space. In the tournament selection mechanism, a binary tournament is typically used, randomly selecting two individuals and comparing their fitness on a specific subproblem; the one with the better fitness wins. This selection mechanism maintains both selection pressure and population diversity.

[0036] When calculating the weighted total score, the scheme selection submodule of the dynamic execution feedback module needs to normalize the three objective values ​​corresponding to each scheme in the Pareto candidate set. This is because the three objectives usually have different dimensions and numerical ranges. Normalization can eliminate the influence of dimensions, making the weighted sum meaningful. Commonly used normalization methods include min-max normalization, which maps each objective value to the interval between 0 and 1. The stochastic gradient descent algorithm of the online learning submodule does not use all historical data to calculate the gradient when updating parameters. Instead, it randomly selects a small batch from recently accumulated feedback data and uses the average gradient of this batch to approximate the overall gradient. This strategy greatly improves learning efficiency, enabling its application to large-scale data scenarios. The parameter update cycle is set to every 100 orders; this number is the result of a trade-off between model stability and adaptability. A cycle that is too short may cause model oscillation, while a cycle that is too long may cause the model to fail to keep up with operational changes.

[0037] Data interaction between system modules is achieved through clearly defined interface protocols. The data-aware fusion module provides a standardized data flow interface to the spatiotemporal feature extraction module, which specifies the data format, fields, transmission frequency, and protocol. The spatiotemporal feature extraction module provides a feature output interface to the multi-objective decision-making module, transmitting the association strength weight matrix and the spatiotemporal reachability score tensor. The multi-objective decision-making module provides a candidate solution set interface to the dynamic execution feedback module, containing a set of solutions and their corresponding objective function values. The dynamic execution feedback module provides an instruction interface to the warehouse management system and feeds back learning signals to the system internally. All these interfaces are designed with corresponding error handling mechanisms, such as retransmission or enabling cached data when data times out, ensuring that the system can still operate under degraded conditions in some abnormal situations.

[0038] Example 2: This example provides another specific implementation of a deep learning-based intelligent warehouse location allocation system. Its core lies in the in-depth development of alternative technical paths for the spatiotemporal feature extraction module and the multi-objective decision-making module, so as to demonstrate the flexibility of the technical solution of the present invention.

[0039] In the spatiotemporal feature extraction module, the order product association submodule can employ a correlation mining method based on time-series prediction as a supplement or alternative to the graph neural network approach. This method does not directly construct a product co-occurrence graph; instead, it treats the historical demand sequence of each product as a time series. The system first aggregates historical order data according to fixed time windows, generating a sequence of frequency of occurrence or demand for each product within different time windows. Then, a Long Short-Term Memory (LSTM) network model is used to model these time series. LTM, with its internal gating mechanism, can effectively capture long-term dependencies and periodic patterns in time series.

[0040] For any two products, the synchronicity and similarity of their demand patterns can be assessed by calculating the dynamic time-warped distance or Pearson correlation coefficient between their latent state vectors encoded by a Long Short-Term Memory (LSTM) network. This time-based correlation measurement method can identify product pairs that do not always appear in the same order but have highly consistent demand rhythms, such as different product categories whose sales increase simultaneously during weekly Friday promotions. The resulting time-series correlation strength can serve as an effective supplement to the spatial co-occurrence correlation strength output by the graph neural network, or, in certain scenarios, as an independent correlation feature input into a multi-objective decision-making module.

[0041] Accordingly, the storage location spatiotemporal accessibility submodule can employ an evaluation method based on real-time dynamic path planning. Unlike traditional methods based on fixed time slices and static weighted scoring, this method, upon receiving the list of storage locations to be assigned, simulates in real-time the dynamic path planning process from the locations of all currently active pickers or handling equipment to each candidate storage location. The simulation process fully considers the current global environment, including the real-time task queues of all equipment, known temporary obstacles in passageways, and traffic flow changes predicted by the system in the next few minutes. The path planning algorithm can employ the A* algorithm or its variants, which incorporates real-time traffic information, to calculate the estimated shortest time and path complexity to reach each candidate storage location under the current conditions. This estimated time itself is a very intuitive and accurate measure of accessibility. This method eliminates the overhead of pre-calculating and storing a large number of time slice scores, achieving fully on-demand, real-time accessibility assessment, and is particularly suitable for warehouse scenarios where the operating environment changes drastically and predictive models are difficult to accurately model.

[0042] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based intelligent warehouse location allocation system, characterized in that, include: The data perception and fusion module is used to collect and integrate various data from warehouse IoT devices in real time, including historical order data, product master data, warehouse 3D layout data, real-time inventory data, real-time location and movement trajectory data of pickers and handling equipment, and congestion status data of operation channels. Through preset data cleaning rules and timestamp alignment algorithms, it unifies heterogeneous data sources into a standardized data stream with consistent spatiotemporal reference. The spatiotemporal feature extraction module is connected to the data perception fusion module and is used to extract and construct multidimensional feature representations from standardized data streams. The spatiotemporal feature extraction module includes an order product association submodule and a warehouse location spatiotemporal reachability submodule; The multi-objective decision-making module receives the association strength weights and storage location spatiotemporal accessibility scores from the spatiotemporal feature extraction module and performs multi-objective optimization calculations. This module first defines three core optimization objectives: the first objective is to maximize order picking efficiency by minimizing the overall picking path length between related products; the second objective is to balance the load in the storage area by controlling the inventory turnover variance in each shelf partition to not exceed a preset threshold; and the third objective is to dynamically adapt to the real-time operating status by incorporating the current location of the picker and the equipment congestion status as decision variables into the optimization model. The multi-objective decision-making module uses a decomposition-based multi-objective evolutionary algorithm to optimize the above objectives in parallel, and finally outputs a set of Pareto optimal storage location allocation scheme candidates. The dynamic execution feedback module, connected to the multi-objective decision-making module, is used to select the final execution plan from the Pareto optimal candidate set and monitor the execution effect; The specific operation process of the order product association submodule is as follows: Construct a product co-occurrence graph from the historical order sequence, where nodes represent products and edges represent the frequency of a product appearing in the same order; Graph attention networks are used to encode node features, and a multi-head attention mechanism is used to learn the representation vector of each product node in its neighbor context. By calculating the cosine similarity between the product representation vectors and mapping it through an S-shaped function, the association strength weight in the range of 0 to 1 is obtained. The construction method of the spatiotemporal reachability submodule of the storage location is as follows: The physical space of the warehouse is discretized into a grid composed of storage location units, with each grid unit corresponding to one storage location; Each time slice is defined as 15 minutes long, and a total of 96 time slices are planned to cover the 24-hour operation cycle; For each storage location at each time slice, its spatiotemporal reachability score is calculated. This score is a weighted sum of three indicators: the reciprocal of the path base distance, the reciprocal of the expected travel time, and the complement of the equipment load factor. The weighting coefficients are optimized and determined on historical data using a grid search method.

2. The deep learning-based intelligent warehouse location allocation system according to claim 1, characterized in that, The dynamic execution feedback module includes a scheme selection submodule and an online learning submodule. The scheme selection submodule assigns different weights to the three optimization objectives based on the current operational strategy preferences, calculates the weighted total score of each candidate scheme, and selects the scheme with the highest total score as the final warehouse location allocation instruction to be issued to the warehouse management system. The online learning submodule continuously collects the actual operational indicators after the scheme is executed, including the actual picking path length, task completion time, and equipment idle rate, and compares these actual effect data with the expected objectives to generate feedback signals for adjusting the model parameters in the spatiotemporal feature extraction module and the objective weights in the multi-objective decision-making module, thereby forming a closed-loop optimization system.

3. The deep learning-based intelligent warehouse location allocation system according to claim 1, characterized in that, The execution steps of the decomposition-based multi-objective evolutionary algorithm in the multi-objective decision-making module are as follows: Initialize a population of 100 individuals, each representing a complete storage allocation scheme; use simulated binary crossover and polynomial mutation operators to generate offspring individuals; The Chebyshev decomposition method is used to transform the original multi-objective problem into multiple single-objective sub-problems, and a uniformly distributed weight vector is assigned to each sub-problem. In each generation of evolution, the fitness value of each individual on each subproblem is calculated, and excellent individuals are retained through a tournament selection mechanism. The evolutionary process continues until the maximum number of iterations, 500, is reached. At this point, the non-dominated solution set in the population is the Pareto optimal candidate set.

4. The deep learning-based intelligent warehouse location allocation system according to claim 2, characterized in that, The parameter update mechanism of the online learning submodule is as follows: The loss function is defined as the mean squared error between the expected and actual operating indicators. The stochastic gradient descent algorithm is used to minimize this loss function, with the learning rate set to 0.001 and the batch size set to 32. The parameter update period is set to be executed once after processing 100 new orders to ensure that the model can adapt to changes in the operating mode in a timely manner.

5. The deep learning-based intelligent warehouse location allocation system according to claim 2, characterized in that, The graph attention network employs a multi-head attention mechanism to enhance its learning ability. This mechanism runs multiple independent attention computation processes in parallel. Each head learns the relationships between nodes in a different representation subspace. Afterward, the outputs of all heads are concatenated or averaged to form the final comprehensive representation vector of the nodes.

6. The deep learning-based intelligent warehouse location allocation system according to claim 1, characterized in that, The calculation process of the spatiotemporal accessibility score integrates three key factors: static distance, dynamic time cost, and equipment resource status. The basic path distance is a fixed geometric distance, the expected travel time is predicted based on historical data from the same period and real-time congestion status, and the equipment load factor reflects the busyness of the equipment.

7. The deep learning-based intelligent warehouse location allocation system according to claim 3, characterized in that, The Chebyshev decomposition method transforms the original multi-objective problem into multiple single-objective sub-problems. Each weight vector is a three-dimensional vector, corresponding to the weights of the three objectives of picking efficiency, load balancing, and real-time adaptability. These weight vectors are uniformly distributed in the three-dimensional objective space to ensure that the final Pareto front can cover the entire objective space well.

8. A deep learning-based intelligent allocation method for warehouse storage locations, characterized in that, The deep learning-based intelligent warehouse location allocation system described in any one of claims 1-7 enables intelligent allocation of warehouse locations.

Images