Order clustering and sorting sequence optimization method for goods-to-person sorting system

Abstract

The invention relates to the field of logistics automation and storage control, and discloses an order clustering and sorting sequence optimization method for a goods-to-person sorting system. According to the method, order detail data containing order numbers and SKU numbers are acquired according to wave order, an SKU complete set is constructed, orders are mapped into binary feature vectors, and order similarity is measured by adopting cosine distance; determining a clustering cluster number under the constraint of order position parallel capacity, initializing a clustering center by adopting K-means + +, and then optimizing the clustering center by using an in-cluster dispersion objective function in combination with a particle swarm optimization algorithm to obtain an order clustering grouping result; and further calculating the similarity between the cluster center vector and the inter-cluster cosine, constructing a cluster access sequence according to a greedy iteration rule, taking the cluster access sequence as a sorting task issuing sequence, and generating an order bit task queue to be executed by a system. According to the invention, joint optimization of order grouping and execution sequence is realized, the parallel utilization rate of order positions is improved, and the switching burden of sorting positions is reduced.

Description

Technical Field

[0001] This invention relates to the field of logistics automation and warehouse control technology, specifically to an order clustering and picking order optimization method applied to a goods-to-person sorting system in an automated warehouse. Background Technology

[0002] In the "goods-to-person" sorting system of an automated warehouse, the conveyor system or stacking / handling equipment transports storage units, such as bins and pallets, to the picking station according to task instructions. Operators then pick the target goods from the picking station and place them into the order box at the order location. The operational efficiency of this type of system depends not only on the speed of a single picking action but also heavily on the order batch organization method and execution sequence: if the SKU overlap between adjacent orders is low, the types of storage units required to reach the picking station and the arrival rhythm will fluctuate more, leading to frequent storage unit switching, increased waiting time, and decreased picking station utilization; if the order batch division does not match the parallel capacity of the order locations, uneven occupancy of order locations can easily occur, with some order locations idle while others are congested, thus reducing system throughput.

[0003] To improve picking efficiency, existing technologies typically employ order batching and sorting strategies to organize workflows. Common solutions include: rule-based batching, such as grouping by order row count, arrival time, region, or fixed threshold; similarity-based heuristic grouping, using the degree of overlap in the SKU sets contained in an order as the merging criterion; and optimized grouping based on clustering or metaheuristic algorithms, mapping orders to feature vectors, performing clustering, and then combining greedy or other heuristic methods to generate a processing order. While these methods can improve local efficiency at a certain scale, they still have significant shortcomings under the actual constraints of "goods-to-person" systems.

[0004] On the one hand, order SKU structures typically exhibit high-dimensional, sparse, and discrete set characteristics. Within the same batch, there are many SKU types, but each order contains a relatively small number of SKUs. This makes traditional clustering-based grouping results highly sensitive to initial centers and parameter settings, easily leading to insufficient intra-cluster similarity or unstable clustering results. Consequently, the demand for storage units corresponding to picking locations fluctuates significantly, increasing job switching costs. On the other hand, existing methods rely heavily on statistical indicators or empirical rules when determining the number of clusters or batches, making it difficult to establish direct constraints with the number of order locations and their parallel processing capabilities. This can easily result in unclear batch-order location mappings or uneven load distribution, leading to low order location utilization and prolonged batch completion times. Furthermore, even after order grouping, many existing solutions often fail to systematically optimize the execution order between clusters to account for picking location switching costs. This results in significant SKU differences between adjacent clusters, further exacerbating storage unit switching frequency and waiting time.

[0005] Therefore, existing technologies urgently need to solve the following technical problems: In the "goods-to-person" sorting system of automated warehouses, under the conditions of high-dimensional sparseness of order SKUs within a wave, limited parallel capacity of order positions, and significant switching costs of storage units for picking positions, how to achieve effective clustering and grouping of orders and generate optimized picking execution order while satisfying the parallel constraints of order positions, so as to improve the SKU similarity within the cluster, reduce the switching costs between adjacent batches / clusters, balance the load of order positions, and improve the overall system throughput and picking position utilization. Summary of the Invention

[0006] To address the issues raised in the background section regarding the high-dimensional and sparse SKUs within a wave of goods-to-person sorting systems, limited parallel capacity of order positions, and significant switching costs for picking position storage units, existing order batching and sorting methods suffer from unstable clustering results, low SKU similarity within clusters, difficulty in matching the number of clusters with the order position capacity, and a lack of order optimization to address inter-cluster switching costs. These problems lead to uneven order position load, frequent picking position switching, and decreased system throughput. This invention proposes an order clustering and picking order optimization method for goods-to-person sorting systems. This method uses feature vectorization to represent wave order data and calculates the similarity between orders. Under the constraint of parallel order positions, it determines the number of clusters, employs an improved cluster center initialization and swarm intelligence optimization strategy to obtain stable and high-quality order clustering results, and constructs cluster center representations and inter-cluster similarity metrics to further generate a picking execution order that reduces inter-cluster switching costs. This achieves the goals of improving intra-cluster order similarity, reducing picking position switching frequency, balancing order position workload, and enhancing overall system picking efficiency and throughput.

[0007] The technical solution adopted in this invention is as follows: A method for order clustering and picking order optimization in a goods-to-person sorting system includes the following steps: S1. Obtain order detail data within a preset time period from the order management system according to the wave. The order detail data must include at least the order number and SKU number, and form a set of pending orders for that wave. S2. Count all SKUs appearing in the set of orders to be processed, construct the complete SKU set, and define a fixed-dimensional feature space based on the complete SKU set; S3. Map each order in the set of orders to be processed to a binary order feature vector in the feature space, where the values ​​of each dimension of the order feature vector are used to indicate whether the corresponding SKU is included in the order; S4. Calculate the cosine similarity between any two orders based on the order feature vectors, and convert the cosine similarity into cosine distance to obtain the order distance metric. S5. Determine the number of clusters under the parallel capacity constraint of order positions. The parallel capacity constraint of the order position must at least satisfy The number of orders should not exceed the number of order slots, and after clustering is completed, each cluster will be assigned to different order slots for parallel processing. S6. Using the K-means++ initialization strategy with cosine distance as the metric, select k initial cluster centers. Specifically: randomly select one order as the first initial cluster center; for the remaining orders, construct a selection probability distribution based on the minimum cosine distance between the order and the selected cluster centers, and iteratively select subsequent initial cluster centers until the desired cluster center is obtained. One initial cluster center; S7, will An initial set of cluster centers is encoded as individual particles in a swarm. The particle swarm optimization algorithm is used to iteratively update the particle positions to optimize the cluster centers. The fitness of the particle swarm optimization algorithm is given by the intra-cluster discreteness objective function. The intra-cluster discreteness objective function includes the aggregation calculation of the cosine distance between the order feature vectors within each cluster and the corresponding cluster centers, resulting in the optimized set of cluster centers. S8. Based on the optimized cluster center set, each order is assigned to the corresponding cluster according to the principle of minimizing the distance to the cosine, thus obtaining the order clustering grouping result, and the clusters are mapped to the order positions according to the allocation rules in step S5. S9. Aggregate the order feature vectors within each cluster to obtain the cluster center vector, and calculate the cosine similarity between any two clusters based on the cluster center vector; S10. Construct a cluster access sequence based on the cosine similarity between clusters. The construction method is as follows: determine the initial clusters; iteratively select the next cluster with the largest cosine similarity to the current cluster from the clusters that have not been added to the access sequence, until all clusters are added to the access sequence; use the cluster access sequence as the order of picking tasks issued by the goods-to-person sorting system. S11. Output the order clustering results and cluster access sequence, and generate the order task queue for delivery to the sorting system based on this.

[0008] Furthermore, the SKU set and order feature vector in steps S2 and S3 are constructed as follows: All SKUs appearing within a wave are grouped into a set. ; ; For any order Construct binary order feature vectors ; ; Among them, when the order Include hour ,otherwise ; In the formula, The SKU set represents the complete set of all SKUs that have appeared in the same batch of pending orders. SKU complete set The first in SKUs, of which For dimensional indexing, ; SKU complete set The number of elements, i.e. the number of SKU types appearing in this wave; "Order" refers to any order in the set of orders to be processed. For orders The order feature vector has a dimension of ; For orders The expanded form of the eigenvectors, which includes One component; For orders The eigenvectors in the th Components on the dimension.

[0009] Furthermore, in step S4, the cosine similarity and cosine distance between orders are calculated as follows: Any two orders and The cosine similarity is: ; The cosine distance is: ; In the formula, For the first One order; For the first One order; and For order indexing, ; For orders The corresponding order feature vector has a dimension equal to the size of the entire SKU set. ; For orders The corresponding order feature vector also has the same dimension. ; For orders With orders The cosine similarity is used to characterize the directional similarity between two orders in the SKU feature space; a larger value indicates a higher degree of similarity. ; For vectors The modulus is used to normalize the dot product; For vectors The modulus length; For orders With orders The cosine distance, obtained from the cosine similarity transform, is used to satisfy the distance metric that the smaller the distance, the higher the similarity. .

[0010] Furthermore, the number of clusters in step S5 Under the constraint of parallel capacity for order slots, it is determined as follows: ; In the formula, This indicates the number of orders pending processing in this wave; This indicates the number of order positions that can work in parallel in a goods-to-person sorting system; This indicates the maximum number of orders that can be processed simultaneously in a single parallel processing batch; This indicates the rounding up operation.

[0011] Furthermore, in step S6, the K-means++ initialization strategy selects the initial cluster centers as follows: Randomly select one order vector from the order set as the first cluster center; for any order vector that is not selected as a center... Define it to the set of selected centers. The minimum cosine distance is: ; The next cluster center is selected according to the following probability distribution: ; In the formula, Indicates the first The order feature vectors corresponding to each order are used as candidate vectors for initializing centers in K-means++. For order indexing, , This represents the number of orders in this wave. This represents a cluster center vector within the selected set of cluster centers; Indicates the first The set of selected cluster centers at the next selection is the set of vectors of the currently selected cluster centers. , The index is the count of the number of selected centers / iteration rounds; Represents the candidate order vector Gather at the selected center The minimum cosine distance, i.e. With sets The minimum cosine distance between all center vectors in the vector; Indicates the set of the selected center. All center vectors Take the result that minimizes the expression within the parentheses; Represents the candidate order vector The probability of being selected as the next cluster center; This indicates that the calculation is performed on all candidate order vectors. The summation result is used as a normalization factor to make the sum of the probabilities of all candidate points equal to 1. The index of the candidate order vector.

[0012] Furthermore, in step S7, the particle swarm optimization algorithm uses the following velocity and position update formulas to iteratively optimize the cluster centers: ; ; In the formula, For the first The particle in the first The velocity vector at the next iteration is calculated using the velocity update formula; The inertial weight is used to adjust the degree to which particles inherit the velocity of the previous iteration, affecting the global search and local convergence characteristics of the algorithm. No. The particle in the first The velocity vector at each iteration is used to control the update direction and step size of the particle position in the search space; This is an individual learning factor used to adjust the particle's trajectory towards its individual historical best position. The degree of closeness; This is a swarm learning factor used to adjust the particle's trajectory toward the swarm's global optimal position. The degree of closeness; , is a random number factor, consisting of independent and uniformly distributed random numbers within the interval [0,1], used to introduce random perturbations to enhance search diversity. Corresponding individual items, Corresponding group terms; For the first The particle in the iteration to the _ The individual's historical optimal position vector at this time is the particle position that makes the fitness function optimal during the historical iteration process. For the first The particle in the first The position vector at the next iteration. Used to characterize a set of candidate cluster centers; For the particle swarm optimization in the iteration to the th The global optimal position vector at this time is the position of the particle that makes the fitness function optimal among all the historical positions of all particles in the particle swarm. For the first The particle in the first The position vector at the next iteration is calculated using the position update formula; For particle indexing, , The number of particles; Index for iteration count, , This represents the maximum number of iterations or the number of iterations to terminate.

[0013] Furthermore, the intra-cluster discreteness objective function in step S7 is constructed as follows: Let the clustering results include The nth cluster, the th The center vector of each cluster is The set of order vectors within the cluster is The intra-cluster dispersion is: ; The fitness function is defined as follows: ; In the formula, Intra-cluster dispersion is the sum of the squared distances from all samples within a cluster to the center of their respective clusters, used to evaluate cluster compactness. The number of clusters represents the number of clusters into which the clustering result is divided. This is a cluster index used to identify the first... There are 5 clusters, with values ​​ranging from 1 to 2. ; Let be any order feature vector within the cluster, representing the set An element whose dimension is same; For order vectors with cluster center vector Distance metric between; The fitness function value is used in the particle swarm optimization algorithm to evaluate the quality of a set of cluster centers. It is a positive constant, used to avoid the denominator. A value of zero or too small will cause the value to be unstable.

[0014] Furthermore, the cluster center vector in step S9 is calculated as follows: For the There are n clusters, and the cluster center vector is: ; In the formula, Represents a set The number of order vectors in the middle; Let be any order feature vector within the cluster, representing the set An element whose dimension is same; Indicates the first A set of order vectors within a cluster.

[0015] Furthermore, in step S10, the cluster access sequence is constructed according to a greedy iteration rule: Let the set of clusters be The corresponding cluster center vector is Initialize by selecting any cluster as the first cluster of the sequence. ; for the first Selected clusters Select the next cluster from the unselected cluster set: ; Until an access sequence containing all clusters is obtained. ; In the formula, The set of clusters represents the set of all clusters obtained from clustering. For cluster set The various clusters within; This represents the number of clusters. To cluster The set of cluster center vectors with one-to-one correspondence, where Represents clusters The cluster center vector is used to characterize the SKU structure features of the cluster; To access the next cluster in the sequence, as described above Rules are selected from candidate clusters; The operator that takes the value of the argument that maximizes the objective function; Cluster It belongs to the set of unselected clusters and is a candidate cluster that can be selected as the next cluster. This is a candidate cluster index, used to represent a candidate cluster in the set of unselected clusters. ; The set of unselected clusters represents the set of clusters. Exclude clusters that have already been added to the access sequence. The remaining clusters, where \ represents the set difference operation; For the current cluster The cluster center vector; Candidate clusters The cluster center vector; , Representing vectors respectively and The modulus is used to normalize the dot product and form the cosine similarity.

[0016] Furthermore, the order task queue in step S11 is generated as follows: According to the cluster access sequence obtained in step S10, the clusters in the access sequence are mapped to the task distribution order of the goods-to-person sorting system in turn, and the order set in each cluster is converted into the task queue entry of the corresponding order position; wherein, the task queue entry contains at least the order position identifier, the cluster identifier, the order number set, and the SKU number set corresponding to the order number.

[0017] Compared with existing technical solutions based on rule-based batching, single similarity merging, or only order grouping without system-level sequence control, the order clustering and picking order optimization method for goods-to-person sorting systems proposed in this invention has at least the following beneficial effects: 1. Enhanced batch executability driven by parallel constraints of order positions: This invention introduces a parallel capacity constraint for order slots when determining the number of clusters, so that the number of clusters matches the number of order slots and the parallel processing capability of a single order slot. After clustering, each cluster is mapped to a different order slot for parallel processing, so that the order grouping results can be directly converted into an order slot task queue and used for on-site execution. This reduces the idle and congested phenomena of order slots caused by the mismatch between the number of groups and the order slot capacity, and improves the parallel utilization rate of order slots.

[0018] 2. Improving clustering stability and grouping quality for high-dimensional sparse SKU features: Orders, constructed from the complete set of SKUs, exhibit high-dimensional sparse binary feature vectors. Traditional random initialization can easily lead to cluster center concentration or cluster collapse, resulting in insufficient intra-cluster similarity and unstable results. This invention employs a K-means++ center initialization strategy based on minimum cosine distance to improve the dispersion of initial centers. Furthermore, particle swarm optimization is used for global search and iterative updates of cluster centers, reducing the sensitivity of the clustering process to initial centers and parameters, resulting in more stable clustering results, more homogeneous SKU structures within clusters, and improved order grouping quality.

[0019] 3. The correlation between the clustering objective function and warehousing operation costs has been strengthened: This invention uses the intra-cluster dispersion, calculated based on the cosine distance between the order vector and the cluster center, as the fitness criterion for particle swarm optimization. This allows cluster optimization to directly affect the consistency of the order SKU structure within the cluster, thereby making the clustering results more consistent with the operational characteristics of "orders in the same cluster sharing SKUs / sharing storage units" in goods-to-person sorting, providing a basis for reducing storage unit access switching.

[0020] 4. Inter-cluster order optimization results in a system-level task delivery order, reducing the switching burden: This invention not only obtains order clustering grouping results, but also further constructs a cluster access sequence based on the cosine similarity of the cluster center vector, and uses this access sequence as the order of picking tasks; through greedy iteration, it selects the next cluster with the highest similarity to the current cluster from the unselected clusters, so that the SKU structure between adjacent clusters is more similar, reducing the frequency of SKU switching and storage unit switching when switching picking positions between different clusters, and reducing the job fluctuations caused by waiting and switching.

[0021] 5. Improve system throughput and picking location utilization, and shorten wave completion time: By combining "parallel-executable grouping results + sequential control oriented towards switching costs", this invention can more effectively organize order batch processing in wave operations, making the order position load distribution more balanced, the picking position operation more continuous, reducing the waiting time caused by fluctuations in the arrival rhythm of storage units, increasing the effective operation ratio of picking positions, and improving the overall throughput capacity of the goods-to-person sorting system and shortening the wave completion time. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0023] Figure 1 This is the overall flowchart of the order clustering and picking order optimization method for goods-to-person sorting systems of the present invention; Figure 2 This is a schematic diagram showing the working objects and constraints of the goods-to-person sorting system of the present invention; Figure 3 This is a flowchart illustrating the order task queue generation and distribution process of the present invention. Figure 4 This is a flowchart illustrating the determination of the number of clusters k and the mapping of order position constraints in this invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] To address the issues raised in the background section regarding the high-dimensional sparseness of order SKUs within a wave, limited parallel capacity of order positions, and significant switching costs of picking position storage units in goods-to-person sorting systems, existing order batching and sorting methods suffer from unstable clustering results, low SKU similarity within clusters, difficulty in matching the number of clusters with the order position capacity, and a lack of order optimization for inter-cluster switching costs. This leads to uneven order position load, frequent picking position switching, and decreased system throughput. This embodiment proposes an order clustering and picking order optimization method for goods-to-person sorting systems.

[0026] like Figure 1 As shown, the order clustering and picking order optimization method for the goods-to-person sorting system includes the following steps: like Figure 2 As shown, the system objects of the goods-to-person sorting system described in this embodiment include orders, picking locations, storage units, and order locations, and the number of clusters is... Subject to order position parallel capacity Maximum batch processing capacity per order and the total number of orders per wave Constraints, etc.

[0027] Step S1, Acquisition of wave order data and construction of the set of orders to be processed: In this embodiment, order data originates from an Order Management System (OMS) or a Warehouse Management System (WMS). The system retrieves order detail data within a preset time period in waves and forms a set of orders to be processed. Order detail data uses "order lines" as the basic record unit, and each order line contains at least the following fields: 1. Order ID, used to uniquely identify a single order; 2. SKU ID, used to uniquely identify the Stock Keeping Unit; Optional fields also include: SKU quantity (Qty), order time (CreateTime), order priority, wave ID (WaveID), order status (Status), etc., used for filtering and expanding scheduling strategies.

[0028] The wavelet definition and generation rules are as follows: A wave is a centralized processing unit in a goods-to-person sorting system, used to merge multiple scattered orders received within a certain time window into a single batch for parallel picking. In this embodiment, waves can be generated in any of the following ways: a. Generation by time window: Confirmed orders within a preset time window, such as every 5 minutes or 10 minutes, are summarized to form a wave; b. Generation by threshold: Waves are generated when the number of pending orders reaches a preset threshold or the expected SKU coverage reaches a threshold; c. Generation by business rules: Waves are generated after dividing the order into categories such as store / region / temperature zone / promised delivery time. Regardless of the generation method used, the wave serves as the calculation boundary for subsequent clustering and sequence optimization, thereby ensuring that the calculation results are consistent with the on-site work rhythm.

[0029] The order details data were obtained and cleaned as follows: After retrieving order detail data from the same batch from OMS / WMS, the data undergoes consistency processing, including at least: a. Status filtering: removing order lines that have been cancelled, completed, or will not enter the sorting process; b. Duplicate and anomaly handling: merging duplicate order lines and removing or completing abnormal records with missing SKU numbers or order numbers; c. Order line merging: when the same SKU appears in multiple lines within the same order, merging them into one line (the quantity field can be summed) to ensure consistency in subsequent feature representations. After the above processing, the valid order detail set for that batch is obtained.

[0030] The construction and statistical output of the set of orders to be processed are as follows: The cleaned order details are aggregated by order number to form the set of orders to be processed in this wave: ,in This represents the number of orders pending processing in this wave.

[0031] Simultaneously, the size of the candidate set of the complete SKU set within each wave was also analyzed. This refers to the number of SKU types appearing within a wave; the number of SKU rows per order, i.e., the number of SKUs in an order, used to characterize the sparsity feature of "few SKUs per order, many SKUs per wave"; and the order-SKU association table, for example, using the order number as the key to associate a set of SKU numbers, which serves as the direct input for constructing the order feature vector in subsequent steps S2 and S3.

[0032] In summary, the purpose of step S1 is to establish a unified data processing boundary on a wave basis, so that subsequent clustering and picking order optimization can match the actual operating rhythm of the goods-to-person sorting system; to form structured and computable "order-SKU" association data, providing a foundation for subsequent order vectorization representation and similarity calculation; and to output the number of orders per wave. These statistics are used to determine the number of clusters under the parallel capacity constraint of order positions. "Provide the required input."

[0033] In this embodiment, the number of clusters The selection is not based on adaptive statistical indicators such as the contour coefficient, but is determined by the parallel capacity constraint of the order slots; therefore, the number of wave orders output in step S1 is... With system configuration parameters, i.e., the number of order slots Single order unit parallel limit Joint composition determined The key basis is to ensure that the clustering results can be directly mapped to the order position parallel processing task queue, avoiding the problem of "the clustering is good in algorithm but cannot be executed on site".

[0034] The engineering connection between step S1 and subsequent algorithm selection is as follows: the order-SKU association data obtained in step S1 typically exhibits the characteristics of "many SKU types within a wave, and relatively few SKUs in a single order," making orders appear as high-dimensional, sparse, and discrete features in the SKU space. Based on this data characteristic, cosine distance is subsequently used to measure order similarity, and K-means++ initialization is employed to improve the dispersion of initial cluster centers and reduce the risk of cluster collapse caused by random initialization. Simultaneously, particle swarm optimization is combined to perform iterative search in wavelet-level computational scenarios to meet the engineering requirements of the goods-to-person system for computational convergence speed and interruptibility.

[0035] Step S2: Construct the complete set of SKUs and define a fixed-dimensional feature space: Steps S2 and S3 are used to map wave orders to a fixed feature space with the entire SKU set as its dimension, forming binary order feature vectors and providing a unified input for cosine distance measurement. The purpose of step S2 is to uniformly number all SKUs appearing within the same wave, constructing a fixed-dimensional feature space so that each order can be represented in the same dimensional system and similarity / distance calculations can be performed. Since there are usually many types of SKUs in goods-to-person waves and order SKUs exhibit a set-like discrete characteristic, without first establishing the entire SKU set, it is difficult to compare different orders in a unified coordinate system, and their similarity measurement and clustering will lack a consistent dimensional basis.

[0036] The specific implementation method is as follows: The set of orders to be processed formed in step S1 is traversed, and all SKU numbers that have appeared in each wave are extracted and deduplicated to obtain the complete set of SKUs: ; In the formula, The SKU set represents the complete set of all SKUs that have appeared in the same batch of pending orders. SKU complete set The first in SKUs, of which For dimensional indexing, ; SKU complete set The number of elements, i.e. the number of SKU types appearing in this wave.

[0037] With SKU complete set Define a feature space of fixed dimensions, where the number of dimensions of the feature space is equal to... A mapping table of "SKU number - dimension index" is established to quickly map order details to the corresponding dimension.

[0038] Table 1. SKU Number-Dimension Index Mapping Table (based on the complete set of wave SKUs) (For example, =5): To ensure repeatability, this embodiment can generate sequences using one of the following methods. : Arrange the SKUs in ascending order and number them sequentially; or number them sequentially according to the order in which the SKUs first appear.

[0039] All of the above methods can ensure that the dimension definition is unique and traceable within the same wave.

[0040] Step S3: Map the order to a binary order feature vector. The purpose of step S3 is to represent each order as a high-dimensional sparse binary vector, so that the SKU set similarity between orders can be directly calculated through vector operations. The binary representation can highlight the set feature of "whether it contains a certain SKU", which is suitable for the operation logic of "storage unit arrival and SKU set coverage" in goods-to-person sorting; at the same time, this representation is naturally sparsity, which is conducive to the implementation of wavelet fast calculation and real-time scheduling scenarios.

[0041] The specific implementation method is as follows: For any order Based on the feature space established in step S2, construct the binary order feature vector: ; In the formula, For orders The order feature vector has a dimension of ; For orders The expanded form of the eigenvectors, which includes One component; For orders The eigenvectors in the th Components in the dimension; Let be any order in the set of orders to be processed, satisfying: ; In the engineering implementation, the order details can be scanned line by line: for each "Order Number - SKU Number" record, the corresponding dimension can be found through the mapping of "SKU Number - Dimension Index". and will Set to 1; keep it unchanged when the same SKU appears repeatedly in the same order. If the number of SKUs in the order is much smaller than... Sparse storage can be used, for example, storing only the set of dimension indices with a value of 1, to reduce memory usage and improve computational efficiency; however, mathematically it is still equivalent to the above-mentioned length of A binary vector.

[0042] To facilitate understanding, an example is given: Let the complete set of SKUs be... Then the order feature vector can be represented as: Order 1 includes ,but ; Order 2 includes ,but ; Order 3 only contains ,but .

[0043] The output of steps S2 and S3 is: the complete set of SKUs. SKU dimension mapping table, and a set of binary order feature vectors for all orders within a wave. In step S4, the similarity / distance metric between orders is calculated using the order feature vector as input, providing a distance basis for K-means++ initialization, cluster center optimization, and final clustering.

[0044] Based on steps S2 and S3, subsequent steps determine the number of clusters. The algorithm does not rely on statistical indicators such as silhouette coefficients, but is directly determined by the parallel capacity constraints of order positions, enabling each cluster to establish an executable mapping relationship with the parallel processing batches of order positions. Since the order vectors exhibit high-dimensional, sparse, binary characteristics, K-means++ is subsequently used for decentralized initialization of cluster centers, and particle swarm optimization is combined with wavelet optimization to achieve rapid iterative convergence in wavelet-level computation, thereby improving cluster stability and usability. The intra-cluster dispersion is then used as the evaluation criterion for cluster center optimization, and a cluster access sequence is further generated as the WCS / WES task distribution order, realizing a system-level control strategy oriented towards the picking position switching burden.

[0045] Step S4, Order Distance Metric Construction: After completing steps S2 and S3, this embodiment has represented each order in the same wave as a fixed-dimensional binary order feature vector. Step S4 is used to construct a distance metric between orders, which is used to determine the number of clusters in the subsequent process. The selection of initial cluster centers and the iterative optimization of cluster centers provide a unified basis for calculating "similar / dissimilar".

[0046] The purpose of step S4 is to characterize the proximity between orders in a high-dimensional sparse SKU feature space using a similarity index that conforms to the overlapping characteristics of order SKU sets, and further convert it into a distance metric that satisfies the K-means principle of "the smaller the distance, the higher the similarity". Since the SKU structure of goods-to-person sorting orders is usually a set-type discrete feature, with the number of SKUs per order being much smaller than the number of wave SKU types, directly using Euclidean distance is easily affected by dimensionality and sparsity, making it difficult to reflect the degree of overlap of order SKU sets. Using cosine similarity can characterize the consistency of order vector directions, which is equivalent to measuring the degree of overlap and structural similarity of order SKU sets, and is more suitable for subsequent clustering and task organization.

[0047] The methods for calculating cosine similarity and cosine distance are as follows: Let any two orders be respectively and Their order feature vectors are respectively and ,in , , The size of the entire SKU set. The cosine similarity between two orders is defined as: ; In the formula, For the first One order; For the first One order; and For order indexing, ; For orders The corresponding order feature vector has a dimension equal to the size of the entire SKU set. ; For orders The corresponding order feature vector also has the same dimension. ; For orders With orders The cosine similarity is used to characterize the directional similarity between two orders in the SKU feature space; a larger value indicates a higher degree of similarity. ; For vectors The modulus is used to normalize the dot product; For vectors The length of the module.

[0048] To adapt to the calculation logic of K-means clustering algorithm, which states that "the smaller the distance, the more similar the data," cosine similarity is converted to cosine distance: ; In the formula, For orders With orders The cosine distance, obtained from the cosine similarity transform, is used to satisfy the distance metric that the smaller the distance, the higher the similarity. .

[0049] In engineering implementation, to avoid a denominator of zero, it should be ensured that the order vectors involved in the calculation are non-zero vectors, that is, each order contains at least one SKU; if there are abnormal empty orders, they can be removed in the data cleaning stage of step S1, or a very small positive number can be introduced when calculating the modulus. Implement numerical protection.

[0050] The order similarity / distance matrix is ​​constructed as follows: To improve the computational efficiency of subsequent clustering iterations, this embodiment can calculate the similarity or distance between all order pairs and construct an n×n order similarity matrix. or distance matrix , where 𝑛 represents the number of orders in that wave. Taking the similarity matrix as an example, its elements are defined as: ; Accordingly, the elements of the distance matrix are: ; In practical systems, the following strategies can be selected based on the wave size: 1. When 𝑛 is small or system resources are sufficient, pre-calculate the complete matrix. or 1. Facilitates rapid table lookup during the clustering stage; 2. When n is large, an on-demand calculation method is adopted, calculating the cosine distance only for necessary order-center pairs in the cluster center update or allocation steps to reduce storage and computation overhead; 3. When the order vector is stored in a sparse structure, the dot product and modulus can be quickly calculated using the "dimensional set with a value of 1", improving the efficiency of wavelet-level real-time calculation.

[0051] Based on the cosine distance metric obtained in step S4, the number of clusters can be determined subsequently under the parallel capacity constraint of order positions. This provides the foundation for clustering. Unlike schemes that adaptively select the number of clusters based on statistical indicators such as silhouette coefficients, this embodiment subsequently determines... The parallel capacity constraint of order positions is directly introduced to establish an executable mapping relationship between the number of clusters and the number of order positions, ensuring that each cluster can be mapped to a batch input of parallel processing of order positions, so that the clustering results can be directly used for order position task queue generation and WCS / WES scheduling execution.

[0052] Since the order feature vector is a high-dimensional sparse binary vector, randomly initializing cluster centers can easily lead to multiple centers clustering in similar SKU regions or "cluster collapse," resulting in insufficient cluster stability. The cosine distance metric established in step S4 provides the basis for K-means++ to "construct a probability distribution based on the minimum cosine distance," enabling the initial centers to be distributed and covered in the SKU space, improving the initial segmentation quality and reducing fluctuations in subsequent iterations.

[0053] In this embodiment, intra-cluster dispersion is used as the fitness criterion for particle swarm optimization. Since cosine distance directly reflects the differences in order SKU structure, a smaller intra-cluster dispersion means a higher degree of overlap between order SKUs within the same cluster. In goods-to-person sorting operations, this corresponds to a more concentrated range of storage unit types that the picking station needs to reach when processing orders in that cluster, resulting in fewer access switches. Therefore, minimizing intra-cluster dispersion can be characterized in engineering terms as reducing the number of times the picking station switches between different storage units, establishing a direct relationship between the "mathematical objective function" and the "operational cost".

[0054] The cosine similarity / distance metric constructed in step S4 is not only used for order clustering, but also serves as the basis for subsequent calculation of cluster center similarity and construction of cluster access sequences. The cluster access sequence is subsequently used as the task distribution sequence input for WCS / WES to control the scheduling order and arrival rhythm of bins / pallets in the conveying system, thereby reducing load jumps and waiting caused by adjacent cluster switching at the system level.

[0055] Step S5: Determine the number of clusters under the parallel capacity constraint of order positions. : like Figure 4 As shown, when inputting the number of wave orders... Number of order slots And the maximum batch processing limit for a single order. Based on this, calculate the number of clusters. It also determines a one-to-one mapping or capability-based matching relationship between clusters and order positions to ensure that the clustering results can be executed directly in parallel.

[0056] After obtaining the cosine distance metric between orders in step S4, this embodiment proceeds to step S5, which is used to determine the number of clusters. This step is a key technical step that distinguishes this method from existing schemes that determine the number of clusters using statistical indicators or empirical rules.

[0057] The purpose of step S5 is to establish a direct constraint relationship between "order clustering" and the parallel processing capability of the order location in the goods-to-person sorting system, so that the number of clusters... Matching the number of order slots and their parallel capacity ensures that clustering results can be directly executed on the system side in a "cluster-order slot" manner. Specifically, goods-to-person sorting systems typically configure multiple order slots for parallel processing of order boxes. If the number of clusters is too large, the clustering results cannot form a clear mapping with the order slots, easily leading to task queue congestion and uneven order slot occupancy. If the number of clusters is too small, orders with significantly different SKU structures are forcibly merged, increasing the burden of picking slot switching and reducing parallel utilization. Therefore, this method transforms the determination of the number of clusters from "pure algorithm parameter selection" to "determination of scheduling executable parameters constrained by the parallel capacity of order slots," enabling each cluster to be mapped to a batch input for parallel processing of an order slot, thereby providing an executable foundation for subsequent clustering calculations, task queue generation, and WCS / WES deployment.

[0058] Number of clusters Method for determining: Let the number of orders to be processed in this wave be... The number of order locations that can work in parallel in a goods-to-person sorting system is: The maximum number of orders that can be processed simultaneously in a single order slot within a parallel processing batch is [number]. In this embodiment, the number of clusters... Under the constraint of parallel capacity for order slots, it is determined as follows: ; In the formula, This indicates the number of orders pending processing in this wave; This indicates the number of order positions that can work in parallel in a goods-to-person sorting system; This indicates the maximum number of orders that can be processed simultaneously in a single parallel processing batch; This indicates the rounding up operation.

[0059] By using this rule At least satisfy This ensures that each cluster has a "one cluster corresponds to one order" mapping at the parallel scheduling level; simultaneously, Used to characterize the capacity of a single order batch. Under the condition of completing wave order processing, the lower bound of the number of parallel batches required is determined so that the number of clusters is consistent with the order processing capacity.

[0060] In the project configuration, parameters The capacity can be preset by the system, given by order location specifications, or by operational strategies, such as the number of order boxes that the order location can accommodate simultaneously, the system-defined parallel picking limit, or a safety threshold. If the system does not explicitly set the single order batch capacity, it can be used in one optional implementation. =1, then the cluster number is determined to degenerate into It still satisfies the parallel constraint objective that "the number of clusters does not exceed the number of order slots".

[0061] Cluster-order bit mapping and parallel processing rules: In determining Subsequently, in this embodiment, upon completion of clustering, each cluster is assigned to different order positions for parallel processing, forming a cluster-order position mapping relationship. Specifically, one of the following methods can be used: 1. One-to-one mapping: when or At that time, select from the order position set One order slot, and 1. One-to-one allocation among clusters; 2. Constrained allocation: If order slots have different processing capabilities or priorities, clusters can be matched and allocated according to the order slot's capabilities, provided that "one cluster corresponds to one order slot"; 3. Backup order slot strategy: When At that time, unassigned order slots are used as backups or for order insertion and exception handling to maintain the stability of wave execution.

[0062] Through the above mapping, the clusters are no longer just the result of the algorithm grouping, but parallel processing units that can be identified by WCS / WES and used as task queue entries to generate orders, providing structured input for the subsequent system-level control strategy of "cluster access sequence as task issuance order".

[0063] Step S5 determines This serves as the input for the center count in subsequent K-means++ initialization. Since the order feature vector is a high-dimensional sparse binary vector, random initialization is performed on a given... Clustering is prone to problems such as center concentration, cluster collapse, and unstable clustering. K-means++ can be used to select centers based on a probability mechanism of "farthest from the selected center" using cosine distance. The initial centers are distributed more widely to cover the SKU space, improving the performance in a fixed manner. Constrained clustering availability and stability, avoiding the disruption of the engineering goal of "cluster-order position" parallel mapping due to cluster load imbalance caused by unreasonable centers.

[0064] exist Once the parallel capacity constraint of the order location is determined, the goal of cluster optimization is no longer to find an "optimal number of clusters" in any statistical sense, but rather to obtain high-quality cluster centers that satisfy the job constraints as quickly as possible under a fixed number of clusters. The wave scheduling of the goods-to-person system has real-time requirements, necessitating fast algorithm convergence and the ability to output executable results within a limited time. This method uses particle swarm optimization to iteratively update the cluster centers. Compared to methods such as genetic algorithms, it has fewer parameters and a computational structure more suitable for rapid iteration and mid-process termination, making it suitable for wave-level real-time computation.

[0065] In subsequent steps of this invention, intra-cluster dispersion is used as the clustering optimization objective. Under the cosine distance metric, a smaller intra-cluster dispersion indicates a more consistent SKU structure and higher SKU overlap within the same cluster. Corresponding to goods-to-person sorting operations, this means that the type of storage units required by the picking station when processing orders from that cluster is more concentrated, and the picking station has fewer switching requirements between different storage units. Therefore, under fixed... Under the premise of satisfying the parallel constraint of order positions, minimizing the intra-cluster dispersion can be equivalently represented as reducing the number of times the picking position switches to different storage units, thus establishing a consistency constraint between the "mathematical objective function" and the "engineering operation cost".

[0066] Since step S5 ensures that "each cluster can be mapped to order positions for parallel processing", the subsequent construction of the cluster access sequence is no longer a simple sorting output, but is directly used as the sequential input for WCS / WES task issuance. This is used to control the scheduling rhythm and arrival order of bins / pallets in the conveying system, reduce the load jump caused by switching between different SKU-dense areas of the picking position, and thus achieve joint executable optimization of grouping and ordering.

[0067] Step S6: Initialize the selection using K-means++ under the cosine distance metric. Initial cluster centers: In step S5, the number of clusters has been determined based on the parallel capacity constraint of the order position. After establishing the cosine distance metric between orders in step S4, this embodiment proceeds to step S6, which is used to select orders based on the cosine distance metric. The initial cluster centers provide a stable initial state for subsequent cluster center optimization and order allocation.

[0068] The purpose of step S6 is to select dispersed and representative initial cluster centers in the high-dimensional sparse SKU binary feature space, reduce the sensitivity of the clustering process to the initial centers, avoid center clustering, cluster collapse, and insufficient intra-cluster similarity caused by random initialization, thereby improving the stability and repeatability of the clustering results. Since the SKU structure of goods-to-person sorting wave orders typically exhibits "high dimensionality, strong sparsity, and significant discrete set characteristics," random initialization, under the cosine distance metric, easily leads to multiple centers falling into nearby SKU regions, resulting in some cluster gaps or unstable inter-cluster partitioning, thus violating the executable constraint of step S5, "each cluster can be mapped to one order position for parallel processing." K-means++ initialization selects centers by "constructing a probability distribution based on the square of the minimum cosine distance," making subsequent centers more likely to originate from orders far from the selected centers, forming a more dispersed center coverage in the SKU space, and improving the convergence quality and stability of subsequent iterations.

[0069] The specific implementation method for K-means++ initialization is as follows: Suppose that the set of wave orders contains 𝑛 order vectors. ,in The binary order feature vector constructed in step S3; The number of clusters determined in step S5. The K-means++ initialization process includes the following sub-steps: First, randomly select the first cluster center: randomly select an order vector from the set of order vectors as the first initial cluster center, denoted as: ; in The randomly selected order vector; initialize the selected cluster center set as follows: ; Then, calculate the minimum cosine distance from the candidate order to the set of selected centers: for any order vector that was not selected as a center. Define it to the currently selected center set. The minimum cosine distance is: ; In the formula, Indicates the first The order feature vectors corresponding to each order are used as candidate vectors for initializing centers in K-means++. For order indexing, , This represents the number of orders in this wave. This represents a cluster center vector within the selected set of cluster centers; Indicates the first The set of selected cluster centers at the next selection is the set of vectors of the currently selected cluster centers. , The index is the count of the number of selected centers / iteration rounds; Represents the candidate order vector Gather at the selected center The minimum cosine distance, i.e. With sets The minimum cosine distance between all center vectors in the vector; Indicates the set of the selected center. All center vectors Take the result that minimizes the expression within the parentheses.

[0070] Then, construct a probability distribution based on the squared distance and select the next center: after obtaining all candidate orders... Then, the next cluster center is selected according to the following probability distribution: ; In the formula, Represents the candidate order vector The probability of being selected as the next cluster center; This indicates that the calculation is performed on all candidate order vectors. The summation result is used as a normalization factor to make the sum of the probabilities of all candidate points equal to 1. The index of the candidate order vector.

[0071] The denominator is the sum of the squared distances over all candidate order vectors, used for normalization to make the sum of probabilities equal to 1. A sampling is then performed based on this probability distribution to select an order vector. As the next cluster center And update the central set: ; Iterate until k initial cluster centers are obtained. Repeat the above steps until the number of selected centers reaches [number missing]. This yields the initial set of cluster centers: ; The above set This serves as the initial center input for subsequent cluster center optimization and order allocation.

[0072] Since the order vector is a binary non-negative sparse vector, the dot product... The modulus can be quickly calculated using the "number of dimensions that are all equal to 1". Pre-calculation and caching can reduce the computational overhead of the K-means++ initialization phase, meeting the real-time computation requirements of wavelet-level processing. If a sparse storage structure is used, the calculation of the minimum cosine distance can be accelerated by set intersection and union. When the data scale is large, batch sampling of the candidate set can be performed to further reduce the computational load without changing the algorithm principle. The number of centers selected in the initialization phase is strictly equal to the number determined in step S5. This ensures that the subsequent clustering results are structurally consistent with the parallel capacity constraints of the order positions, thus guaranteeing from the source that the subsequent clustering results are consistent with the constraints of the order positions in terms of structure.

[0073] The number of clusters in this method Instead of being determined through statistical indicators such as silhouette coefficients or the elbow rule, it is directly constrained by the parallel capacity of order positions, ensuring that each cluster can be mapped to a single batch input of parallel processing for an order position at the system execution level. Under this premise, the K-means++ initialization in step S6 is based on... Input the number of centers to ensure that the initial centers are distributed to cover the SKU space, and avoid the extreme unevenness of cluster size caused by unreasonable center distribution, which would undermine the feasibility and load balancing of the "cluster-order position" mapping.

[0074] Since SKU features are high-dimensional sparse binary vectors, random initialization can easily lead to multiple centers falling into the same dense SKU region, causing cluster collapse or cluster instability, resulting in insufficient SKU similarity within the cluster and uneven distribution of order load. K-means++ selects centers based on the probability mechanism of minimum cosine distance squared, ensuring that the centers remain dispersed in the high-dimensional sparse SKU space, thereby improving cluster stability and reducing result fluctuations, making it more suitable for the online generation and scheduling of goods-to-person wave tasks.

[0075] K-means++ provides dispersed and stable initial centers, making subsequent iterations less likely to fall into obviously inferior solutions. Based on this, step S7 introduces particle swarm optimization to iteratively update the centers, meeting the real-time scheduling requirements of the goods-to-person system for "fast convergence and interruptible output." Particle swarm optimization has fewer parameters and a computational structure more suitable for rapid iteration, making it easier to obtain an executable set of cluster centers within a wavelet-level time window compared to methods such as genetic algorithms.

[0076] When using intra-cluster dispersion as the fitness criterion, the dispersion of the initialization at the center can reduce the initial value of intra-cluster dispersion and reduce the number of iterations. At the same time, under the cosine distance metric, the smaller the intra-cluster dispersion, the higher the degree of overlap of the SKU set of orders in the same cluster, and the more concentrated the types of storage units that the picking position needs to access. From an engineering perspective, this is equivalent to reducing the number of times the picking position switches between different storage units, reflecting the direct correlation between mathematical objectives and operational costs.

[0077] Stable cluster centers and clustering results provide a basis for subsequent calculation of cluster center vectors and inter-cluster similarity, enabling cluster access sequences to be used as the sequential input for WCS / WES task issuance. This allows for control of the scheduling rhythm of bins / pallets in the conveying system and reduction of load jumps between picking locations in different SKU-dense areas.

[0078] Step S7: Iteratively optimize the cluster centers using the particle swarm optimization algorithm. In step S6, the following has been obtained An initial set of cluster centers Then, this embodiment proceeds to step S7. Step S7 encodes the initial set of cluster centers as individual particles in a swarm, and iteratively optimizes the cluster centers using the Particle Swarm Optimization (PSO) algorithm under the cosine distance metric to obtain an optimized set of cluster centers, providing more stable and more cost-effective central locations for subsequent order allocation clusters.

[0079] The purpose of step S7 is to determine the number of clusters in step S5. Without altering the initial clustering center, further optimization is achieved to make the order SKU structure within clusters more consistent and the distances within clusters more compact. This reduces the sensitivity of the clustering results to the initial center and improves the repeatability and stability of the clustering quality. Compared to local updates that rely solely on K-means iterations, PSO explores center location combinations over a larger area through population search. It can achieve rapid convergence with fewer parameters within a wavelet time window and supports outputting the current optimal center upon reaching a preset number of iterations or a time threshold, meeting the engineering requirements of the goods-to-person system for "real-time calculation and interruptible output."

[0080] The particle encoding process is as follows: Let the dimension of the order feature vector be... The number of clusters is In this embodiment, a set of candidate cluster centers is selected. The positions of the particles are concatenated and encoded in a fixed order. Its dimensions are For example, the first center can be... of Each component is placed in The former Wei, the second center To be placed later Dimension, and so on, yields: ; in, A scheme used to characterize a set of cluster centers. Correspondingly, the particle velocity vector. and Having the same dimensions It is used to control the direction and step size of the center position update.

[0081] During the initialization phase, the initial cluster center obtained in step S6 can be used as the position vector of at least one particle. The remaining particles can be generated by applying small random perturbations near the initial center to enhance population diversity.

[0082] The particle swarm velocity and position update process is as follows: Let the particle swarm size be The number of iterations is For the first The particle, in the... The position vector at the next iteration is The velocity vector is Its velocity and position are updated according to the following formula: ; ; In the formula, For the first The particle in the first The velocity vector at the next iteration is calculated using the velocity update formula; The inertial weight is used to adjust the degree to which particles inherit the velocity of the previous iteration, affecting the global search and local convergence characteristics of the algorithm. In engineering, it can be set to around 0.8. No. The particle in the first The velocity vector at each iteration is used to control the update direction and step size of the particle position in the search space; This is an individual learning factor used to adjust the particle's trajectory towards its individual historical best position. The degree of convergence is around 2.0, which can also be adjusted according to the wave duration and convergence requirements; This is a swarm learning factor used to adjust the particle's trajectory toward the swarm's global optimal position. The degree of convergence is around 2.0, which can also be adjusted according to the wave duration and convergence requirements; , is a random number factor, consisting of independent and uniformly distributed random numbers within the interval [0,1], used to introduce random perturbations to enhance search diversity. Corresponding individual items, Corresponding group terms; For the first The particle in the iteration to the _ The individual's historical optimal position vector at this time is the particle position that makes the fitness function optimal during the historical iteration process. For the first The particle in the first The position vector at the next iteration. Used to characterize a set of candidate cluster centers; For the particle swarm optimization in the iteration to the th The global optimal position vector at this time is the position of the particle that makes the fitness function optimal among all the historical positions of all particles in the particle swarm. For the first The particle in the first The position vector at the next iteration is calculated using the position update formula; For particle indexing, , The number of particles; Index for iteration count, , This represents the maximum number of iterations or the number of iterations to terminate.

[0083] To ensure the availability of the center vector, it can be adjusted after the position is updated. Boundary constraints or normalization processes can be applied, such as limiting the range of components or normalizing the magnitude of the center vector, to avoid numerical divergence and ensure the stability of cosine distance calculation.

[0084] The construction process of the fitness function and the intra-cluster discreteness objective function WCSS is as follows: The fitness function is used to evaluate the merits of the cluster center scheme represented by each particle and drives the PSO search direction. In this embodiment, the fitness is constructed using the intra-cluster discreteness objective function, establishing a correspondence between the optimization objective and "mathematical compactness" and "operational cost".

[0085] Let the clustering results include The nth cluster, the th The center vector of each cluster is The set of order vectors within the cluster is Under the cosine distance metric, the intra-cluster scatter is defined as: ; Where the cosine distance is: ; In the formula, Intra-cluster dispersion is the sum of the squared distances from all samples within a cluster to the center of their respective clusters, used to evaluate cluster compactness. The number of clusters represents the number of clusters into which the clustering result is divided. This is a cluster index used to identify the first... There are 5 clusters, with values ​​ranging from 1 to 2. ; Let be any order feature vector within the cluster, representing the set An element whose dimension is same; For order vectors with cluster center vector The distance between them is measured by summing the squares of the distances between orders in the same cluster and the cluster center. Used to evaluate cluster compactness, The smaller the value, the more consistent the SKU structure of orders within the cluster.

[0086] To minimize This is transformed into "maximizing fitness". In this embodiment, the fitness function is defined as: ; In the formula, The fitness function value is used in the particle swarm optimization algorithm to evaluate the quality of a set of cluster centers. It is a positive constant, used to avoid the denominator. A value of zero or too small leads to numerical instability. PSO uses fitness during iteration. Larger is better, updated individual is optimal with global optimal .

[0087] Since the order vector is represented binary by whether the SKU appears, the cosine distance characterizes the differences in the order SKU structure, and the intra-cluster dispersion... A smaller SKU size indicates a higher degree of overlap within the same order cluster. In goods-to-person sorting operations, the more SKUs shared by orders within the same cluster, the more concentrated the types of storage units the picking station needs to access when processing orders in that cluster, and the fewer the number of switches between different storage units. Therefore, minimizing intra-cluster dispersion is crucial. At the operational level, this can be characterized by reducing the number of times picking stations switch between different storage units and the waiting costs, so that the optimization of cluster centers is consistent with the actual operational cost target of the sorting system.

[0088] The PSO iteration may terminate when any of the following conditions are met: 1. Reach the maximum number of iterations 2. The global optimal fitness improvement is less than the threshold; 3. The upper limit of wave computation time is reached. Output the global optimal position upon termination. The corresponding set of cluster centers is input as the optimized set of cluster centers into step S8, which is used to allocate orders to the corresponding clusters according to the principle of minimum cosine distance.

[0089] Step S7 only optimizes the location of cluster centers, without changing the number of clusters. . Number of clusters Determined by the parallel capacity constraint of the order bit, each cluster can be mapped to a batch input for parallel processing of an order bit; the PSO is fixed in this way. Under the premise of improving intra-cluster consistency and load balancing, we can ensure that the clustering results are both "well-clustered" and "can be executed in parallel".

[0090] The optimized cluster centers in step S7 make the cluster structure more stable, providing a reliable foundation for subsequent calculations of cluster center vectors and inter-cluster similarity, and for constructing cluster access sequences. The cluster access sequence, as the sequential input for WCS / WES task distribution, controls the scheduling rhythm of bins / pallets in the conveying system, reducing load jumps between picking locations in areas with high SKU density, thereby achieving system-level continuous operation and throughput improvement.

[0091] Step S8: Orders are clustered and mapped to order positions: After obtaining the optimized cluster center set in step S7, this embodiment proceeds to step S8. Step S8 is used to assign each order within a wave to the corresponding cluster based on the optimized cluster center set, forming order clustering grouping results, and mapping each cluster to order positions for parallel processing according to the order position parallel capacity constraint determined in step S5.

[0092] The purpose of step S8 is to: with a fixed number of clusters Given the constraints on the parallel capacity of order slots, the assignment of "order-cluster" is determined, transforming the cluster center optimization results into executable order groupings. Each cluster is then further mapped to a parallel processing object for an order slot, generating the structured input required for subsequent order slot task queues. This step ensures that the clustering results are not merely algorithm outputs, but rather operational units that directly interface with the order slot resource organization of the goods-to-person sorting system.

[0093] The order allocation rules and implementation process are as follows: Let the optimized set of cluster centers be... For any order vector Calculate the cosine distance between it and each cluster center: ; The order is then assigned to the cluster that minimizes the cosine distance. ; Therefore, we obtain Clusters ,in To be allocated to the The set of order vectors for each cluster center. In engineering implementation, the vectors can be calculated and categorized one by one for each order, or the order-center distance matrix can be calculated in batch processing mode and the allocation can be completed all at once.

[0094] The mapping process from clusters to order positions is as follows: After obtaining the clustering results, based on the parallel capacity constraints and mapping rules for order positions determined in step S5, the clusters are assigned to different order positions for parallel processing. One of the following mapping methods can be used: 1. One-to-one correspondence mapping: when At that time, from Select from each order position One order slot, and Each cluster is mapped one-to-one.

[0095] 2. Constrained Mapping: When there are differences in the processing capacity of order positions, clusters are matched and allocated according to the order position capacity or priority, while maintaining the parallel processing relationship of "different clusters are assigned to different order positions".

[0096] The mapping relationship can be recorded as Yes, among them For the first Cluster identifier, This is the order identifier, used to generate the order task queue.

[0097] Cluster number in this embodiment Instead of selecting based on statistical indicators such as the silhouette coefficient, the selection is determined by the parallel capacity constraint of the order positions, ensuring that each cluster can be mapped to a batch input for parallel processing of order positions. Step S8 is performed on this fixed... Under the premise of completing order allocation and cluster-order bit mapping, the clustering results can be directly used in the execution layer of the goods-to-person system.

[0098] Step S9, calculate the cluster center vector and inter-cluster similarity: After obtaining the cluster division in step S8, this embodiment proceeds to step S9. Step S9 is used to aggregate the order vectors within each cluster to obtain the cluster center vector, and calculate the cosine similarity between any two clusters based on the cluster center vector, providing a computational basis for the subsequent construction of the cluster access sequence.

[0099] The purpose of step S9 is to represent the SKU structure features of each cluster using a computable vector and to measure the proximity between clusters by the similarity between cluster centers. This result is used to generate a cluster access sequence, enabling the task delivery order to be organized based on the continuity of the SKU structure between clusters, rather than solely based on order arrival time or rule thresholds.

[0100] The cluster center vector is calculated as follows: For the There are clusters, and the set of order vectors within each cluster is . The number of orders within the cluster is The cluster center vector is calculated as follows: ; In the formula, Represents a set The number of order vectors in the middle; Let be any order feature vector within the cluster, representing the set An element whose dimension is same; Indicates the first The set of order vectors within each cluster. This results in... The value in each dimension can be understood as the "frequency of occurrence" or "coverage" of the corresponding SKU for each order pair within the cluster, which can be used to describe the SKU structure distribution characteristics of the cluster.

[0101] The method for calculating the cosine similarity between clusters is as follows: For any two clusters and Their cluster center vectors are respectively and Then the cosine similarity between clusters is: ; The greater the similarity, the closer the two clusters are in terms of SKU structure, and they will be preferentially arranged as adjacent clusters for subsequent processing.

[0102] The optimization of intra-cluster dispersion in step S7 makes the intra-cluster SKU structure more concentrated. Step S9 further describes this concentration using cluster center vectors and measures the cost of switching between adjacent clusters using inter-cluster similarity. The greater the inter-cluster similarity, the more continuous the SKU structure of adjacent clusters, and the smaller the change in the type of storage unit that the corresponding picking position needs to switch when processing across clusters.

[0103] Step S10: Construct a cluster access sequence based on inter-cluster similarity and use it as the task distribution order: After obtaining the cluster center vector and the cosine similarity between clusters in step S9, this embodiment proceeds to step S10. Step S10 is used to construct an access sequence covering all clusters, and input this access sequence into WCS / WES as the picking task distribution order of the goods-to-person sorting system.

[0104] The purpose of step S10 is to determine the processing order of clusters based on the "cluster-order position parallel mapping" structure, so that adjacent processing clusters maintain continuity in the SKU structure, thereby controlling the SKU structure jumps and storage unit arrival rhythm fluctuations when picking positions process different clusters. This step elevates the "clustering result" to a "system-level scheduling control strategy," allowing the output to be directly used as the task issuance order of WCS / WES, driving the conveying system or handling equipment to schedule and arrange the arrival order of bins / pallets.

[0105] The greedy iterative construction rule for cluster access sequences is as follows: Let the set of clusters be: ; The corresponding set of cluster center vectors is Initially, any cluster is selected as the first cluster of the sequence. And establish a set of unselected clusters: ; For the Selected clusters In the unselected set Choose the next cluster that has the highest similarity to the cosine: ; And add the cluster to the access sequence, remove it from the unselected set, and let ← Iterate until the unselected set is empty, to obtain the access sequence containing all clusters: ; In the formula, The set of clusters represents the set of all clusters obtained from clustering. For cluster set The various clusters within; This represents the number of clusters. To cluster The set of cluster center vectors with one-to-one correspondence, where Represents clusters The cluster center vector is used to characterize the SKU structure features of the cluster; To access the next cluster in the sequence, as described above Rules are selected from candidate clusters; The operator that takes the value of the argument that maximizes the objective function; Cluster It belongs to the set of unselected clusters and is a candidate cluster that can be selected as the next cluster. This is a candidate cluster index, used to represent a candidate cluster in the set of unselected clusters. ; The set of unselected clusters represents the set of clusters. Exclude clusters that have already been added to the access sequence. The remaining clusters, where \ represents the set difference operation; For the current cluster The cluster center vector; Candidate clusters The cluster center vector; , Representing vectors respectively and The modulus is used to normalize the dot product and form the cosine similarity. To access the cluster index in the sequence.

[0106] In one alternative implementation, to improve sequence quality, the first cluster selection rule can be extended from "any cluster" to "select the cluster with the highest average similarity to other clusters" or "select the cluster with the largest cluster center modulus", without changing the basic idea of ​​greedy construction based on inter-cluster similarity.

[0107] Task distribution order generation and system execution interface: Cluster access sequence The picking task distribution sequence for a goods-to-person sorting system is as follows: The corresponding clusters are extracted sequentially from the clusters in the data. →…→ Combined with the "cluster-order position mapping relationship" from step S8, an order position task queue entry is generated and submitted to WCS / WES. The task queue entry must contain at least: an order position identifier, a cluster identifier, a set of order numbers, and a set of SKU numbers including priority / timestamps. WCS / WES then drives the conveyor system or handling equipment to schedule the storage unit to arrive at the picking position and complete parallel picking according to this order.

[0108] Number of clusters Determined by the parallel capacity constraint of the order position, the cluster access sequence is naturally aligned with the parallel processing structure of the order position, and the access sequence can be directly mapped to the executable task issuance rhythm. K-means++ and PSO improve the stability of clusters and the consistency within clusters, making the cluster center vector representation more reliable. The greedy sequence construction that maximizes the similarity between clusters makes the SKU structure of adjacent clusters more continuous, the change of the storage unit type required for picking positions during task issuance is smoother, and there is less storage unit switching and waiting caused by cross-cluster switching. The cluster access sequence, as the task issuance order of WCS / WES, belongs to the system-level control input and is used to control the scheduling rhythm of bins / pallets in the conveying system, rather than just outputting the sorting result.

[0109] Step S11: Output the clustering results and access sequence, and generate an order position task queue. like Figure 3 As shown, according to the cluster access sequence Traverse each cluster, extract the order set and SKU list within the cluster, generate task queue entries based on the cluster-order position mapping relationship, and add them to the corresponding order position task queue. Finally, output the order position task queue set for system execution.

[0110] After obtaining the order clustering results in step S8 and the cluster access sequence in step S10, this embodiment proceeds to step S11. Step S11 is used to output the order clustering results and the cluster access sequence, and generate an order task queue accordingly, which is then submitted to the warehouse control system WCS / WES of the goods-to-person sorting system for execution.

[0111] The purpose of step S11 is to convert the "clustering grouping results + cluster access sequence" obtained from the algorithm into a task queue data structure that can be directly executed by the system, enabling the goods-to-person sorting system to organize order box processing according to the parallel capability of order positions and control the task issuance rhythm according to the cluster access sequence. This step completes the mapping from "clustering optimization output" to "order position task queue / task issuance instruction", ensuring that the method of the present invention can form an executable operation sequence and parallel processing arrangement in the actual goods-to-person sorting system.

[0112] The input for step S11 includes at least: 1. Order clustering results: The cluster identifier for each order, and the set of order numbers contained in each cluster; 2. Cluster access sequence 3. Cluster-order position mapping relationship: The correspondence between clusters and order positions obtained in step S8 based on the parallel capacity constraint of order positions in step S5.

[0113] To facilitate system execution, this embodiment defines the task queue as a set of queues divided by order position. ,in For the number of order slots; Indicates order position The task queue. Each task queue consists of several task items, and each task item corresponds to a cluster.

[0114] Step S11 generates the order task queue as follows: 1. Determine the order in which tasks are issued: Based on the cluster access sequence obtained in step S10, the clusters in the access sequence are mapped sequentially to the task distribution order of the goods-to-person sorting system, i.e., according to... → →…→ The clusters are processed and distributed in a specific order.

[0115] 2. Generate task queue entries: For each cluster in the access sequence Extract the set of order numbers corresponding to this cluster. and the set of SKU numbers corresponding to the order number And determine the corresponding order position identifier according to the cluster-order position mapping relationship. Construct a task queue entry. and add it to the order space. Task queue .

[0116] 3. Form an order queue set and output it externally: Complete all After traversing each cluster, output the order task queue set. It also outputs the order clustering results and cluster access sequence, which are then processed and distributed by WCS / WES according to the queue.

[0117] Each task queue entry must contain at least the following fields: 1. Order ID (Bin_ID): Indicates which order position will process this entry in parallel; 2. Cluster ID: Used to identify the cluster corresponding to this entry; 3. Order ID Collection Order_ID_List: A list of order IDs contained within this cluster; 4. SKU ID List: The SKU detail set corresponding to the order ID set, which must contain at least the SKU ID; it may also include SKU quantity, picking quantity, storage location / storage unit identifier, etc.

[0118] In an optional implementation, task queue entries may also include fields such as priority, wave number, estimated processing time, and exception handling flag, for WCS / WES to perform more granular scheduling control, but without changing the basic structure of "organizing and distributing by cluster access sequence and executing by order-based parallel queue".

[0119] Interface with WCS / WES for distribution and execution: After receiving the order position task queue set, WCS / WES can execute it in the following way: Parallel execution: Each order position retrieves its own task from its own task queue. The first item in the queue is retrieved and processed in parallel, enabling simultaneous picking of multiple order positions. Sequence constraints: At the task issuance level, WCS / WES controls the pace at which task items enter the execution state according to the global order of the cluster access sequence, or sets an executable flag for items entering the queue to maintain the continuity of the SKU structure between adjacent clusters. Driven scheduling: For task queue items in the execution state, WCS / WES generates scheduling instructions for the storage unit (bin / pallet) to arrive at the picking position based on the SKU list, and completes the delivery from the picking position to the corresponding order bin.

[0120] In one alternative implementation, if the system adopts a "first issue queues, then activate in sequence" mechanism, all task queue entries can be written into the queue, and the corresponding entries can be activated sequentially according to the cluster access sequence. This approach is beneficial for pre-fetching data and pre-scheduling on the system side, but still maintains the constraint of the access sequence on the execution order.

[0121] The number of clusters in this method Instead of being determined by statistical indicators such as the silhouette coefficient, the constraint is derived from the parallel capacity of the order positions, ensuring that each cluster structurally corresponds to a batch input for parallel processing of order positions. The order position task queue generated in step S11 establishes a "cluster-order position" mapping and parallel queue structure based on this constraint, and is an executable scheduling input rather than just an algorithm output.

[0122] Since the order SKU features are high-dimensional sparse binary vectors, K-means++ is used to select the initial centers of the dispersed SKU space under the cosine distance metric, and PSO is used to quickly iterate and optimize the centers within the wavelet secondary time window and support outputting the optimal solution midway. The above design enables the optimization of the SKU space under a fixed time window. The clusters obtained under constraints are more stable, and the order task queue fluctuates less between waves, making it easier for the system to continue execution.

[0123] Minimizing intra-cluster dispersion corresponds to a higher degree of overlap in the intra-cluster order SKU set, a more concentrated SKU list in the task queue entry, a more concentrated type of storage unit that WCS / WES needs to schedule when executing the entry, and a lower frequency of storage unit switching and waiting.

[0124] The cluster access sequence generated in step S10 is solidified as the task issuance sequence input in step S11. It is used to control the scheduling rhythm and arrival order of bins / pallets in the conveying system, reduce the picking position load fluctuation caused by SKU structure jumps between adjacent task clusters, and make "sorting" an executable control quantity of WCS / WES.

[0125] In summary, this order clustering and picking order optimization method for goods-to-person sorting systems uses feature vectorization to represent wave order data and calculates the similarity between orders. Under the constraint of parallel order positions, it determines the number of clusters, adopts an improved cluster center initialization and swarm intelligence optimization strategy to obtain stable and high-quality order clustering results, and builds cluster center representation and inter-cluster similarity measurement on this basis. It further generates a picking execution order that reduces the switching cost between clusters, thereby achieving the goals of improving intra-cluster order similarity, reducing picking position switching frequency, balancing order position workload, and improving the overall picking efficiency and throughput of the system.

[0126] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for order clustering and picking order optimization in a goods-to-person sorting system, characterized in that: Includes the following steps: S1. Obtain order detail data within a preset time period from the order management system according to the wave. The order detail data must include at least the order number and SKU number, and form a set of pending orders for that wave. S2. Count all SKUs appearing in the set of orders to be processed, construct the complete SKU set, and define a fixed-dimensional feature space based on the complete SKU set; S3. Map each order in the set of orders to be processed to a binary order feature vector in the feature space, where the values ​​of each dimension of the order feature vector are used to indicate whether the corresponding SKU is included in the order; S4. Calculate the cosine similarity between any two orders based on the order feature vectors, and convert the cosine similarity into cosine distance to obtain the order distance metric. S5. Determine the number of clusters under the parallel capacity constraint of order positions. The parallel capacity constraint of the order position must at least satisfy The number of orders should not exceed the number of order slots, and after clustering is completed, each cluster will be assigned to different order slots for parallel processing. S6. Selecting parameters using the K-means++ initialization strategy under the cosine distance metric. There are several initial cluster centers, where: one order is randomly selected as the first initial cluster center; for the remaining orders, a selection probability distribution is constructed based on the minimum cosine distance between the order and the selected cluster center, and subsequent initial cluster centers are iteratively selected until a suitable cluster center is obtained. One initial cluster center; S7, will An initial set of cluster centers is encoded as individual particles in a swarm. The particle swarm optimization algorithm is used to iteratively update the particle positions to optimize the cluster centers. The fitness of the particle swarm optimization algorithm is given by the intra-cluster discreteness objective function. The intra-cluster discreteness objective function includes the aggregation calculation of the cosine distance between the order feature vectors within each cluster and the corresponding cluster centers, resulting in the optimized set of cluster centers. S8. Based on the optimized cluster center set, each order is assigned to the corresponding cluster according to the principle of minimizing the distance to the cosine, thus obtaining the order clustering grouping result, and the clusters are mapped to the order positions according to the allocation rules in step S5. S9. Aggregate the order feature vectors within each cluster to obtain the cluster center vector, and calculate the cosine similarity between any two clusters based on the cluster center vector; S10. Construct a cluster access sequence based on the cosine similarity between clusters. The construction method is as follows: determine the initial clusters; iteratively select the next cluster with the largest cosine similarity to the current cluster from the clusters that have not been added to the access sequence, until all clusters are added to the access sequence; use the cluster access sequence as the order of picking tasks issued by the goods-to-person sorting system. S11. Output the order clustering results and cluster access sequence, and generate the order task queue for delivery to the sorting system based on this.

2. The order clustering and picking order optimization method for a goods-to-person sorting system according to claim 1, characterized in that: In steps S2 and S3, the complete SKU set and order feature vector are constructed as follows: All SKUs appearing within a wave are grouped into a set. ; ； For any order Construct binary order feature vectors ; ； Among them, when the order Include hour ,otherwise ; In the formula, The SKU set represents the complete set of all SKUs that have appeared in the same batch of pending orders. SKU complete set The first in SKUs, of which For dimensional indexing, ; SKU complete set The number of elements, i.e. the number of SKU types appearing in this wave; "Order" refers to any order in the set of orders to be processed. For orders The order feature vector has a dimension of ; For orders The expanded form of the eigenvectors, which includes One component; For orders The eigenvectors in the th Components on the dimension.

3. The order clustering and picking order optimization method for a goods-to-person sorting system according to claim 1, characterized in that: In step S4, the cosine similarity and cosine distance between orders are calculated as follows: Any two orders and The cosine similarity is: ； The cosine distance is: ； In the formula, For the first One order; For the first One order; and For order indexing, ; For orders The corresponding order feature vector has a dimension equal to the size of the entire SKU set. ; For orders The corresponding order feature vector also has the same dimension. ; For orders With orders The cosine similarity is used to characterize the directional similarity between two orders in the SKU feature space; a larger value indicates a higher degree of similarity. ; For vectors The modulus is used to normalize the dot product; For vectors The modulus length; For orders With orders The cosine distance, obtained from the cosine similarity transform, is used to satisfy the distance metric that the smaller the distance, the higher the similarity. .

4. The order clustering and picking order optimization method for a goods-to-person sorting system according to claim 1, characterized in that: Number of clusters in step S5 Under the constraint of parallel capacity for order slots, it is determined as follows: ； In the formula, This indicates the number of orders pending processing in this wave; This indicates the number of order positions that can work in parallel in a goods-to-person sorting system; This indicates the maximum number of orders that can be processed simultaneously in a single parallel processing batch; This indicates the rounding up operation.

5. The order clustering and picking order optimization method for a goods-to-person sorting system according to claim 1, characterized in that: In step S6, the K-means++ initialization strategy selects the initial cluster centers as follows: Randomly select one order vector from the order set as the first cluster center; for any order vector that is not selected as a center... Define it to the set of selected centers. The minimum cosine distance is: ； The next cluster center is selected according to the following probability distribution: ； In the formula, Indicates the first The order feature vectors corresponding to each order are used as candidate vectors for initializing centers in K-means++. For order indexing, , This represents the number of orders in this wave. This represents a cluster center vector within the selected set of cluster centers; Indicates the first The set of selected cluster centers at the next selection is the set of vectors of the currently selected cluster centers. , The index is the count of the number of selected centers / iteration rounds; Represents the candidate order vector Gather at the selected center The minimum cosine distance, i.e. With sets The minimum cosine distance between all center vectors in the vector; Indicates the set of the selected center. All center vectors Take the result that minimizes the expression within the parentheses; Represents the candidate order vector The probability of being selected as the next cluster center; This indicates that the calculation is performed on all candidate order vectors. The summation result is used as a normalization factor to make the sum of the probabilities of all candidate points equal to 1. The index of the candidate order vector.

6. The order clustering and picking order optimization method for a goods-to-person sorting system according to claim 1, characterized in that: In step S7, the particle swarm optimization algorithm uses the following velocity and position update formulas to iteratively optimize the cluster centers: ； ； In the formula, For the first The particle in the first The velocity vector at the next iteration is calculated using the velocity update formula; The inertial weight is used to adjust the degree to which particles inherit the velocity of the previous iteration, affecting the global search and local convergence characteristics of the algorithm. No. The particle in the first The velocity vector at each iteration is used to control the update direction and step size of the particle position in the search space; This is an individual learning factor used to adjust the particle's trajectory towards its individual historical best position. The degree of closeness; This is a swarm learning factor used to adjust the particle's trajectory toward the swarm's global optimal position. The degree of closeness; , is a random number factor, consisting of independent and uniformly distributed random numbers within the interval [0,1], used to introduce random perturbations to enhance search diversity. Corresponding individual items, Corresponding group terms; For the first The particle in the iteration to the _ The individual's historical optimal position vector at this time is the particle position that makes the fitness function optimal during the historical iteration process. For the first The particle in the first The position vector at the next iteration. Used to characterize a set of candidate cluster centers; For the particle swarm to iterate to the th The global optimal position vector at this time is the position of the particle that makes the fitness function optimal among all the historical positions of all particles in the particle swarm. For the first The particle in the first The position vector at the next iteration is calculated using the position update formula; For particle indexing, , The number of particles; Index for iteration count, , This represents the maximum number of iterations or the number of iterations to terminate.

7. The order clustering and picking order optimization method for a goods-to-person sorting system according to claim 1, characterized in that: In step S7, the intra-cluster discreteness objective function is constructed as follows: Let the clustering results include The nth cluster, the th The center vector of each cluster is The set of order vectors within the cluster is The intra-cluster dispersion is: ； The fitness function is defined as follows: ； In the formula, Intra-cluster dispersion is the sum of the squared distances from all samples within a cluster to the center of their respective clusters, used to evaluate cluster compactness. The number of clusters represents the number of clusters into which the clustering result is divided. This is a cluster index used to identify the first... There are 5 clusters, with values ​​ranging from 1 to 2. ; Let be any order feature vector within the cluster, representing the set An element whose dimension is same; For order vectors with cluster center vector Distance metric between; The fitness function value is used in the particle swarm optimization algorithm to evaluate the quality of a set of cluster centers. It is a positive constant, used to avoid the denominator. A value of zero or too small will cause the value to be unstable.

8. The order clustering and picking order optimization method for a goods-to-person sorting system according to claim 1, characterized in that: In step S9, the cluster center vector is calculated as follows: For the first There are n clusters, and the cluster center vector is: ； In the formula, Represents a set The number of order vectors in the middle; Let be any order feature vector within the cluster, representing the set An element whose dimension is same; Indicates the first A set of order vectors within a cluster.

9. The order clustering and picking order optimization method for a goods-to-person sorting system according to claim 1, characterized in that: In step S10, the cluster access sequence is constructed according to a greedy iteration rule: Let the set of clusters be The corresponding cluster center vector is Initialize by selecting any cluster as the first cluster of the sequence. ; for the first Selected clusters Select the next cluster from the unselected cluster set: ； Until an access sequence containing all clusters is obtained. ; In the formula, The set of clusters represents the set of all clusters obtained from clustering. For cluster set The various clusters within; This represents the number of clusters. To cluster The set of cluster center vectors with one-to-one correspondence, where Represents clusters The cluster center vector is used to characterize the SKU structure features of the cluster; To access the next cluster in the sequence, as described above Rules are selected from candidate clusters; The operator that takes the value of the argument that maximizes the objective function; Cluster It belongs to the set of unselected clusters and is a candidate cluster that can be selected as the next cluster. This is a candidate cluster index, used to represent a candidate cluster in the set of unselected clusters. ; The set of unselected clusters represents the set of clusters. Exclude clusters that have already been added to the access sequence. The remaining clusters, where \ represents the set difference operation; For the current cluster The cluster center vector; Candidate clusters The cluster center vector; , Representing vectors respectively and The modulus is used to normalize the dot product and form the cosine similarity.

10. The order clustering and picking order optimization method for a goods-to-person sorting system according to claim 1, characterized in that: In step S11, the order task queue is generated as follows: According to the cluster access sequence obtained in step S10, the clusters in the access sequence are mapped to the task distribution order of the goods-to-person sorting system in turn, and the order set in each cluster is converted into the task queue entry of the corresponding order position; wherein, the task queue entry contains at least the order position identifier, the cluster identifier, the order number set, and the SKU number set corresponding to the order number.

Images