RFID warehousing logistics management method and system

Abstract

The present invention discloses an RFID warehousing and logistics management method and system, which relates to the technical field of warehousing and logistics management. The present invention constructs a historical commodity information database, uses GAT to extract features from commodity data, and uses HDBSCAN to cluster commodities. The commodities in each cluster are trained using a kernel ridge regression model, and the parameters are optimized using a particle swarm to obtain an optimal storage location model. When a new commodity enters the warehouse, information is obtained through the RFID tag, and its features are extracted using the trained GAT model. The cluster Cmin corresponding to the minimum distance between the cluster feature center and the cluster feature center is found and substituted into Qmin to obtain the optimal storage location for the new commodity. A backtracking algorithm is used to generate all alternative paths from the storage location to the warehouse door, and the Pareto algorithm is used to optimize the alternative paths. The present invention can quickly locate the storage location of new commodities and pre-plan the optimal path to the warehouse door when the commodity is shipped out.

Description

Technical Field

[0001] The present invention relates to the technical field of warehousing and logistics management, and in particular to an RFID warehousing and logistics management method and system. Background Art

[0002] In today's complex and ever-changing warehousing and logistics environment, traditional warehouse management methods face numerous challenges. With the rapid increase in the variety and quantity of goods, manual management struggles to accurately determine optimal storage locations and efficient retrieval routes for goods, leading to problems such as low warehouse space utilization, low logistics efficiency, and rising costs. The emergence of RFID technology has brought new opportunities to warehouse management. While RFID can quickly obtain product identification information, utilizing this information to make more efficient warehousing decisions, such as optimizing storage locations and planning retrieval routes, is difficult to achieve solely with RFID technology. Effective data analysis methods are needed to further unlock its value. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention provides an RFID warehousing logistics management method and system, which solves the problem that when new goods enter the warehouse, the storage location cannot be found quickly and accurately, and the optimal path to the warehouse door cannot be found based on the current location of the goods.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions: an RFID warehouse logistics management method, comprising:

[0005] S1. Construct a product feature database and an optimal storage location database based on warehouse historical information;

[0006] S2. Use the Graph Attention Network (GAT) to extract features from the data in the product feature database;

[0007] S3. Based on the features extracted by GAT, the HDBSCAN density clustering algorithm is used to cluster the products and obtain r clusters;

[0008] S4. Use the kernel ridge regression model to train the products in each cluster Ct(t∈[1,r]), and optimize its regularization parameters and kernel parameters through the particle swarm optimization algorithm to obtain the optimal storage location model QTt of Ct;

[0009] S5. When a new product enters the warehouse, the RFID tag of the new product is scanned by an RFID reader to obtain all its information. The trained GAT model is then used to extract its features and calculate the distance between it and the feature center of each cluster. The cluster with the smallest distance Cmin is found and the features are substituted into the optimal storage location model Qmin corresponding to Cmin to obtain the optimal storage location for the new product.

[0010] S6. Based on the optimal storage location of the new product, a backtracking algorithm is used to generate all alternative paths from the warehouse entrance to the storage location, and the Pareto algorithm is used to perform multi-objective optimization on the alternative paths to obtain the optimal alternative path.

[0011] As a further solution of the present invention, the specific steps of using GAT to extract features from commodity data are as follows:

[0012] Standardize product data;

[0013] Construct a correlation graph G = (V, E) between each product;

[0014] Determine the hyperparameters of the GAT layer, including the number of attention heads K and the output feature dimension dout;

[0015] Initialize the weight matrix and attention coefficient vector of each attention head;

[0016] The Latin Hypercube (LHS) sampling method is used to sample the neighboring nodes of each node;

[0017] Calculate each node vi and its neighbor node vj∈Ni after LHS sampling sampled (Ni sampled is the set of sampled neighbor nodes) the attention coefficient eij under each attention head k k :

[0018] Among them, LeakyReLU is the activation function, and its formula is LeakyReLU(a)=max(0.01a,a);

[0019] Get the aggregated features of node vi under each attention head k in, is the normalized attention weight;

[0020] The feature zi under each attention head k k Perform concatenation operation to obtain the final output feature of node vi as zi=concat(zi 1 ,zi 2 ,...,zi K ), where concat() is the concatenation function.

[0021] As a further solution of the present invention, the specific process of LHS sampling includes:

[0022] Number the nodes in the neighbor node set Ni in random order;

[0023] Divide each dimension (which can be understood as a feature dimension of each neighbor node) into m equal probability intervals;

[0024] For each divided interval, a random number generator is used to generate a random number within the interval;

[0025] Match the randomly selected points in each interval with the values ​​of the corresponding dimensions in the neighbor node set;

[0026] In multiple dimensions, select nodes that match the interval of the randomly selected point multiple times.

[0027] As a further solution of the present invention, the HDBSCAN density clustering algorithm is used to cluster the products:

[0028] After the n product data are processed by GAT, the feature vector of each product is zi, with a dimension of dout. They are assembled into an n×dout matrix Y, where Ypq represents the value of the p-th product in the q-th feature dimension;

[0029] For any two products i and j, the similarity is calculated according to the formula d(Yi,Yj) = ||Yi-Yj||;

[0030] Determine a neighborhood radius ε and the minimum number of points MinPts contained in the neighborhood;

[0031] Find all points within the neighborhood radius ε, that is, point j that satisfies d(Yi,Yj)<=ε, and record it as set N ε (i);

[0032] Calculate the set N ε The distances from the point i to the point i in (i) are MinPts. After sorting these distances, the distance with the largest MinPts value is taken as the core distance coredist(i). If the number of points in the set is less than MinPts, the core distance is defined as infinite.

[0033] Based on the core distance, the direct density accessibility, density accessibility and density connection between commodities are judged;

[0034] For any core point i, all points densely connected to it form a cluster, and all core points are traversed until all points are divided into corresponding clusters or marked as noise points (those that neither belong to any cluster core point nor are densely connected to any cluster point).

[0035] As a further embodiment of the present invention, the steps for confirming the optimal storage location model QTt for each cluster Ci are as follows:

[0036] Construct the dataset E1={(Y1,U1),(Y2,U2),...,(Yn,Un)}, where Yi∈R dout is the input feature vector extracted by GAT, and Ui is the data in the optimal storage location database;

[0037] The dataset is divided into 70% training set Etrain and 30% test set Etest;

[0038] Assuming the particle swarm size is N, the inertia weight is w, and the learning factors are c1 and c2, the position of particle i can be expressed as Q i =(λ i ,σ1 i ), speed is expressed as V i =(v λ,i ,v σ1,i ), initialize the particle's velocity Vi to 0 vector, Q i Randomly generated within the set parameter range, the initial individual optimal position pbesti=Q i , the global optimal position is gbest;

[0039] For the [0.7*n] ([] is the rounding symbol) samples in the training set Etrain, according to the selected Gaussian kernel function, Get the kernel matrix K∈R [0.7*n]×[0.7*n] ;

[0040] For each particle, its position parameter Q i Substituted into the kernel ridge regression model, the objective function of the kernel ridge regression model is By (K T K+λ i I) β=K T U is calculated to obtain the coefficient vector β;

[0041] Using k-fold cross validation, the obtained kernel ridge regression model is trained and validated on the training set, and the mean square error (MSE) of the model is calculated as the fitness value Fi, that is, Among them, MESj is the mean square error of the j-fold validation, nj is the number of samples in the j-fold validation set, Us true is the true value, Us ypred is the predicted value;

[0042] Compare the current fitness value of each particle and update pbesti and gbest;

[0043] Update the particle's velocity and position:

[0044]

[0045]

[0046] The obtained global optimal position gbest is used as the optimal regularization parameter and kernel parameter of the kernel ridge regression model, and the training set Etrain is substituted into the retrained model to obtain QTt;

[0047] Use the test set Etest to evaluate QTt. If the effect is not good, continue to iterate until the optimal QTt is found.

[0048] As a further solution of the present invention, the specific steps of using the Pareto algorithm to select the optimal path are as follows:

[0049] The values ​​of factors such as the path length, number of turns, labor cost, equipment loss cost, and waiting time of each alternative path are organized into a matrix form. If there are m1 alternative paths in total, the matrix A1 is constructed as follows: Where li represents the path length of the i-th path, ci represents the number of turns, hi represents the labor cost, ei represents the equipment loss cost, fi represents the waiting time, and i∈[1,m];

[0050] Compare the paths. If path i is not worse than path j in all factors and has at least one factor better than path j, then path i dominates path j. Find all paths that are not dominated by other paths to form the first-level non-dominated set F1. Repeat the process for the remaining paths until all paths are divided into the corresponding non-dominated layer.

[0051] For each path in the non-dominated layer (Fs, s = 1, 2, ...), calculate its congestion distance;

[0052] Prioritize the paths at the front of the non-dominated layer. If there are many selectable paths in the same non-dominated layer, choose the path with the larger congestion distance.

[0053] An RFID warehousing and logistics management system, comprising: a historical information database construction module, a feature extraction module, a commodity clustering module, an optimal storage location construction module, a new commodity storage location determination module, and an alternative path selection module;

[0054] A historical information database construction module, which constructs a commodity feature database and an optimal storage location database based on the warehouse historical information and transmits them to the feature extraction module;

[0055] Feature extraction module, which uses the graph attention network (GAT) to extract features from the data in the product feature database and transmits them to the product clustering module;

[0056] Commodity clustering module: This module clusters commodities based on the features extracted by GAT using the HDBSCAN density clustering algorithm to obtain r clusters, and then transfers the clusters to the optimal storage location construction module;

[0057] The optimal storage location construction module uses the kernel ridge regression model to train the products in each cluster Ct (t∈[1,r]) and optimizes its regularization parameters and kernel parameters through the particle swarm optimization algorithm to obtain the optimal storage location model QTt corresponding to the products in Ct, and transmits QTt to the new product storage location confirmation module;

[0058] The new product storage location confirmation module is based on the fact that when a new product enters the warehouse, this module scans the RFID tag of the new product through an RFID reader to obtain its full information. It then uses the trained GAT model to extract its features and calculates the distance between it and the feature center of each cluster. It finds the cluster with the smallest distance Cmin and substitutes the features into the optimal storage location model Qmin corresponding to Cmin to obtain the optimal storage location for the new product. This is then transmitted to the alternative path selection module.

[0059] The alternative path selection module uses a backtracking algorithm to generate all alternative paths from the warehouse entrance to the storage location based on the optimal storage location of the new product, and uses the Pareto algorithm to perform multi-objective optimization on the alternative paths to obtain the optimal alternative path.

[0060] The present invention provides an RFID warehouse logistics management method and system, which has the following beneficial effects compared with the existing technology:

[0061] (1) When using GAT to extract key features of commodity data, the present invention innovatively introduces LHS to sample the neighboring nodes of the current node to avoid redundant calculations caused by high overlap of neighboring nodes. This can better ensure that the selected neighboring nodes are random and representative in the feature space.

[0062] (2) The present invention uses HDBSCAN to cluster historical commodities, and constructs different optimal storage location models for different clusters, and optimizes them using the particle swarm optimization algorithm to obtain a relatively accurate optimal storage location model.

[0063] (3) Based on the optimal storage location of new products, the present invention finds several alternative paths to the warehouse door through a backtracking algorithm, and optimizes these alternative paths through the Pareto algorithm to find the best alternative path, which effectively saves transportation time and improves the efficiency of product delivery. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 Flow chart of the steps of the present invention;

[0065] Figure 2 This is a block diagram of the system principle of the present invention. DETAILED DESCRIPTION

[0066] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0067] like Figure 1 The present invention provides an RFID warehousing logistics management method, comprising:

[0068] S1. Construct a comprehensive product feature database and an optimal storage location database. The product feature database covers multiple dimensions, including size, weight, sales volume, inventory turnover, gross profit margin, and positive review rate. These features reflect the product's attributes and market performance from different perspectives. For example, a smartphone might have dimensions of 160mm long, 75mm wide, and 8mm high, weigh 200g, and cost 3,000 yuan. In the past quarter, it sold 5,000 units, had an inventory turnover rate of 0.8, a gross profit margin of 40%, and a positive review rate of 92%.

[0069] The optimal storage location database is the location coordinates of the goods on the shelf. A specific corner of the warehouse can be selected as the origin. For example, the lower left corner of the warehouse can be selected. The horizontal right direction is determined as the positive direction of the X-axis to represent the lateral position of the goods in the warehouse plane. The direction perpendicular to the X-axis and in the same plane (usually horizontally forward) is set as the positive direction of the Y-axis to represent the longitudinal position of the goods in the warehouse plane. The direction vertically upward away from the ground is defined as the positive direction of the Z-axis to reflect the shelf layer or height position of the goods. When measuring, the X-axis and Y-axis are divided into scales with appropriate length units (such as meters, decimeters, etc.) to accurately determine the position of the goods in the plane. For the Z-axis, if the height of each shelf layer is fixed, it can be marked according to the number of layers, for example, the bottom layer is layer 0, and the upper layers can be layer 1, layer 2, etc., so as to clearly represent the vertical position information of the goods.

[0070] S2. Use the Graph Attention Network (GAT) to deeply extract data from the product feature database. The GAT model can automatically capture the complex relationships between products by learning the attention weights between nodes (products), thereby identifying key features. The specific steps are as follows:

[0071] Standardize the data in the product feature database. The standardization formula is: Among them, x is the original feature, x' is the standardized feature, μ is the mean, σ

[0072] This step helps improve the training effect and stability of the GAT model, makes different features comparable, and avoids adverse effects on the model due to differences in data dimensions.

[0073] When constructing the product graph G = (V, E), in addition to building edge connections based on certain obvious similarities between products (such as products in the same category), factors such as sales associations can also be considered. For example, if two products are often purchased together, an edge can be established between them. Let the node set H = {v1, v2, ..., vn}, the edge set E = {(vi, vj) | i ≠ j}, and assign an initial weight to each edge (if necessary). The weight can be set based on a quantitative indicator of the degree of association;

[0074] Determine the hyperparameters of the GAT layer, including the number of attention heads K and the output feature dimension dout. By setting different numbers of attention heads, you can focus on the node's neighbor information from different perspectives. Each attention head will generate an independent feature representation, which will then be fused.

[0075] Initialize the weight matrix of each attention head and the attention coefficient vector α k ∈R 2dout / K ,These parameters are usually initialized by random initialization, but it is necessary to ensure that the parameter values ​​are within a suitable range to avoid gradient explosion or vanishing problems. Therefore, the Xavier or He initialization method can be used to initialize the weight matrix;

[0076] In the GAT calculation process, the node's neighbor node information is crucial, but the high overlap of neighbor nodes will cause redundant calculations and reduce computational efficiency. To solve this problem, the Latin Hypercube (LHS) sampling method is introduced to sample the neighbor nodes of each node.

[0077] For each node vi, before calculating the attention coefficient, use LHS to select some nodes from its neighbor node set Ni, and set the number of neighbor nodes to be sampled to m (m<|Ni|);

[0078] The specific sampling process of LHS is as follows: the nodes in the neighbor node set Ni are numbered in a certain order (random order or ascending order of similarity between nodes); according to the Latin hypercube design principle, each dimension (here it can be simply understood as a feature dimension of each neighbor node or a quantitative indicator of the comprehensive feature) is divided into m equal probability intervals; a point is randomly selected in each interval, and the neighbor node corresponding to the interval where the selected point is located is the sampled node. LHS sampling can ensure a certain degree of randomness while making the sampled neighbor nodes more representative in the feature space, reducing redundant calculations that may be caused by the high overlap of neighbor nodes;

[0079] Calculate each node vi and its neighbor node vj∈Ni after LHS sampling sampled (Ni sampled is the set of sampled neighbor nodes) the attention coefficient eij under each attention head k k :

[0080] Among them, LeakyReLU is the activation function, and its formula is LeakyReLU(a)=max(0.01a,a);

[0081] Calculate the aggregate features of node vi under each attention head k in, is the normalized attention weight, and each node vi gets a new feature representation zi under each attention head k k , which integrates the information of neighbor nodes and performs weighted summation based on the attention weight;

[0082] The feature under each attention head k is represented as zi k Perform concatenation operation to obtain the final output feature of node vi as zi=concat(zi 1 ,zi 2 ,...,zi K ), the final output feature dimension is dout, where concat() is a concatenation function, which connects multiple vectors together in sequence to form a new vector with higher dimension.

[0083] S3. Based on the key features extracted by GAT, the HDBSCAN density clustering algorithm is used to cluster the products and obtain r clusters:

[0084] After GAT processing, there are n products, each with a feature vector zi and dimension dout, which can be assembled into an n×dout matrix Y, where Ypq represents the value of the p-th product in the q-th feature dimension;

[0085] The similarity between two products is measured by calculating the Euclidean distance between their feature vectors. That is, for any two products i and j, the distance is calculated according to the formula d(Yi,Yj) = ||Yi-Yj||;

[0086] The steps of determining a neighborhood radius ε and the minimum number of points MinPts contained in the neighborhood and calculating the core distance coredist(i) of product i include:

[0087] Find all points within the neighborhood radius ε, that is, point j that satisfies d(Yi,Yj)<=ε, and record it as set N ε (i);

[0088] Calculate the set N ε The distances from the point i to the point i in (i) are MinPts. After sorting these distances, the distance with the largest MinPts value is taken as the core distance coredist(i). If the number of points in the set is less than MinPts, the core distance is defined as infinite.

[0089] Based on the core distance, we determine the direct density reachability between products. For two products i and j, if the following two conditions are met, then j is said to be directly density reachable from i:

[0090] j∈N ε (i), that is, j is within the neighborhood radius ε of i;

[0091] d(Yi,Yj)<=coredist(i), that is, the distance from j to i does not exceed the core distance of i;

[0092] For two commodities u and v, if there exists a commodity g such that both u and v are density-reachable from g, then u and v are said to be density-connected;

[0093] For any core point i, all points densely connected to it form a cluster, and all core points are traversed until all points are divided into corresponding clusters or marked as noise points (those that neither belong to any cluster core point nor are densely connected to any cluster point).

[0094] S4. For the r clusters that have been clustered, the kernel ridge regression model is used to train the products in each cluster Ct(t∈[1,r]) one by one, and the particle swarm algorithm is used to optimize its regularization parameters and kernel parameters to obtain the corresponding product optimal storage location model QTt:

[0095] Construct the dataset E1={(Y1,U1),(Y2,U2),...,(Yn,Un)}, where Yi∈R doutis the input feature vector extracted by GAT, and Ui is the data in the optimal storage location database;

[0096] The dataset is divided into 70% training set Etrain and 30% test set Etest;

[0097] Assuming the particle swarm size is N, the inertia weight is w, and the learning factors are c1 and c2, the position of particle i can be expressed as Q i =(λ i ,σ1 i ), speed is expressed as V i =(v λ,i ,v σ1,i ), initialize the particle's velocity Vi to 0 vector, Q i Randomly generated within the set parameter range, the initial individual optimal position pbesti=Q i , the global optimal position is gbest;

[0098] For the [0.7*n] ([] is the rounding symbol) samples in the training set Etrain, according to the selected Gaussian kernel function, Get the kernel matrix K∈R [0.7*n]×[0.7*n] ;

[0099] For each particle, its position parameter Q i Substituted into the kernel ridge regression model, the objective function of the kernel ridge regression model is By (K T K+λ i I) β=K T U is calculated to obtain the coefficient vector β;

[0100] Using k-fold cross validation, the obtained kernel ridge regression model is trained and validated on the training set, and the mean square error (MSE) of the model is calculated as the fitness value Fi, that is, Among them, MESj is the mean square error of the j-fold validation, nj is the number of samples in the j-fold validation set, Us true is the true value, Us ypred is the predicted value;

[0101] For each particle i, compare its current fitness value Fi with the fitness value corresponding to the individual optimal position pbesti. If Fi is smaller, update pbesti = Q i , compare the current fitness values ​​of all particles, find the particle position corresponding to the minimum fitness value, and update the global optimal position gbest;

[0102] Update the particle's velocity and position:

[0103]

[0104] Check whether the maximum number of iterations has been reached, or whether the change in fitness value in several consecutive iterations is less than a certain threshold. If the termination condition is met, stop the optimization; otherwise, continue the iteration.

[0105] The obtained global optimal position gbest is used as the optimal regularization parameter and kernel parameter of the kernel ridge regression model, and the training set Etrain is substituted into the retrained model to obtain QTt;

[0106] The test set Etest is used to evaluate QTt, and the evaluation indicators are mean square error (MSE) and determination coefficient (R 2 ).

[0107] S5. When a new product enters the warehouse, use the trained GAT model to extract its features and obtain the feature vector xnew. Calculate the distance between it and the feature center of each cluster. For cluster Ct, its feature center vector is hi and the distance is

[0108] Find the cluster Cmin with the smallest distance, which indicates that the product has the highest similarity with the products in this cluster. Use the product optimal storage location model Qmin corresponding to Cmin to calculate the optimal storage location of the new product.

[0109] S6. Based on the optimal storage location for the new product, a backtracking algorithm is used to generate all possible paths from the warehouse entrance to that storage location. In the warehouse layout diagram, shelves and aisles are abstracted as nodes and edges of the graph. All possible paths are traversed using a backtracking search algorithm. For each alternative path, its path length, number of turns, labor cost, equipment loss cost, and waiting time are recorded. The Pareto algorithm is then used to perform multi-objective optimization on these alternative paths:

[0110] The values ​​of factors such as the path length, number of turns, labor cost, equipment loss cost, and waiting time of each alternative path are organized into a matrix form. If there are m1 alternative paths in total, the matrix A1 is constructed as follows: Where li represents the path length of the i-th path, ci represents the number of turns, hi represents the labor cost, ei represents the equipment loss cost, fi represents the waiting time, and i∈[1,m];

[0111] For any two paths i and j (i≠j), if for all targets, there is A1 k,i <=A1 k,j (k∈[1,5]), and there exists at least one target l such that A l,i l,j ​, then path i is said to dominate path j. Find all paths that are not dominated by other paths. These paths constitute the first-level non-dominated set, denoted as F1. From the remaining paths (excluding the paths in F1), continue to find non-dominated paths according to the above rules to form the second-level non-dominated set F2. And so on, until all paths are divided into the corresponding non-dominated layers;

[0112] For each path in the non-dominated layer (Fs, s = 1, 2, ...), calculate its congestion distance to measure the distribution of solutions around the path, so as to further screen the paths in the same non-dominated layer:

[0113] For each target column (5 columns in total, corresponding to 5 factors), sort the paths in Fs from small to large according to the value corresponding to the column;

[0114] For the first and last paths after sorting, their congestion distance is assigned to infinity;

[0115] For the middle path i, calculate its congestion distance Among them, Ak max and Ak min are the maximum and minimum values ​​of the target value in the kth column in the non-dominated layer Fs, respectively;

[0116] Prioritize paths at the front of the non-dominated layer. If there are many paths available in the same non-dominated layer, further screening can be done based on the congestion distance, giving priority to paths with larger congestion distances.

[0117] If only one path is obtained after screening, then this path is the best path after comprehensive consideration of various factors and can be determined as the final transportation path for the goods;

[0118] If multiple paths remain after filtering by congestion distance, the final path is determined based on the actual warehouse operation focus. For example, if cost reduction is currently the priority, the labor costs and equipment loss costs of these paths can be compared and the path with the lowest cost can be selected. If transportation efficiency is more important, the path length and waiting time can be focused on and the path that shortens the overall transportation time can be selected.

[0119] like Figure 2 , an RFID warehousing logistics management system, comprising: a historical information database construction module, a feature extraction module, a commodity clustering module, an optimal storage location construction module, a new commodity storage location determination module, and an alternative path selection module;

[0120] A historical information database construction module, which constructs a commodity feature database and an optimal storage location database based on the warehouse historical information and transmits them to the feature extraction module;

[0121] Feature extraction module, which uses the graph attention network (GAT) to extract features from the data in the product feature database and transmits them to the product clustering module;

[0122] Commodity clustering module: This module clusters commodities based on the features extracted by GAT using the HDBSCAN density clustering algorithm to obtain r clusters, and then transfers the clusters to the optimal storage location construction module;

[0123] The optimal storage location construction module uses the kernel ridge regression model to train the products in each cluster Ct (t∈[1,r]) and optimizes its regularization parameters and kernel parameters through the particle swarm optimization algorithm to obtain the optimal storage location model QTt corresponding to the products in Ct, and transmits QTt to the new product storage location confirmation module;

[0124] The new product storage location confirmation module is based on the fact that when a new product enters the warehouse, this module scans the RFID tag of the new product through an RFID reader to obtain its full information. It then uses the trained GAT model to extract its features and calculates the distance between it and the feature center of each cluster. It finds the cluster with the smallest distance Cmin and substitutes the features into the optimal storage location model Qmin corresponding to Cmin to obtain the optimal storage location for the new product. This is then transmitted to the alternative path selection module.

[0125] The alternative path selection module uses a backtracking algorithm to generate all alternative paths from the warehouse entrance to the storage location based on the optimal storage location of the new product, and uses the Pareto algorithm to perform multi-objective optimization on the alternative paths to obtain the optimal alternative path.

[0126] Some of the data in the above formulas are dimensionless and numerically calculated. Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0127] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. An RFID warehouse logistics management method, characterized in that: The following steps are involved: S1. Construct a product feature database and an optimal storage location database based on warehouse historical information; S2. Use the graph attention network (GAT) to extract features from the product feature database. The specific operations are as follows: standardize the product data; construct the association graph G = (V, E) between each product; determine the hyperparameters of the GAT layer, including the number of attention heads K and the output feature dimension dout; initialize the weight matrix and attention coefficient vector of each attention head; and use the Latin Hypercube (LHS) to sample the neighboring nodes of each node. Calculate the attention coefficient of each node vi and its neighbor node vj after LHS under each attention head k in, is the set of neighbor nodes after sampling, the formula of LeakyReLU is LeakyReLU(a)=max(0.01a,a); get the aggregated features of node vi under each attention head k in, is the normalized attention weight; the features under each attention head k are After splicing, the final output feature of node vi is obtained as Among them, concat() is the concatenation function; S3. Based on the features extracted by GAT, the HDBSCAN density clustering algorithm is used to cluster the products and obtain r clusters; S4. Use the kernel ridge regression model to train the commodities in each cluster Ct, and optimize its regularization parameters and kernel parameters through the particle swarm optimization algorithm to obtain the optimal storage location model QTt. The specific operation is as follows: Construct a data set E1 = {(Y1, U1), (Y2, U2), ..., (Yn, Un)}, where Yi∈R dout is the input feature vector extracted by GAT, Ui is the data in the optimal storage location database; it is divided into 70% training set and 30% test set; the particle swarm size is set to N, the inertia weight is w, the learning factors are c1 and c2, and the position of particle i can be expressed as Q i =(λ i ,σ1 i ), speed is expressed as V i =(v λ,i ,v σ1,i ), initialize the particle's velocity V i is a 0 vector, Q i Randomly generated within the set parameter range, the initial individual optimal position pbesti=Q i , the global optimal position is gbest; for [0.7*n] samples in the training set, according to the selected Gaussian kernel function, Get the kernel matrix K∈R [0.7*n]×[0.7*n] , where [] is the rounding symbol; for each particle, its position parameter Q i Substituted into the kernel ridge regression model, the objective function of the kernel ridge regression model is By (K T K+λ i I) β=K T U calculates the coefficient vector β; uses k-fold cross validation to train and verify the obtained kernel ridge regression model on the training set, and calculates the mean square error MSE of the model as the fitness value Fi, that is, Among them, MSE j is the mean square error of the j-fold validation, nj is the number of samples in the j-fold validation set, is the true value, is the predicted value; compare the current fitness value of each particle and update pbesti and gbest; update the speed and position of the particle: The obtained global optimal position gbest is used as the optimal regularization parameter and kernel parameter of the kernel ridge regression model, and the training set is substituted into the retrained model to obtain QTt; QTt is evaluated using the test set. If the effect is not good, the iterative optimization is continued until the optimal QTt is found; S5. When a new product enters the warehouse, the RFID tag of the new product is scanned by an RFID reader to obtain all its information. The trained GAT model is then used to extract its features and calculate the distance between it and the feature center of each cluster. The cluster with the smallest distance Cmin is found and the features are substituted into the optimal storage location model Qmin corresponding to Cmin to obtain the optimal storage location for the new product. S6. Based on the optimal storage location of the new product, a backtracking algorithm is used to generate all alternative paths from the warehouse entrance to the storage location, and the Pareto algorithm is used to perform multi-objective optimization on the alternative paths to obtain the optimal alternative path.

2. The RFID warehouse logistics management method according to claim 1, characterized in that: The specific process of LHS sampling includes: Number the nodes in the neighbor node set Ni in random order; Divide each dimension into m equal probability intervals; For each divided interval, a random number generator is used to generate a random number within the interval; Match the randomly selected points in each interval with the values ​​of the corresponding dimensions in the neighbor node set; In multiple dimensions, select nodes that match the interval of the randomly selected point multiple times.

3. The RFID warehouse logistics management method according to claim 1, characterized in that: Use the HDBSCAN density clustering algorithm to cluster products: After n product data are processed by GAT, the feature vector of each product is z i , the dimension is dout, and it is assembled into an n×dout matrix Y, where Ypq represents the value of the p-th product on the q-th feature dimension; For any two products i and j, the similarity is calculated according to the formula d(Yi,Yj) = ||Yi-Yj||; Determine a neighborhood radius ε and the minimum number of points MinPts contained in the neighborhood; Find all points within the neighborhood radius ε, that is, point j that satisfies d(Yi,Yj)<=ε, and record it as set N ε (i); Calculate the set N ε The distances from the point i to the point i in (i) are MinPts. After sorting these distances, the distance with the largest MinPts value is taken as the core distance coredist(i). If the number of points in the set is less than MinPts, the core distance is defined as infinite. Based on the core distance, the direct density accessibility, density accessibility and density connection between commodities are judged; For any core point i, all points densely connected to it form a cluster, and all core points are traversed until all points are divided into corresponding clusters or marked as noise points.

4. The RFID warehouse logistics management method according to claim 1, characterized in that: The specific steps of using the Pareto algorithm to select alternative paths are: The values ​​of path length, number of turns, labor cost, equipment loss cost, and waiting time factors of each alternative path are organized into a matrix form. If there are m1 alternative paths in total, the matrix A1 is constructed as Where li represents the path length of the i-th path, ci represents the number of turns, hi represents the labor cost, ei represents the equipment loss cost, fi represents the waiting time, and i∈[1,m]; Compare the paths. If path i is not worse than path j in all factors and has at least one factor better than path j, then path i dominates path j. Find all paths that are not dominated by other paths to form the first-level non-dominated set F1. Repeat the process for the remaining paths until all paths are divided into the corresponding non-dominated layer. For each path in the non-dominated layer, calculate its congestion distance; Consider the paths at the front of the non-dominated layer. If there are many selectable paths in the same non-dominated layer, choose the path with the larger congestion distance.

5. An RFID warehouse logistics management system, used to implement the RFID warehouse logistics management method according to any one of claims 1 to 4, characterized in that: include: Historical information database construction module, feature extraction module, product clustering module, optimal storage location construction module, new product storage location determination module, and alternative path selection module; A historical information database construction module, which constructs a commodity feature database and an optimal storage location database based on the warehouse historical information and transmits them to the feature extraction module; Feature extraction module, which uses the graph attention network (GAT) to extract features from the product feature database and transmits them to the product clustering module; Commodity clustering module: This module clusters commodities based on the features extracted by GAT using the HDBSCAN density clustering algorithm to obtain r clusters, and then transfers the clusters to the optimal storage location construction module; The optimal storage location construction module uses the kernel ridge regression model to train the products in each cluster Ct, and optimizes its regularization parameters and kernel parameters through the particle swarm optimization algorithm to obtain the optimal storage location model QTt corresponding to the products in Ct, and transmits QTt to the new product storage location confirmation module, where t∈[1,r]; The new product storage location confirmation module is based on the fact that when a new product enters the warehouse, this module scans the RFID tag of the new product through an RFID reader to obtain its full information. It then uses the trained GAT model to extract its features and calculates the distance between it and the feature center of each cluster. It finds the cluster with the smallest distance Cmin and substitutes the features into the optimal storage location model Qmin corresponding to Cmin to obtain the optimal storage location for the new product. This is then transmitted to the alternative path selection module. The alternative path selection module uses a backtracking algorithm to generate all alternative paths from the warehouse entrance to the storage location based on the optimal storage location of the new product, and uses the Pareto algorithm to perform multi-objective optimization on the alternative paths to obtain the optimal alternative path.

Images