Warehouse logistics sniffing method, device and medium based on big data analysis

Abstract

This application discloses a warehouse logistics sniffing method, device, and medium based on big data analysis, relating to the field of logistics technology. The warehouse logistics sniffing method includes: determining a global adjacency matrix of goods based on the local adjacency matrix of goods in all warehouses; determining global multidimensional attribute features of each good based on the local multidimensional attribute features of each good in all warehouses; constructing a probabilistic graphical model using the global adjacency matrix and global multidimensional attribute features as observation data; inferring the probability distribution of sales type of each good through the probabilistic graphical model; constructing a product association graph based on the global adjacency matrix and the probability distribution of sales type of each good; each node in the product association graph carries the local multidimensional attribute features of the good in each warehouse and the historical sales sequence; inputting the product association graph into a time-series graph convolutional network to output the estimated demand of each good in each warehouse; and completing the inventory allocation of each warehouse based on the estimated demand.

Description

Technical Field

[0001] This application relates to the field of warehousing and logistics technology, and in particular to a warehousing and logistics sniffing method, equipment and medium based on big data analysis. Background Technology

[0002] Inventory is a core element for businesses to balance market demand and operating costs, and its proper allocation directly impacts a company's profitability and market competitiveness. Excessive inventory significantly increases holding costs such as warehousing rent, capital tied up, and losses, reducing capital turnover efficiency; while insufficient inventory may lead to stockouts, order delays, and other problems, damaging customer experience and the company's brand reputation.

[0003] Currently, the core logic of existing inventory allocation schemes is to estimate the supply of goods in the future based on historical demand data (such as sales volume and order volume) of individual products over a period of time, using statistical methods (such as moving average and exponential smoothing). However, such schemes are difficult to accurately match actual market demand, easily leading to inventory backlogs or stockouts, thus hindering the improvement of warehousing and logistics efficiency. Summary of the Invention

[0004] This application provides a warehousing and logistics sniffing method, equipment, and medium based on big data analysis. Through big data analysis and intelligent algorithm modeling, it achieves accurate prediction of commodity demand at the warehouse level, effectively reducing inventory backlog and stockout losses caused by prediction errors, and significantly improving the response speed and operational efficiency of warehousing and logistics.

[0005] Firstly, this application provides a warehouse logistics sniffing method based on big data analysis, including:

[0006] Historical sales data of all products in each warehouse are obtained, and the local adjacency matrix of products in each warehouse and the local multidimensional attribute features of each product in each warehouse are calculated based on the historical sales data of all products in each warehouse. The local adjacency matrix of products is determined by the common purchase frequency among the products in the corresponding warehouse.

[0007] A global adjacency matrix for each product is determined based on the local adjacency matrix of all the product warehouses, and global multidimensional attribute features for each product are determined based on the local multidimensional attribute features corresponding to each product in all warehouses. A probabilistic graphical model is constructed using the global adjacency matrix and the global multidimensional attribute features of each product as observation data. The sales type probability distribution of each product is obtained through reasoning using the probabilistic graphical model. The probabilistic graphical model assumes that the global adjacency matrix of the product is generated based on the sales type probability distribution of the product and the correlation strength between sales types, and assumes that the global multidimensional attribute features of the product are generated based on the sales type probability distribution of the product and the correlation strength between sales type and attribute features. The sales type probability distribution represents the probability distribution of a product belonging to different sales types.

[0008] A product association graph is constructed based on the global adjacency matrix of the products and the probability distribution of the sales types of each product. Each node in the product association graph represents a product, and each node carries the local multidimensional attribute features of the corresponding product in each warehouse and the historical sales sequence in each warehouse. Each edge represents the association relationship between products. The product association graph is input into a pre-trained temporal graph convolutional network to output the estimated demand of each product in each warehouse. Based on the estimated demand of each product in each warehouse, the replenishment quantity of each product in each warehouse is determined, and the inventory allocation of each warehouse is completed according to the replenishment quantity.

[0009] In a second aspect, this application provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the warehouse logistics sniffing method based on big data analysis as provided in the first aspect.

[0010] Thirdly, this application provides a storage medium, which is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the warehouse logistics sniffing method based on big data analysis as provided in the first aspect.

[0011] The warehouse logistics sniffing method, equipment, and medium based on big data analysis provided in this application have the following beneficial effects:

[0012] By constructing local adjacency matrices and local multidimensional attribute features of goods based on warehouses as the basic unit, the differentiated characteristics such as regional consumption preferences of each warehouse are preserved. Then, by globally fusing the local adjacency matrices and local multidimensional attribute features of goods from each warehouse, a global adjacency matrix and global multidimensional attribute features of goods are obtained. This integrates the general laws of product association and effectively compensates for the lack of historical data for small sample warehouses and new products. Subsequently, a probabilistic graphical model is used to make structured assumptions, deeply binding product associations, attribute features, and sales types to uncover implicit association patterns between sales types. Finally, a product association graph integrating global product features and regional warehouse features is constructed. Then, by... The product relationship graph is input into a time-series convolutional network. The model performs correlation analysis on the historical sales sequences and local multi-dimensional attribute features of products carried between nodes in each warehouse. On the one hand, it can capture the demand coordination patterns across warehouses. On the other hand, it enables the model to make differentiated predictions by integrating the product sales characteristics of each warehouse based on the global demand pattern. This effectively avoids the problem of global modeling masking regional characteristics, so that the final output of warehouse-level product estimated demand can accurately match the actual demand of each warehouse. The final predicted demand can guide warehouse replenishment and inventory allocation, effectively reducing inventory backlog and stockout losses caused by prediction errors, and significantly improving the response speed and operational efficiency of warehousing and logistics. Attached Figure Description

[0013] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0014] Figure 1 A flowchart illustrating a warehousing logistics sniffing method based on big data analysis provided in an embodiment of this application;

[0015] Figure 2 A framework diagram of the probabilistic graphical model provided in the embodiments of this application;

[0016] Figure 3 This is an example diagram of the product association relationship provided in the embodiments of this application;

[0017] Figure 4 This application provides a network architecture diagram of a temporal graph convolutional network.

[0018] Figure 5 This is a structural block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0020] The core logic of current inventory allocation schemes is to estimate the supply of goods in the future period based on historical demand data (such as sales volume and order volume) of a single product over a past period, using statistical methods (such as moving average and exponential smoothing).

[0021] The applicant found that the above-mentioned scheme analyzed individual commodities in isolation, ignoring the complementary relationships between commodities and failing to capture the interconnected fluctuations in commodity demand. Furthermore, it did not consider the differences in consumer preferences across different distribution areas and did not employ a uniform demand estimation logic, resulting in low alignment between the predicted results and actual regional demand. In addition, for distribution areas with limited data samples, conducting demand analysis in isolation would lead to unreliable results due to insufficient effective data, further exacerbating the bias in demand estimation.

[0022] Based on this, this application provides a warehousing logistics sniffing method based on big data analysis. It takes warehouses as the basic unit, integrates the relationships between commodities and the multi-dimensional attribute characteristics of commodities in each warehouse, and realizes commodity demand prediction from global correlation and regional adaptation through collaborative modeling of probabilistic graphical models and graph neural networks. It not only makes up for the data shortcomings of small sample warehouses through global data fusion, but also realizes differentiated prediction by combining the regional characteristics of warehouses.

[0023] The following section provides a detailed description of the warehousing and logistics sniffing method based on big data analysis proposed in this application, with reference to the accompanying drawings.

[0024] Please see Figure 1 , Figure 1 This is a flowchart illustrating a warehouse logistics sniffing method based on big data analysis, provided as an embodiment of this application. In this embodiment, the warehouse logistics sniffing method based on big data analysis includes steps S10 to S60:

[0025] S10: Obtain historical sales data of all products in each warehouse, and calculate the local adjacency matrix of products in each warehouse and the local multidimensional attribute features of each product in each warehouse based on the historical sales data of all products in each warehouse. The local adjacency matrix of products is determined by the common purchase frequency among the products in the corresponding warehouse.

[0026] It should be noted that "all products" specifically refers to the set of products for which demand forecasting is to be performed. Historical sales data includes the historical sales sequence of products within a preset historical time period (e.g., the past year) and all order numbers associated with each product.

[0027] In this application, the following operations are performed for each sub-warehouse:

[0028] (1) Extract all order numbers associated with each product in warehouse r based on the historical sales data of each product. Based on all order numbers associated with each product in warehouse r, for any product pair (product i, product j), count the number of orders in warehouse r that simultaneously contain product i and product j, and define it as the local co-purchase frequency of product i and product j in warehouse r. As an example, Table 1 shows the frequency of joint purchases of various products in the Guangzhou and Shanghai distribution centers.

[0029] Table 1

[0030]

[0031] Then, the local co-purchase frequency of each product pair in warehouse r is normalized, and the normalized local co-purchase frequency is used as the local association strength of each product pair in warehouse r. Based on the local association strength of all product pairs in warehouse r, a local adjacency matrix of products in warehouse r is constructed. .

[0032] Specifically, the frequency of local co-purchases is calculated using the following formula. Normalization is performed to obtain product i

[0033] Local correlation strength between commodity j and warehouse r :

[0034] ;

[0035] in, This represents the maximum local co-purchase frequency among all product pairs within warehouse r. This represents the minimum frequency of local co-purchase among all product pairs within warehouse r.

[0036] Wherein, the local adjacency matrix of goods in warehouse r. Let r be an N×N matrix, where the element in the i-th row and i-th column represents the local correlation strength between product i and product j in warehouse r. N is the total quantity of goods.

[0037] (2) Feature extraction and quantitative analysis are performed on each product within warehouse r to obtain the local multidimensional attribute features of each product in warehouse r. As an example, the multidimensional attribute features include summer sales ratio, winter sales ratio, sales growth during promotional periods, bundled sales ratio, and monthly sales fluctuation coefficient throughout the year. The definitions of each local attribute feature of the product in warehouse r are as follows:

[0038] Summer sales percentage of a product in warehouse r: The percentage of a product's total annual sales during the summer season within the area covered by warehouse r.

[0039] Winter sales percentage of a product in warehouse r: The percentage of a product's total annual sales during the winter season within the area covered by warehouse r.

[0040] Sales growth rate of a product during the promotional period in warehouse r: The percentage increase in sales of a product during the promotional period relative to sales during the non-promotional period within the coverage area of ​​warehouse r;

[0041] Bundled sales ratio of a product in warehouse r: The proportion of a product's total sales frequency when bundled with other products within the coverage area of ​​warehouse r. The calculation formula is: Bundled sales ratio = Frequency of the product appearing in the same order with other products / Frequency of the product appearing independently in an order + Frequency of the product appearing in the same order with other products.

[0042] Monthly sales volatility coefficient of a product in warehouse r throughout the year: A quantitative value of the degree of fluctuation in monthly sales of a product within the coverage area of ​​warehouse r throughout the year.

[0043] Here, the coverage area of ​​warehouse r refers to the service radius of the goods outbound service corresponding to warehouse r. For example, if warehouses in Guangzhou and Shanghai are set up, when a goods outbound order is generated within the region, orders from cities such as Shenzhen, Zhongshan, and Zhuhai will prioritize retrieving goods from the Guangzhou warehouse, while orders from cities such as Shanghai and Hangzhou will prioritize retrieving goods from the Shanghai warehouse. Table 2 below provides an example of the local multidimensional attribute characteristics of various goods in different warehouses.

[0044] Table 2

[0045]

[0046] S20: Determine the global adjacency matrix of goods based on the local adjacency matrix of goods in all warehouses, and determine the global multidimensional attribute features of each goods based on the local multidimensional attribute features of each goods in all warehouses.

[0047] Specifically, the first weight matrix corresponding to each sub-warehouse is determined, and the first weight matrix of each sub-warehouse is weighted and summed with the local adjacency matrix of the goods in each sub-warehouse to obtain the global adjacency matrix of the goods. At the same time, the multi-dimensional attribute weight coefficients corresponding to each goods in each sub-warehouse are determined, and the local multi-dimensional attribute features of each goods in each sub-warehouse are weighted and summed with the multi-dimensional attribute weight coefficients of each goods in the corresponding sub-warehouse to obtain the global multi-dimensional attribute features of each goods.

[0048] The first weight matrix of a single warehouse is an N×N matrix, where N is the total number of goods. Each element in the global adjacency matrix of the goods is determined by the following formula:

[0049] ,and ;

[0050] in, The value of the element in the i-th row and j-th column of the global adjacency matrix represents the initial global association strength between the i-th and j-th products. Let R represent the value of the element in the i-th row and j-th column of the first weight matrix of warehouse r, corresponding to the first weighting coefficient between the i-th and j-th products in warehouse r, where R is the total number of warehouses. Let i be the frequency of joint purchases of the i-th product and the j-th product in the local warehouse r. The specific meaning of has been defined above and will not be repeated here.

[0051] In addition, the multidimensional attribute weight coefficients of each commodity in each warehouse are determined, and the multidimensional attribute features of each commodity in each warehouse are weighted and summed with the corresponding multidimensional attribute weight coefficients to obtain the global multidimensional attribute features of each commodity.

[0052] Among them, the multidimensional attribute weight coefficients of a single product in a single warehouse are M-dimensional vectors, represented as [ ], Let represent the attribute weight coefficients of the first dimension, the second dimension, ..., the Mth dimension of the i-th product in warehouse r, respectively, where M is the total number of attribute features.

[0053] The global multidimensional attribute features of each product are calculated using the following formula:

[0054] ,and ;

[0055] in, Let R represent the global value of the m-th attribute feature of the i-th product, and R be the total number of warehouses. Let represent the local value of the m-th attribute feature of the i-th product in warehouse r. This represents the attribute weight coefficient of the m-th attribute feature of the i-th product under warehouse r.

[0056] In the above scheme, the global adjacency matrix integrates the correlation patterns of goods in each warehouse, and the global multidimensional attribute features retain the core attribute differences of goods in each warehouse. This avoids the randomness of data from a single warehouse while preserving regional differences, providing highly robust observational data for subsequent model inference. Furthermore, through global data fusion, warehouses with smaller datasets can leverage data on similar goods from larger sample warehouses (e.g., milk powder data from remote small warehouses can be fused with milk powder data from Guangzhou and Beijing warehouses), solving the problem of isolated analysis of small sample warehouses and low reliability of results.

[0057] S30 uses the global adjacency matrix of the products and the global multidimensional attribute features of each product as observation data to construct a probabilistic graphical model, and infers the probability distribution of the sales type of each product through the probabilistic graphical model.

[0058] The probabilistic graphical model assumes that the global adjacency matrix of a product is generated based on the probability distribution of the product's sales type and the correlation strength between sales types. It also assumes that the global multidimensional attribute features of a product are generated based on the probability distribution of the product's sales type and the correlation strength between the sales type and the attribute features. The probability distribution of the sales type represents the probability distribution of a product belonging to different sales types.

[0059] It should be noted that the probability distribution of sales types satisfies the normalization constraint, that is, the sum of the probabilities of all sales types corresponding to a single product is 1.

[0060] In some implementations, sales types include seasonal fluctuations, stable sales, promotion-sensitive products, and bundled add-on products. Seasonally fluctuating products refer to goods significantly affected by seasons, such as air conditioners with high sales in summer and low sales in winter; stable sales products refer to goods with small annual sales fluctuations and no significant seasonal differences, such as face masks; promotion-sensitive products refer to goods with stable sales outside of promotional periods but significant sales increases during promotional periods, such as milk powder; bundled add-on products refer to goods whose sales share bundled with other products is significantly higher than their independent sales share, usually as complementary products or gifts, such as hand cream sold bundled with face masks.

[0061] In some implementations, a probabilistic graphical model is constructed using the global adjacency matrix of products and the global multidimensional attribute features of each product as observation data. The probability distribution of sales type for each product is then obtained through inference using the probabilistic graphical model. This may include:

[0062] Define the nodes and edges of the probabilistic graphical model. The probabilistic graphical model contains three types of nodes: product nodes, sales type nodes, and attribute feature nodes. The edges of the probabilistic graphical model include the association edges between product nodes and sales type nodes, the association edges between sales type nodes, and the association edges between sales type nodes and attribute feature nodes.

[0063] Initialize the model parameters, including: variational parameters of each product node belonging to each sales type node, the first association matrix representing the weight of the association edges between sales type nodes, and the second association matrix representing the weight of the association edges between sales type nodes and attribute feature nodes; execute the iterative optimization process, and in the current iteration of the iterative optimization process, perform the following operations: fix the first and second association matrices of the previous iteration, wherein, in the first iteration, the initialized first and second association matrices are used as the parameters of the previous iteration;

[0064] For each product node, based on its element values ​​corresponding to other products in the global adjacency matrix and its global multidimensional attribute features, the variational parameters of the product node belonging to each sales type node are updated using a variational inference algorithm to obtain the variational approximate distribution of the product's sales type probability distribution in the current iteration round. The variational approximate distributions of each product obtained in the current iteration round are fixed, and the first association matrix of the current iteration round is updated based on the global adjacency matrix of the products, with the goal of maximizing the likelihood of sales type association matching. The second association matrix of the current iteration round is also updated based on the global multidimensional attribute features of each product, with the goal of maximizing the likelihood of sales type matching with attribute features. Based on the variational approximate distributions, the first association matrix, and the second association matrix of the current iteration round, the evidence lower bound of the probabilistic graphical model in the current iteration round is calculated. If the absolute value of the difference between the evidence lower bound of the current iteration round and the evidence lower bound of the previous iteration round is less than a preset threshold, or if the current iteration round reaches a preset maximum number of iterations, the iteration stops, and the probability distribution of each product belonging to different sales types is output based on the variational approximate distributions of each product obtained in the last iteration round.

[0065] Among them, the probabilistic graphical model (PGM) is a mathematical model that uses a graph structure to represent the probabilistic dependencies between variables. In machine learning, it is used to infer unknown information, especially when it is necessary to infer unknown variables from observable variables. This involves the "inference" process of the probabilistic model, the core of which is how to infer the conditional probability distribution P(c|x) of the unknown variable c based on the observable variable x.

[0066] For example, please see Figure 2 , Figure 2 The diagram illustrates the framework of the constructed probabilistic graphical model, where black circles represent attribute feature nodes, white circles represent sales type nodes, and gray circles represent product nodes. Figure 2 In This represents the m-th attribute feature node. This represents the k-th sales type node. This represents the i-th product node.

[0067] In the probabilistic graphical model constructed in this application embodiment, all edges are weighted associated edges and are all undirected edges. The weight of the associated edge represents the strength of the association between nodes, and is specifically defined as follows:

[0068] First type of edge: An association edge between a product node and a sales type node, let... Represents product nodes (i.e., the i-th product node) and the sales type node The greater the weight of the edge connecting the k-th sales type node to the product node, the stronger its representation. Sales type node The higher the probability, the better, and for a single product node, the sum of the weights of the associated edges between that product node and all sales type nodes is 1.

[0069] The second type of edge: the association edge between sales type nodes, let Represents the sales type node With sales type node (i.e., the first) The weight of the edges connecting the sales type nodes, and the larger the weight, the stronger the connection between the sales type nodes. With sales type node The stronger the correlation between them, the more likely they are to be related. The value range is [0, 1].

[0070] The third type of edge: the association edge between the sales type node and the attribute feature node, let Represents attribute feature nodes (i.e., the t-th attribute feature node) and the sales type node The weight of the edges connecting the nodes, and the larger the weight, the more characteristic the node is. With sales type node The stronger the correlation between them, The value range is [0, 1].

[0071] Then, assign uniformly random numbers in the interval [0, 1] to the variational parameters of each product node belonging to each sales type node, and let... Let represent the variational parameter indicating that the i-th product node belongs to the k-th sales type node, used to approximate the probability that the i-th product node belongs to the k-th sales type node. Also, initialize the first association matrix. (It can be initialized randomly or based on prior knowledge), where, Let K be a K×K square matrix, where K is the number of sales types, and the first correlation matrix. elements in This indicates that the k-th sales type node is related to the k-th sales type node. The weights (or strengths) of the association edges between nodes of different sales types. For example, if the sales types include seasonal fluctuations, stable sales, promotion-sensitive sales, and bundled additions, then the second association matrix... Generate a 4×4 matrix. Also, initialize the second incidence matrix. (It can be initialized randomly or based on prior knowledge), where, Let M be a K×M matrix, where M represents the number of attribute features, and the second correlation matrix... elements in Represents attribute feature nodes With sales type node The weights of the edges connecting them.

[0072] Next, model learning is performed. During model learning, the edge weights of the probabilistic graphical model are continuously optimized through alternating iterations of "fixing some parameters → updating other parameters," minimizing the error between the model's observed data (i.e., the global adjacency matrix of the products and multi-dimensional attribute features) and the actual data, thus obtaining the final probability distribution of the sales type for each product. Each iteration's learning process includes step ad:

[0073] Step a: Fix the first association matrix in round t. and the two-fold correlation matrix The value of t (initially t=0) is used in conjunction with real observation data (global adjacency matrix of goods and global multidimensional attribute features of each goods). Mean Field Variational Inference (MFVI) is used to optimize the variational approximation distribution of the sales type probability distribution of goods, so that the variational approximation distribution is closer to the real sales type probability distribution, where t is an integer.

[0074] Make product nodes The variational objective function is expressed as: ,in, Representing variational parameters The variational objective function value, This represents the i-th product node (i.e. ) belongs to the kth sales type node (i.e. The variational parameters of ) Let represent the variational approximation distribution of the i-th product node belonging to the k-th sales type node. This is a learnable parameterized distribution, such as a Gaussian distribution, used to approximate the true posterior distribution. In a given first correlation matrix Second Correlation Matrix Under these conditions, commodity nodes With sales type node The joint probability of simultaneous occurrence, E, is the mathematical expectation.

[0075] Then, solve for the optimal solution of the variational objective function:

[0076] ,

[0077] in, This represents the variational parameter indicating that, in the (t+1)th round, the i-th product node belongs to the k-th sales type node. This represents the initial global association strength between the i-th product and the j-th product. This indicates that the k-th sales type node is related to the k-th sales type node. The weights of the edges connecting nodes of each sales type. This represents the global value of the m-th attribute feature of the i-th product node. This represents the weight of the edge connecting the k-th sales type node and the m-th attribute feature node, where N is the number of product nodes and M is the number of attribute features. This indicates that in the t-th round, the j-th product node belongs to the t-th product node. Variational parameters for each sales type node, It is an exponential function.

[0078] Finally, the variational parameters of all product nodes belonging to each sales type node are integrated, and the variational parameter matrix for round t+1 is output. .

[0079] Step b: Update the first correlation matrix, including:

[0080] Fix the variational parameter values ​​for each product node belonging to each sales type node, i.e., fix Without changing the matrix, update each element in the first correlation matrix, that is:

[0081] ,

[0082] in, This indicates that the k-th sales type node in round t+1 is related to the k-th sales type node. The weights of the edges connecting nodes of each sales type. This represents the variational parameter indicating that, in the (t+1)th round, the i-th product node belongs to the k-th sales type node. This indicates that in the (t+1)th round, the j-th product node belongs to the _th_ ... The variational parameters for each sales type node are defined above, while the definitions of other parameters are provided in the relevant content above.

[0083] By integrating the weights of the edges between all sales type nodes obtained through updates, we obtain the first association matrix for round t+1. .

[0084] Update the second association matrix ,include:

[0085] fixed With the value of unchanged, update each element in the second correlation matrix, that is:

[0086] ,

[0087] in, This represents the weight of the association edge between the k-th sales type node and the m-th attribute feature node in the (t+1)-th round. Other parameters are defined above and will not be repeated here.

[0088] Finally, by updating the weights of the association edges between all sales type nodes and all attribute feature nodes, the second association matrix for the (t+1)th round is obtained. .

[0089] Step c: Calculate the lower bound of evidence (i.e., ELBO value) for the current iteration. The core meaning of the ELBO value is: lower bound of evidence = log-likelihood - KL divergence. The higher the ELBO value, the better the model fits the observed data, and the closer the variational approximation distribution is to the true posterior distribution.

[0090] based on , , Based on real observation data, the ELBO calculation formula is constructed:

[0091] ,

[0092] in, This represents the lower bound of evidence in the (i+1)th round. Let represent the variational parameter indicating that the i-th product node belongs to the k-th sales type node at iteration t+1. This is a latent variable that the model wants to infer. Let represent the variational approximation distribution of the i-th product node belonging to the k-th sales type node at iteration t+1. This is a learnable parameterized distribution, such as a Gaussian distribution, used to approximate the true posterior distribution. The conditional likelihood logarithm measures the probability of observing product data given model parameters and latent variables. This represents the data in the i-th row of the global adjacency matrix for products, indicating the initial global association strength between the i-th product and all other products. This represents the global multidimensional attribute features of the i-th product node. Let represent the first incidence matrix obtained in the (t+1)th iteration (updated from the first incidence matrix in the tth iteration). This represents the second incidence matrix obtained in iteration t+1. Indicates that in the known , , Under the premise of model generation and The probability logarithm; KL is an abbreviation for Kullback-Leibler Divergence, also known as relative entropy, and in this application is used to measure the difference between the approximate distribution and the true distribution of the probability of a product's sales type. By constraining the magnitude of this difference, it is ensured that the variational parameters obtained after iterative optimization accurately reflect the true probability of the product's sales type. This represents the prior distribution (such as empirical prior based on historical data or uniform distribution) that the i-th product node belongs to the k-th sales type node at the (t+1)-th iteration, representing the initial assumption about the product sales type. Represents the variational approximation distribution With prior distribution The difference between them is used to constrain the variational approximation distribution. Do not deviate from the prior distribution. Too far.

[0093] Step d: If the absolute value of the difference between the ELBO values ​​of two adjacent rounds is less than the preset threshold ε or the current iteration round reaches the maximum number of iterations T, then stop the iteration and output the variational approximation distribution of the sales type probability distribution of each product in the last round as the sales type probability distribution of each product. For example, if the condition that the absolute value of the difference between the ELBO values ​​of the 701st round is less than the preset threshold is met, then the variational approximation distribution of the sales type probability distribution of each product in the 701st round is taken as the final sales type probability distribution of each product.

[0094] As an example, T is 1000.

[0095] S40. Construct a product association graph based on the global adjacency matrix of products and the probability distribution of sales types of each product. Each node in the product association graph represents a product, and each node carries the local multidimensional attribute features of the corresponding product in each warehouse and the historical sales sequence in each warehouse. Each edge represents the association between products.

[0096] In some implementations, the initial global association strength between products is obtained based on the global adjacency matrix of products, and the sales type similarity between products is calculated based on the sales type probability distribution of each product. Then, the final global association strength between each product is calculated based on the initial global association strength and sales type similarity between each product. Using products as nodes, association edges are constructed for any two product nodes whose final global association strength is greater than a preset association threshold to obtain a product association graph.

[0097] More specifically, assuming that after determining the probability distribution of sales types for each product through step S30 above, let... , Let represent the probability distribution of the sales type of the i-th product. Let represent the probability that the i-th product belongs to the first sales type, the probability that the i-th product belongs to the second sales type, ..., the probability that the i-th product belongs to the k-th sales type, and let be the probability that the i-th product belongs to the first sales type. =[ ], This represents the probability distribution of the sales type of the j-th product. Let represent the probabilities that the j-th product belongs to the first sales type, the second sales type, ..., the K-th sales type, respectively. Then, the sales type similarity between the i-th and j-th products is calculated using the following formula:

[0098] ,

[0099] in, This represents the similarity in sales type between the i-th product and the j-th product. Let represent the probability that the i-th product belongs to the k-th sales type and the probability that the j-th product belongs to the k-th sales type, respectively.

[0100] Then, the final association strength between the i-th product and the j-th product is calculated using the following formula:

[0101] .

[0102] in, Let represent the final association strength and the initial global association strength between the i-th product and the j-th product, respectively.

[0103] Finally, using products as nodes, edges are created for product pairs whose final global association strength is greater than or equal to a preset association threshold (e.g., 0.6). The weight of the edge is represented by the final global association strength of the corresponding product pair. Figure 3 As shown, assume there are 5 product nodes, namely... , Figure 3 In this context, 'e' represents an edge connecting nodes. Each edge carries its own weight, which is represented by the final global association strength. Figure 3 In These represent the edges between the first and second product nodes, respectively. The meanings of the parameters labeled for the remaining edges follow the same logic and will not be listed individually. Each node carries the local multi-dimensional attribute features of the corresponding product in each warehouse and the sales sequence within the target historical time period of each warehouse. The target historical time period is a preset adjacent historical time period from the current time, such as the past month.

[0104] By integrating global product relationships (edges of the graph), probability distribution of product sales types, local multidimensional attribute features of warehouses, and historical sales time-series data of products in each warehouse, the model can, on the one hand, uncover the synergistic patterns of product demand across warehouses, breaking the isolation limitations of data from a single warehouse; on the other hand, each node in the product relationship graph carries the local multidimensional attributes and historical sales sequences of the corresponding product in each warehouse, capturing both the fluctuation patterns of product demand over time and preserving the regional consumption characteristics differences between different warehouses, thus making subsequent predictions of product demand at the warehouse level more accurate.

[0105] S50 inputs the product relationship graph into a pre-trained temporal graph convolutional network and outputs the estimated demand for each product in each warehouse.

[0106] In some implementations, a product association graph is input into a temporal graph convolutional network, which includes an input layer, multiple stacked spatiotemporal convolutional modules, and an output layer. Then, the input layer performs preprocessing of the graph structure and node features of the product association graph. Next, multiple spatiotemporal convolutional modules extract the temporal features of sales volume of each product in each warehouse and the inter-product association features. The temporal features of sales volume of each product in each warehouse and the inter-product association features are then fused to obtain fused features. Finally, the output layer maps the fused features to the estimated demand of each product in each warehouse.

[0107] Among them, the Temporal Graph Convolutional Network (TGCN) adopts a spatiotemporal stacked architecture, such as... Figure 4 As shown, the temporal graph convolutional network mainly consists of three parts: an input layer, multiple stacked spatiotemporal convolutional modules (…). Figure 4 (Only two are shown, but this application is not limited to these), and the output layer. The following provides a detailed description of each network layer.

[0108] Input Layer: This layer receives the input product relationship graph, performs graph structure normalization and node feature standardization, and provides the basic input data for the subsequent spatiotemporal convolution module. The product relationship graph can be represented as follows: V represents the set of nodes, corresponding to N product nodes, and E represents the set of edges, the weight of which is determined by the final global association strength between products. Let S be the set of edge weights, where the weight of each edge represents the final global association strength of the corresponding product pair. Each product node corresponds to a multi-dimensional feature vector with dimensions R×(M+T), where R is the total number of warehouses, M is the number of attribute features for each product, and T is the length of the historical sales sequence. By summing the multi-dimensional feature vectors of all product nodes, we obtain the feature matrix S, which has dimensions [N, R×(M+T)], where N is the total number of products.

[0109] The steps of input layer normalizing graph structure include: setting edge weights. Construct the target adjacency matrix H, which is an N×N matrix, and If product i and product j have no associated edges, then It is then normalized using the following formula:

[0110] ,

[0111] Where D is the degree matrix of the target adjacency matrix, and I is the identity matrix. This is the normalized target adjacency matrix.

[0112] The steps for standardizing node features in the input layer include: standardizing the feature matrix S by dimension, where attribute features are normalized to [0, 1] and sales sequences are normalized to [−1, 1].

[0113] Spatiotemporal Convolution Module: Each spatiotemporal convolution module consists of two core components: a temporal convolution component and a graph convolution component. In this application, the order of temporal convolution followed by graph convolution is adopted.

[0114] The temporal convolution component performs convolution operations on the historical sales sequence of each product node to capture its temporal dependency patterns. Specifically, it extracts data from the normalized target adjacency matrix. In the middle, attribute feature sub-matrices are split according to the compartments. and sales sequence submatrix .in, All local multidimensional attribute features of each compartment are preserved and do not participate in temporal convolution. The historical sales sequences of all warehouses are used as the input for temporal convolution. Then, the temporal convolution operation is performed in parallel on each warehouse: Restructured according to the warehouse division dimension Each compartment is configured with an independent one-dimensional convolutional (Conv1D) kernel to achieve parallel extraction of temporal features from each compartment. Specific parameter design includes:

[0115] Convolution kernel configuration: kernel size = 3 to capture sales fluctuations over 3 consecutive time steps, number of input channels = 1, number of output channels = stride=1, padding=1, to ensure that the sequence length after convolution is still T;

[0116] Nonlinear activation: After convolution, the ReLU function is used to filter out negative temporal feature contributions;

[0117] Temporal dimension reduction: Convolution output for each compartment (dimension [N, ... Global average pooling is performed on [T] to reduce the time dimension T to 1, thus obtaining the time-series features of the warehousing division. .

[0118] Next, time-series-attribute feature fusion is performed, and the retained attribute feature submatrix is... Remodeling , and time series characteristics By concatenating the data along the channel dimension, we obtain the final output of the temporal convolutional component. : .

[0119] The graph convolution component takes the output of the temporal convolution as input and utilizes the structural information of the product association graph to aggregate the features of core products and neighboring products, uncovering spatial association patterns (such as the increased demand for hot water bottles driving the demand for electric blankets). Specifically, the graph convolution component first processes the output of the temporal convolution... Flattening the graph along the partition dimension yields the input feature matrix for graph convolution: Then, first-order graph convolution (GCN) is used for spatial feature aggregation, with the input being... and The calculation formula is: .in, The learnable weight matrix has dimensions [ ],in This represents the number of output channels of the graph convolution, i.e., the hidden layer dimension, typically set to 64 or 128. The bias term has the following dimensions: , For activation functions, ReLU is preferred to avoid gradient vanishing. This formula uses... Weighting is applied to allocate features from neighboring products to the core product based on their association strength; products with stronger associations receive higher feature weights. Finally, the graph convolutional component outputs... This feature matrix integrates the spatiotemporal characteristics of the product's own distribution and the correlation characteristics of neighboring products, achieving preliminary coordination between temporal and spatial aspects.

[0120] When multiple spatiotemporal convolutional modules are stacked, the graph convolution output of the previous spatiotemporal convolutional module is used as the input of the next spatiotemporal convolutional module, and residual connections are added between adjacent spatiotemporal convolutional modules. The module input and output features are concatenated by dimension and then ReLU activation is performed to avoid the gradient vanishing problem of deep networks.

[0121] Output Layer: The output layer maps the deep spatiotemporal features output by the last spatiotemporal convolutional module to the estimated demand for each commodity in each warehouse at future time steps, achieving the goal of multiple outputs from a single node. The output layer uses a fully connected layer, and the mapping of the fully connected layer uses a single fully connected layer. The input is the output of the last spatiotemporal convolutional module, and the output dimension is designed according to the prediction demand: if it is necessary to predict the demand in the next S time steps, then the output dimension of the fully connected layer is [N, R×S], where R is the number of warehouses and S is the number of prediction time steps.

[0122] The formula for calculating a fully connected layer is: ,in, The output of the output layer. These represent the output of the last spatiotemporal convolutional module, the output layer weight matrix, and the output layer bias term, respectively. The dimension is [ ], The dimension is R×S. Finally, Reshape to [N, R, S].

[0123] In some implementations, the training process of a temporal graph convolutional neural network includes: first acquiring and preprocessing a training set; then inputting the preprocessed training set into a pre-constructed temporal graph convolutional neural network for forward propagation to output demand prediction values; and finally, applying a loss function. The prediction error is quantified, and the network parameters are iteratively optimized based on this error to achieve model training convergence. The loss function is... Represented as:

[0124] ;

[0125] ,

[0126] ,

[0127] ;

[0128] in, To weight the multi-dimensional MSE main loss, in order to balance the prediction accuracy of different sub-accounts and different time steps, This is a non-negative positive expression term that penalizes cases where the predicted value is less than 0 (demand has no negative value). This is an extremum regularization term that penalizes predicted values ​​that far exceed historical peaks, in order to avoid inventory buildup. : represents the regularization coefficient, used to balance the contribution of the main loss and the regularization term; S is the prediction time step, and the minimum value of S is 1. This represents the weight of the sub-warehouse r, which can be determined based on the sales proportion of the sub-warehouse. This represents the proportion of the prediction time step 's'; the smaller 's' is, the larger the weight. This represents the predicted demand for the i-th product at a time step s under warehouse r. This represents the actual demand for the i-th item at time step s under warehouse r, and is a label carried in the training samples. This represents the historical peak sales volume of the i-th product in warehouse r.

[0129] Through the above approach, when the temporal graph convolutional network learns by integrating the global product relationship graph with the product relationship graph of the warehouse features, it can accurately capture the collaborative patterns of product demand across warehouses. For example, the demand for winter warming products and dry season skin care products in all warehouses in the northern region will show a synchronous growth pattern. On the other hand, it can rely on the local multi-dimensional attribute features of each product in the corresponding warehouse (such as the proportion of summer sales and the sales increase during the promotion period) to allow the model to further integrate the regional attribute features of each warehouse on the basis of the global demand pattern for differentiated prediction. This effectively eliminates the masking of regional characteristics by global modeling and ultimately greatly improves the prediction accuracy of warehouse-level product demand.

[0130] S60 determines the replenishment quantity of each product in each warehouse based on the estimated demand for each product in each warehouse, and completes the inventory allocation of each warehouse according to the replenishment quantity.

[0131] Specifically, the current inventory, safety stock, and in-transit inventory of each product in each warehouse are obtained. Then, the replenishment quantity of product i in warehouse r for the most recent forecast time step is determined using the following formula: Estimated demand for product i in warehouse r + Safety stock of product i in warehouse r - Current inventory of product i - In-transit inventory of product i heading to warehouse r. Finally, replenishment is performed for each warehouse according to the corresponding inventory levels of each product.

[0132] This application embodiment constructs a local adjacency matrix and local multidimensional attribute features of goods based on warehouses as the basic unit, in order to retain the differentiated characteristics such as regional consumption preferences of each warehouse. Then, by globally fusing the local adjacency matrix and local multidimensional attribute features of goods in each warehouse, a global adjacency matrix and global multidimensional attribute features of goods are obtained. This not only integrates the general laws of product association, but also effectively makes up for the lack of historical data for small sample warehouses and new products. Then, through a probabilistic graphical model, structural assumptions are made to deeply bind product association, attribute features and sales types, and to explore the implicit association laws between sales types. Finally, a product association graph that integrates global product features and regional features of warehouses is constructed. Subsequently, by inputting the product relationship graph into the time-series graph convolutional network, the model performs correlation analysis on the historical sales sequences and local multi-dimensional attribute features of products carried between nodes in each warehouse. On the one hand, it can capture the demand coordination patterns across warehouses. On the other hand, it enables the model to integrate warehouse characteristics for differentiated prediction based on global demand patterns, effectively avoiding the problem of global modeling masking regional characteristics. This ensures that the final output of warehouse-level product demand estimates can accurately match the actual demand of each warehouse. The final predicted demand can guide warehouse replenishment and inventory allocation, effectively reducing inventory backlog and stockout losses caused by prediction errors, and significantly improving the response speed and operational efficiency of warehousing and logistics.

[0133] Accordingly, see Figure 5 , Figure 5 This is a structural block diagram of an electronic device provided in an embodiment of this application. In this embodiment, the electronic device 10 includes a memory 101, a processor 100, and a computer program 102 stored in the memory 101 and executable on the processor 100. When the processor 100 executes the computer program 102, it implements all the steps of the warehouse logistics sniffing method based on big data analysis provided in the above embodiment.

[0134] The electronic device can be a mobile computing device (such as a mobile phone), a desktop computer, a laptop, a PDA, and a cloud server, etc. This terminal device may include, but is not limited to, a processor 100 and a memory 101. Those skilled in the art will understand that... Figure 5 This is merely an example of an electronic device and does not constitute a limitation on electronic devices. It may include more or fewer components than shown in the illustration, or combinations of certain components, or different components. For example, it may also include input / output devices, network access devices, etc.

[0135] The processor 100 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0136] In some embodiments, memory 101 may be an internal storage unit of an electronic device, such as a hard disk or memory. In other embodiments, memory 101 may be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. Furthermore, memory 101 may include both internal and external storage units. Memory 101 is used to store operating systems, applications, bootloaders, data, and other programs, such as program code for computer programs. Memory 101 can also be used to temporarily store data that has been output or will be output.

[0137] Accordingly, this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the warehouse logistics sniffing method based on big data analysis as provided in the above embodiments.

[0138] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to a breastfeeding aid / electronic device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium.

[0139] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A warehouse logistics sniffing method based on big data analysis, characterized in that, include: Historical sales data of all products in each warehouse are obtained, and the local adjacency matrix of products in each warehouse and the local multidimensional attribute features of each product in each warehouse are calculated based on the historical sales data of all products in each warehouse. The local adjacency matrix of products is determined by the common purchase frequency among the products in the corresponding warehouse. The global adjacency matrix of a product is determined based on the local adjacency matrix of all the products in the warehouses, and the global multidimensional attribute features of each product are determined based on the local multidimensional attribute features of each product in all the warehouses. A probabilistic graphical model is constructed using the global adjacency matrix of the products and the global multidimensional attribute features of each product as observation data. The sales type probability distribution of each product is obtained through inference using the probabilistic graphical model. The probabilistic graphical model assumes that the global adjacency matrix of the products is generated based on the sales type probability distribution of the products and the correlation strength between sales types, and that the global multidimensional attribute features of the products are generated based on the sales type probability distribution of the products and the correlation strength between sales types and attribute features. The sales type probability distribution represents the probability distribution of products belonging to different sales types. A product association graph is constructed based on the global adjacency matrix of the products and the probability distribution of the sales type of each product. Each node in the product association graph represents a product, and each node carries the local multidimensional attribute features of the corresponding product in each warehouse and the historical sales sequence in each warehouse. Each edge represents the association relationship between products. The product association graph is input into a pre-trained temporal graph convolutional network, which outputs the estimated demand for each product in each of the warehouses. The replenishment quantity of each product in each of the aforementioned warehouses is determined based on the estimated demand for each product, and the inventory allocation of each of the aforementioned warehouses is completed according to the replenishment quantity.

2. The big data analytics based warehouse logistics sniffing method of claim 1, wherein, The process of determining the global adjacency matrix of goods based on the local adjacency matrix of goods in all the warehouses, and determining the global multidimensional attribute features of each good based on the local multidimensional attribute features of each good in all the warehouses, includes: Obtain the first weight matrix corresponding to each of the sub-warehouses, and perform a weighted summation operation between the first weight matrix of each sub-warehouse and the local adjacency matrix of the goods in each sub-warehouse to obtain the global adjacency matrix of the goods; In addition, the multidimensional attribute weight coefficients corresponding to each product in each of the sub-warehouses are obtained, and the local multidimensional attribute features of each product in each of the sub-warehouses are weighted and summed with the multidimensional attribute weight coefficients of each product in the corresponding sub-warehouses to obtain the global multidimensional attribute features of each product.

3. The big data analytics based warehouse logistics sniffing method as claimed in claim 1, wherein, The step of constructing a probabilistic graphical model using the global adjacency matrix of the products and the global multidimensional attribute features of each product as observation data, and inferring the probability distribution of the sales type of each product through the probabilistic graphical model, includes: Define the nodes and edges of the probabilistic graphical model, wherein the probabilistic graphical model includes three types of nodes: product nodes, sales type nodes, and attribute feature nodes, and the edges of the probabilistic graphical model include the association edges between product nodes and sales type nodes, the association edges between sales type nodes, and the association edges between sales type nodes and attribute feature nodes. Initialize the model parameters, including: variational parameters of each product node belonging to each sales type node, the first association matrix representing the weight of the association edges between sales type nodes, and the second association matrix representing the weight of the association edges between sales type nodes and attribute feature nodes; Execute the iterative optimization process. In the current iteration of the iterative optimization process, perform the following operations: Fix the first and second correlation matrices of the previous iteration round. In the first iteration, the first and second correlation matrices set in the initialization are used as the first and second correlation matrices of the previous iteration round. For each product node, based on the element values ​​of its corresponding other products in the global adjacency matrix of the product and the global multidimensional attribute features, the variational parameters of the product node belonging to each sales type node are updated based on the variational inference algorithm to obtain the variational approximate distribution of the sales type probability distribution of the product in the current iteration round. The variational approximation distribution of each product obtained in the current iteration round is fixed. With the goal of maximizing the likelihood of sales type association matching, the first association matrix of the current iteration round is updated based on the global adjacency matrix of the product. And, with the goal of maximizing the likelihood of sales type matching with attribute features, the second association matrix of the current iteration round is updated based on the global multidimensional attribute features of each product. Based on the variational approximation distribution, the first correlation matrix, and the second correlation matrix of the current iteration, calculate the evidence lower bound of the probabilistic graphical model in the current iteration. If the absolute value of the difference between the lower bound of evidence in the current iteration and the lower bound of evidence in the previous iteration is less than a preset threshold, or if the current iteration reaches the preset maximum number of iterations, then the iteration stops, and the probability distribution of each product belonging to different sales types is output based on the variational approximation distribution of each product obtained in the last iteration.

4. The big data analytics based warehouse logistics sniffing method according to any one of claims 1 to 3, characterized in that, The step of constructing a product association graph based on the global adjacency matrix of the products and the probability distribution of the sales types of each product includes: The initial global association strength between products is obtained based on the global adjacency matrix of the products; Based on the probability distribution of the sales type of each product, calculate the sales type similarity between the products. The final global association strength between each product is calculated based on the initial global association strength and sales type similarity. Then, using the product as a node, an association edge is constructed between any two product nodes whose final global association strength is greater than a preset association threshold to obtain the product association graph.

5. The big data analytics based warehouse logistics sniffing method according to any one of claims 1 to 3, wherein, The sales types include seasonal fluctuations, stable sales, promotion-sensitive sales, and bundled sales.

6. The warehouse logistics sniffing method based on big data analysis as described in any one of claims 1 to 3, characterized in that, The multidimensional attribute characteristics of the product include summer sales share, winter sales share, sales growth during promotional periods, bundled sales share, and monthly sales fluctuation coefficient throughout the year.

7. The warehouse logistics sniffing method based on big data analysis as described in any one of claims 1 to 3, characterized in that, The step of inputting the product association graph into a pre-trained temporal graph convolutional network and outputting the estimated demand for each product in each of the warehouses includes: The product association graph is input into the temporal graph convolutional network, wherein the temporal graph convolutional network includes an input layer, multiple stacked spatiotemporal convolutional modules, and an output layer; The input layer completes the preprocessing of the graph structure and node features of the product association graph. Then, multiple spatiotemporal convolution modules extract the sales time-series features of each product in each warehouse and the product association features. The sales time-series features of each product in each warehouse and the product association features are then fused to obtain the fused features. The output layer maps the fused features to the estimated demand for each commodity in each of the respective warehouses.

8. The warehouse logistics sniffing method based on big data analysis as described in any one of claims 1 to 3, characterized in that, The process of determining the replenishment quantity of each commodity in each of the aforementioned warehouses based on the estimated demand for each commodity in each warehouse includes: Obtain the current inventory, safety stock, and in-transit inventory of each product in each of the aforementioned warehouses; Based on the current inventory, safety stock, in-transit inventory, and estimated demand of each commodity in each of the aforementioned warehouses, the replenishment quantity of each commodity in each of the aforementioned warehouses is determined.

9. An electronic device, characterized in that, include: The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the warehouse logistics sniffing method based on big data analysis as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the warehouse logistics sniffing method based on big data analysis as described in any one of claims 1 to 8.

Images