Goods management method and system based on intelligent warehouse

Abstract

The invention relates to the technical field of warehouse management, in particular to a cargo management method and system based on an intelligent warehouse. Comprising the following steps: S1, acquiring historical sales data, and extracting sales features of different commodities based on the historical sales data; s2, classifying the commodities based on the sales features, and obtaining sales types of the commodities; s3, according to the sales types and a preset storage fragmentation rule, storage fragmentation is carried out on an inventory database, and different commodity information is distributed to different data fragmentation according to the sales types; s4, subscribing and transmitting the messages of the storage fragments, and updating an inventory database; and updating preset historical inventory monitoring data by using the inventory database to obtain real-time inventory monitoring data. According to different commodity types, different storage fragment distributions are carried out, and on the basis of improving the high-concurrency read-write effect, the throughput and response speed of the whole inventory database can be remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of warehouse management, and particularly relates to a goods management method and system based on an intelligent warehouse. BACKGROUND

[0002] A warehouse management system (WMS) is a software system used to manage and control the daily operations of a warehouse. It can help warehouse managers efficiently track inventory, optimize warehouse layout, manage orders, handle incoming and outgoing shipments, and achieve functions such as inventory monitoring, warehouse operation management, order processing, material scheduling, and goods tracking. To optimize inventory management and improve enterprise operational efficiency, enterprises need to use WMS for inventory monitoring. However, in actual application, during shopping festivals or large-scale promotional activities, the system will face instantaneous high-concurrency access. At this time, the instantaneous order flood may cause the system to process tens of thousands of concurrent requests per second, which involves "inventory reduction" operations for inventory locking, "write operations" for order status synchronization, and "read operations" for real-time data queries to form a mixed high-concurrency scenario. Frequent and intensive warehouse operations will cause the WMS to need to handle a large number of read and write requests at the same time, which may cause inventory data to be inaccurate, processing delays, and other problems. SUMMARY

[0003] In view of the above defects, the purpose of the present application is to provide a goods management method and system based on an intelligent warehouse, which solves the problems of inaccurate inventory data and processing delays when facing a large amount of data read and write.

[0004] To achieve this purpose, the present application adopts the following technical solution: a goods management method based on an intelligent warehouse, comprising the following steps:

[0005] Step S1: Obtain historical sales data, and extract sales features of different goods based on the historical sales data;

[0006] Step S2: Classify the goods based on the sales features, and obtain the sales categories of the goods;

[0007] Step S3: According to the sales categories and a preset storage sharding rule, perform storage sharding on an inventory database, and different goods information is allocated to different data shards according to the sales categories;

[0008] Step S4: Subscribe to and transmit messages of the storage shards, and update the inventory database; update the preset historical inventory monitoring data by using the inventory database to obtain real-time inventory monitoring data.

[0009] Preferably, the step of extracting sales features in step S1 is as follows:

[0010] Step S11: Obtain the time stamp of the historical sales data of the commodity in a time period, and sort the historical sales data according to the time stamp to obtain a data sequence;

[0011] Step S12: Output the spatial correlation of the data sequence in the cycle time by using a geographically weighted regression model as a first feature;

[0012] Step S13: Perform time series recursive convolution on the data sequence to obtain a short-term inventory feature as a second feature;

[0013] Step S14: Perform attention encoding to obtain a third feature;

[0014] Step S15: Use a fully connected neural network and a Softmax function to learn the attention weights of the first feature, the second feature and the third feature;

[0015] Step S16: Use the attention weights to weight and sum the first feature, the second feature and the third feature respectively to obtain the sales feature.

[0016] Preferably, the geographically weighted regression model in step S12 is:

[0017]

[0018] where β0 is a constant term, x i,k represents the historical data of the i-th cycle at time k, ε i is the error term of the i-th cycle, β k is the regression coefficient of time k.

[0019] Preferably, the specific steps in step S14 are as follows:

[0020] Step S141: Take a number of data at the next time point in the time period as new input data, and calculate the query vector, the key vector and the value vector respectively through the new input data;

[0021] Wherein the acquisition formula of the query vector, the key vector and the value vector are as follows respectively:

[0022] q t =W Q x t 、k t =W K x t 、v t =W V x t ;

[0023] W Q 、W K 、W VProjection matrix of query vector key vector and value vector, respectively, x t is the new input data;

[0024] Step S142: Insert the new input data into the data sequence, determine whether the length of the data sequence is greater than the length threshold, if greater, obtain the difference between the length of the data sequence and the length threshold as the length difference, and delete the earliest historical sales data of the data sequence, the deletion length is the length difference, and obtain a new data sequence;

[0025] If less, continue to supplement the new input data until the length of the data sequence is equal to the length threshold, and obtain a new data sequence;

[0026] Step S143: Obtain a set of historical key vectors from the new data sequence, obtain a current time query vector, and use the attention score of the current time query vector and each key vector in the new data sequence;

[0027] The attention score acquisition formula is as follows:

[0028] Wherein q now is the current query vector, k i represents the i-th key vector, represents a scaling factor to prevent dot product from being too large;

[0029] Step S144: Normalize the attention score to obtain a first weight, sum the value vectors according to the first weight, and obtain the third feature;

[0030] The formula for obtaining the weight by normalization is as follows:

[0031] Wherein K is the total length of the new data sequence, a j The attention score of the j-th time query vector in the new data sequence and each key vector in the new data sequence;

[0032] The acquisition formula of the third feature is as follows:

[0033]

[0034] Preferably, the step S2 of classifying the goods based on the sales feature is as follows:

[0035] Input the sales feature into a prediction large model to obtain the sales quantity at a future time;

[0036] Determine the sales category of the goods based on the range in which the sales quantity at the future time is located;

[0037] The prediction large model is specifically as follows:

[0038] Input layer: Sales information is input into the embedding layer and transformed into a high-dimensional vector. The specific formula for obtaining the high-dimensional vector is as follows:

[0039] zi = Embedding(w i );

[0040] Where Embedding() is the embedding layer transformation function, w i The i-th sales feature;

[0041] Transformer layer: The Transformer architecture captures the long-short-term dependencies of high-dimensional vectors and updates the model state. The specific formula for updating the model state is as follows;

[0042] h t =Transformer(h t-1 ,z i );

[0043] h t-1 Let z represent the total hidden state at time t-1. i Represents the i-th high-dimensional vector;

[0044] Output layer: An autoregressive generation method is used to predict the sales quantity at future moments based on historical sales behavior and the current high-dimensional input vector. The formula for predicting the sales quantity at future moments is as follows;

[0045]

[0046] Where y t-1，i z represents the actual sales value of the i-th product within the corresponding time period of the previous year. i Let represent the i-th high-dimensional vector, θ be the model parameters, and f() be the prediction model function.

[0047] Preferably, the storage sharding includes low-load sharding, medium-load sharding, and high-load sharding;

[0048] The QPS of the load sharding is 500 times / second;

[0049] The QPS of the load sharding is 2000 times / second;

[0050] The QPS of the high-load sharding is 5000 times / second.

[0051] Preferably, the update of the inventory database in step S4 is as follows:

[0052] Monitor the query volume for medium-load shards and high-load shards respectively;

[0053] If the query amount of the medium-load shard reaches or is greater than 80% QPS, the storage shard of the commodity is modified to a high-load shard;

[0054] If the query amount of the high-load shard reaches or is greater than 80% QPS, the commodity is split according to the SKU, and each SKU of the commodity is allocated a high-load sub-shard. After a time threshold, it is determined whether the query amount of each high-load sub-shard is greater than a query threshold. If yes, the high-load sub-shard is retained. If no, the corresponding SKU is re-allocated to the same high-load shard.

[0055] A cargo management system based on an intelligent warehouse uses the cargo management method based on the intelligent warehouse, which comprises a feature acquisition module, a classification module, a storage allocation module, and a monitoring module.

[0056] The feature acquisition module is configured to acquire historical sales data and extract sales features of different commodities based on the historical sales data.

[0057] The classification module is configured to classify commodities based on the sales features and acquire sales categories of the commodities.

[0058] The storage allocation module is configured to allocate different commodity information to different data shards for storage according to the sales categories and a preset storage shard rule.

[0059] The monitoring module is configured to subscribe to and transmit messages of the storage shards and update the inventory database.

[0060] The inventory database is used to update preset historical inventory monitoring data to obtain real-time inventory monitoring data.

[0061] One of the above technical solutions has the following advantages or beneficial effects: different storage shards are allocated to different commodity categories, which can improve the high-concurrency read-write effect, significantly improve the overall inventory database throughput and response speed, and effectively avoid inventory data inaccuracy and processing delay. BRIEF DESCRIPTION OF DRAWINGS

[0062] Fig. 1 is a flowchart of an embodiment of the method of the present application.

[0063] Fig. 2 is a structural schematic diagram of an embodiment of the system of the present application. DETAILED DESCRIPTION

[0064] Embodiments of the present application are described below in the accompanying drawings, in which like or similar elements / characters / symbols refer to like or similar elements / characters / symbols throughout the figures. The embodiments described below are exemplary only, and are not to be taken in a limiting sense. The embodiments described below are not intended to be exhaustive or to be limited to the precise form disclosed and many modifications and variations are possible in light of the above teachings.

[0065] In the description of the embodiments of the present application, the terms "first", "second", etc. are used only for the purpose of description, and are not to be construed as indicating or implying relative importance or implying the number of the technical features indicated. Thus, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the embodiments of the present application, the meaning of "a plurality of" is two or more, unless otherwise specifically limited.

[0066] In addition, the terms "first", "second" are used only for the purpose of description, and are not to be construed as indicating or implying relative importance or implying the number of the technical features indicated. Thus, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, unless otherwise stated, the meaning of "a plurality of" is two or more. The specific meaning of the above terms in the present application can be understood in specific cases by those of ordinary skill in the art.

[0067] As shown in Figs. 1-2 A smart warehouse-based cargo management method includes the following steps:

[0068] Step S1: Obtain historical sales data, and extract sales features of different goods based on the historical sales data; the historical sales data includes sales quantity, sales amount, sales channel, return data, and sku data, which can reflect the sales situation of the goods.

[0069] Step S2: Classify the goods based on the sales features, and obtain the sales categories of the goods;

[0070] Step S3: According to the sales categories and a preset storage sharding rule, perform storage sharding on the inventory database, and different goods information is allocated to different data shards according to the sales categories;

[0071] Step S4: Subscribe to the messages of the storage shards and update the inventory database; update the preset historical inventory monitoring data using the inventory database to obtain real-time inventory monitoring data.

[0072] In the prior art, the warehouse information of all goods is stored in the same inventory database. When an order flood occurs, the inventory database processes tens of thousands of concurrent requests, including "inventory locking" operations, "write operations" for order status synchronization, and "read operations" for real-time data queries, forming a mixed high-concurrency scenario. Frequent and intensive warehouse operations can cause the WMS to simultaneously process a large number of read and write requests, which can cause inventory data to be inaccurate, processing delays, and other problems. Therefore, in the present application, based on the sales data of the previous year, the sales characteristics of the same product in this sales are extracted. The sales characteristics can reflect the sales rules of the product. Based on the sales characteristics, the goods can be classified, including low-volume goods, medium-volume goods, and high-volume goods. Generally, the sales volume of a product is related to the page view of the product. High-volume products are browsed and ordered relatively more in a time period, which can generate more data changes. Therefore, the sales categories can be divided according to the sales volume.

[0073] After the sales categories are divided, the inventory database is stored according to the preset storage sharding rule, for example, in an embodiment, including load sharding, medium-load sharding, and high-load sharding. Correspondingly, low-volume goods, medium-volume goods, and high-volume goods are processed using load sharding, medium-load sharding, and high-load sharding, respectively. At this time, high-volume goods can be allocated to high-performance shards to support high-concurrency read and write; and low-volume goods are allocated to shards with lower costs. Avoid hot data competition and significantly improve the overall inventory database throughput and response speed.

[0074] Then, the message queue (such as Kafka, RabbitMQ) is used to subscribe to the change events (warehousing, warehousing, allocation, return) of each storage shard. Any inventory change within a shard is published in the form of a message, which is consumed and updated by a unified inventory database. This ensures:

[0075] 1. Real-time: Changes are almost in real-time reflected to the central inventory database, eliminating batch processing delays.

[0076] 2. Reliability: The persistence and retry mechanism of the message queue ensures that changes are not lost, even in the event of partial system failure.

[0077] 3. Decoupling and flexibility: The front-end business system (such as WMS, ERP) is decoupled from the back-end inventory update logic. The front-end only needs to publish messages, and the back-end processes them as needed. The system is more robust and more flexible.

[0078] Finally, the real-time inventory database manager can real-time master the accurate inventory level and turnover status of each sales category and each shard of goods.

[0079] It is worth mentioning that the sharding architecture is easy to horizontally expand. When the data volume or access volume of a certain type of commodity (such as a new hit product) increases sharply, the resources of the shard where the commodity is located can be expanded independently without the need for overall expansion, reducing the cost and complexity of expansion.

[0080] Preferably, the step of extracting the sales feature in step S1 is as follows:

[0081] Step S11: Obtain the time stamp of the historical sales data of the commodity in a time period, and sort the historical sales data according to the time stamp to obtain a data sequence;

[0082] Step S12: Output the spatial correlation of the data sequence in the cycle time by using a geographically weighted regression model as the first feature; Since the time and regional differences need to be considered in the process of selling the commodity, especially considering overseas sales, for example, in June, the northern hemisphere is in summer, and the sales of shorts and shorts increase synchronously, while the southern hemisphere is in winter, and the sales of these commodities are flat and scattered. Therefore, by using the geographically weighted regression method to quantify the spatial correlation of commodity sales in the time cycle, the time dimension and the implicit association between geographical space are explicitly modeled (the first feature), avoiding prediction bias caused by ignoring regional linkage.

[0083] Step S13: Perform time series recursive convolution on the data sequence to obtain a short-term inventory feature as the second feature; The convolution kernel is used to capture the local short-term pattern of the data sequence (such as impulse sales surge caused by promotion, pre-holiday stockpiling peak), and the recursive structure retains the historical state to perceive the continuous trend of inventory consumption.

[0084] Step S14: Perform attention encoding to obtain a third feature; Automatically focus on key time nodes (such as before and after “Double 11”, product life cycle inflection point), generate context-aware features (third feature), and solve the problem of noise interference and information dilution in long sequences.

[0085] Step S15: Use a fully connected neural network and a Softmax function to learn the attention weights of the first feature, the second feature, and the third feature;

[0086] Step S16: Use the attention weights to weight and sum the first feature, the second feature, and the third feature respectively to obtain the sales feature.

[0087] The sales feature obtained by weighting can effectively integrate time and space, seasonal promotion, and other factors, effectively and accurately feedback the sales rules of the commodity, and provide a solid foundation for subsequent storage sharding.

[0088] Preferably, the geographically weighted regression model in step S12 is:

[0089]

[0090] wherein β0is a constant term, x i,k denotes the historical data at time k of the i-th period, ε i is the error term of the i-th period, β k is the regression coefficient at time k.

[0091] Preferably, the specific steps in step S14 are as follows:

[0092] Step S141: taking the data at the next time point in the time period as new input data, and calculating the query vector, key vector and value vector from the new input data respectively;

[0093] wherein the formulas for obtaining the query vector, key vector and value vector are as follows:

[0094] q t = W Q x t , k t = W K x t , v t = W V x t ;

[0095] W Q , W K , W V are the projection matrices of the query vector, key vector and value vector respectively, x t is the new input data;

[0096] Step S142: inserting the new input data into the data sequence, and determining whether the length of the data sequence is greater than the length threshold value; if yes, obtaining the difference between the length of the data sequence and the length threshold value as the length difference value, and deleting the earliest historical sales data of the data sequence, the deletion length being the length difference value, to obtain a new data sequence;

[0097] if no, continuously supplementing the new input data until the length of the data sequence is equal to the length threshold value, to obtain a new data sequence;

[0098] Step S143: obtaining the historical key vector set from the new data sequence, obtaining the current time query vector, and using the attention score of the current time query vector and each key vector in the new data sequence;

[0099] wherein the formula for obtaining the attention score is as follows:

[0100] wherein q now is the current query vector, k i denotes the i-th key vector, denotes a scaling factor to prevent the dot product from being too large;

[0101] Step S144: normalizing the attention score to obtain a first weight, and summing the value vectors according to the first weight to obtain the third feature;

[0102] The formula for obtaining the weight by normalization is as follows:

[0103] where K is the total length of the new data sequence, a j The attention score of the jth time query vector in the new data sequence and each key vector in the new data sequence;

[0104] The formula for obtaining the third feature is as follows:

[0105]

[0106] Preferably, the step S2 of classifying the goods based on the sales features is as follows:

[0107] The sales features are input into a prediction large model to obtain the sales quantity at the future time;

[0108] The sales category of the goods is determined based on the range in which the sales quantity at the future time is located;

[0109] The prediction large model is specifically as follows:

[0110] Input layer: sales information is input into an embedding layer to be converted into a high-dimensional vector, and the formula for obtaining the high-dimensional vector is specifically as follows:

[0111] zi=Embedding(w i );

[0112] where Embedding() is an embedding layer conversion function, w i is the ith sales feature;

[0113] Transformer layer: the long and short term dependencies of the high-dimensional vector are captured through a Transformer architecture to update the model state, and the formula for updating the model state is specifically as follows:

[0114] h t =Transformer(h t-1 ,z i );

[0115] h t-1 denotes the total hidden state at t-1 time, and z i denotes the ith high-dimensional vector;

[0116] Output layer: using autoregressive generation method, according to historical sales behavior and current input high-dimensional vector, predicting the sales quantity at future time, the formula for predicting the sales quantity at future time is as follows:

[0117]

[0118] Where y t-1，i is the actual sales value of the i-th commodity in the corresponding time of the previous year, z i represents the i-th high-dimensional vector, theta is the model parameter, and f() is the prediction model function.

[0119] Preferably, the storage shards include low-load shards, medium-load shards, and high-load shards.

[0120] The QPS of the load shard is 500 times / sec.

[0121] The QPS of the medium-load shard is 2000 times / sec.

[0122] The QPS of the high-load shard is 5000 times / sec.

[0123] Of course, users can choose to use different specifications of shards according to their own sales needs.

[0124] Preferably, the updating of the inventory database in step S4 is as follows:

[0125] The query amount of the medium-load shard and the high-load shard is monitored respectively.

[0126] If the query amount of the medium-load shard reaches or is greater than 80% QPS, the storage shard of the commodity is modified to a high-load shard.

[0127] If the query amount of the high-load shard reaches or is greater than 80% QPS, the commodity is split according to the SKU, each SKU of the commodity is allocated a high-load sub-shard, and after a time threshold, it is determined whether the query amount of each high-load sub-shard is greater than a query threshold, if yes, the high-load sub-shard is retained, and if no, the corresponding SKU is re-allocated to the same high-load shard.

[0128] The QPS of the high-load sub-shard is consistent with the QPS of the high-load shard.

[0129] Since the storage shards allocated to the goods in the prior art are allocated by the prediction results, when the operation is performed, other factors can affect the current actual sales of the goods and the predicted sales, and the difference between the actual sales and the predicted sales can also cause different inventory changes, affecting the read-write speed and the like. Therefore, in the present application, the medium load shard and the high load shard are also monitored. When the query amount of the goods page reaches or is greater than 80% QPS, it means that the goods are browsed by many people. For the goods of the medium load shard, a high load shard can be allocated to the goods. For the high load shard, since the QPS of the shard cannot be further improved, the goods can be further divided according to the SKU of the goods. In some goods, the browsing and ordering conditions of the unpopular SKU and the popular SKU are quite different. Therefore, when the query amount of the goods of a high load shard reaches or is greater than 80% QPS, the operation of re-sharding needs to be performed. At this time, since it is not clear which SKU of the goods is popular, each SKU needs to be allocated to a high load sub-shard. After the time threshold, it can be known which SKU is popular according to the query amount of each high load sub-shard. Therefore, the popular SKU is retained in the high load sub-shard to ensure that the high read-write can continue, and the SKU less than the query threshold is re-allocated to the same high load shard to solve the cost problem.

[0130] A goods management system based on an intelligent warehouse uses the goods management method based on the intelligent warehouse, which comprises a feature acquisition module, a classification module, a storage allocation module, and a monitoring module.

[0131] The feature acquisition module is configured to acquire historical sales data and extract sales features of different goods based on the historical sales data.

[0132] The classification module is configured to classify the goods based on the sales features and obtain sales categories of the goods.

[0133] The storage allocation module is configured to allocate the goods information to different data shards according to the sales categories and a preset storage shard rule.

[0134] The monitoring module is configured to subscribe to and transmit the messages of the storage shards and update the inventory database.

[0135] The inventory database is used to update the preset historical inventory monitoring data to obtain real-time inventory monitoring data.

[0136] In the description of the specification, reference to "one embodiment", "some embodiments", "an exemplary embodiment", "an example", "a specific example", or "some examples" means that a particular feature, structure, material, or characteristic being described is included in at least one embodiment or example of the application. The appearances of the phrases "in one embodiment", "in some embodiments", "in an exemplary embodiment", "an example", "a specific example", or "some examples" in various places in the specification are not necessarily referring to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0137] Although embodiments of the application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and alterations can be made hereto without departing from the principles and the scope of the application, which is defined by the claims and their equivalents.

Claims

1. A method for managing goods based on an intelligent warehouse, characterized in that, Includes the following steps: Step S1: Obtain historical sales data and extract the sales characteristics of different products based on the historical sales data; Step S2: Classify the products based on sales characteristics to obtain the sales categories of the products; Step S3: Based on the sales categories and preset storage sharding rules, the inventory database is sharded, and different product information is allocated to different data shards according to the sales categories; Step S4: Subscribe to and transmit messages for the storage shards, and update the inventory database; The real-time inventory monitoring data is obtained by updating the preset historical inventory monitoring data using the inventory database.

2. The cargo management method based on a smart warehouse according to claim 1, characterized in that, The steps for extracting sales features in step S1 are as follows: Step S11: Obtain the timestamps of the historical sales data of the product within a time period, and sort the historical sales data according to the timestamps to obtain the data sequence; Step S12: Use a geographically weighted regression model to output the spatial correlation of the data sequence over the periodic time as the first feature; Step S13: Perform temporal recursive convolution on the data sequence to obtain short-term inventory features, which are used as the second feature; Step S14: Perform attention encoding to obtain the third feature; Step S15: Use a fully connected neural network and a softmax function to learn the attention weights for the first, second, and third features; Step S16: Use attention weights to perform weighted summation on the first feature, the second feature, and the third feature respectively to obtain the sales feature.

3. The cargo management method based on a smart warehouse according to claim 2, characterized in that, The geographically weighted regression model in step S12 is as follows: Where β0 is a constant term, x i,k ε represents the historical data at time k in the i-th period. i Let β be the error term for the i-th period. k Let be the regression coefficient at time k.

4. A method for managing goods based on an intelligent warehouse according to claim 2, characterized in that, The specific steps in step S14 are as follows: Step S141: Using some data from the next time point of this time period as new input data, calculate the query vector, key vector, and value vector respectively using the new input data; The formulas for obtaining the query vector, key vector, and value vector are as follows: q t =W Q x t 、k t =W K x t 、v t =W V x t ; W Q W K W V The projection matrices of the query vector key vector and value vector, respectively, and x t For new input data; Step S142: Insert the new input data into the data sequence, determine whether the length of the data sequence is greater than the length threshold, if it is greater, obtain the difference between the length of the data sequence and the length threshold as the length difference, and delete the earliest historical sales data in the data sequence. The length of the deleted data is the length difference, and a new data sequence is obtained. If the length is less than the threshold, continue to add new input data until the length of the data sequence equals the length threshold, thus obtaining a new data sequence. Step S143: Obtain the set of historical key vectors from the new data sequence, obtain the current time query vector, and use the current time query vector with the attention score of each key vector in the new data sequence; The formula for obtaining the attention score is as follows: Where q now k is the current query vector. i This represents the i-th key vector. This represents the scaling factor to prevent the dot product from becoming too large. Step S144: Normalize the attention score to obtain the first weight, and sum the value vector according to the first weight to obtain the third feature; The formula for obtaining the weights through normalization is as follows: Where K is the total length of the new data sequence, a j Attention score between the query vector at time j in the new data sequence and each key vector in the new data sequence; The formula for obtaining the third feature is as follows:

5. A method for managing goods based on an intelligent warehouse according to claim 1, characterized in that, The steps for classifying goods based on sales characteristics in step S2 are as follows: The sales characteristics are input into the large prediction model to obtain the sales volume at future time points; The product categories are determined based on the range of sales volume at future times; The specific prediction model is as follows: Input layer: Sales information is input into the embedding layer and transformed into a high-dimensional vector. The specific formula for obtaining the high-dimensional vector is as follows: zi=Embedding(w i ); Where Embedding() is the embedding layer transformation function, w i The i-th sales feature; Transformer layer: The Transformer architecture captures the long-short-term dependencies of high-dimensional vectors and updates the model state. The specific formula for updating the model state is as follows; h t =Transformer(h t-1 ,z i ); h t-1 Let z represent the total hidden state at time t-1. i Represents the i-th high-dimensional vector; Output layer: An autoregressive generation method is used to predict the sales quantity at future moments based on historical sales behavior and the current high-dimensional input vector. The formula for predicting the sales quantity at future moments is as follows; Where y t-1，i z represents the actual sales value of the i-th product within the corresponding time period of the previous year. i Let represent the i-th high-dimensional vector, θ be the model parameters, and f() be the prediction model function.

6. A method for managing goods based on an intelligent warehouse according to claim 1, characterized in that, The storage sharding includes low-load sharding, medium-load sharding, and high-load sharding; The QPS of the load sharding is 500 times / second; The QPS of the load sharding is 2000 times / second; The QPS of the high-load sharding is 5000 times / second.

7. A method for managing goods based on an intelligent warehouse according to claim 6, characterized in that, The update of the inventory database in step S4 is as follows: Monitor the query volume for medium-load shards and high-load shards respectively; If the query volume of a medium-load shard reaches or exceeds 80% of the QPS, then the storage shard for that product will be changed to a high-load shard. If the query volume of a high-load shard reaches or exceeds 80% of the QPS, the product is split according to its SKU. Each SKU is assigned to a high-load sub-shard. After a time threshold, the query volume of each high-load sub-shard is checked to see if it exceeds the query threshold. If it does, the high-load sub-shard is retained; if it does, the corresponding SKU is reassigned to the same high-load shard.

8. A cargo management system based on an intelligent warehouse, characterized in that, The intelligent warehouse-based goods management method according to any one of claims 1 to 7 includes a feature acquisition module, a classification module, a storage allocation module, and a monitoring module; The feature acquisition module is used to acquire historical sales data and extract sales features of different products based on the historical sales data; The classification module is used to classify goods based on sales characteristics and obtain the sales categories of the goods; The storage allocation module is used to perform storage sharding on the inventory database according to the sales type and the preset storage sharding rules. Different product information is allocated to different data shards for storage according to the sales type. The monitoring module is used to subscribe to and transmit messages from the storage shards and update the inventory database. The real-time inventory monitoring data is obtained by updating the preset historical inventory monitoring data using the inventory database.

Images