Warehousing intelligent inventory prediction and replenishment method and system based on artificial intelligence

Abstract

The invention discloses a storage intelligent stock prediction and replenishment method and system based on artificial intelligence, and the method comprises the steps: firstly obtaining a stock cargo label set, extracting a stock cargo feature set and a storage operation feature set, integrating the stock cargo feature set and the storage operation feature set to obtain an integrated feature set, and then extracting a periodic correlation feature set; and extracting corresponding link feature enhancement information from preset link feature enhancement information, and refining the periodic association feature set to obtain a target feature set. And the target feature set is used to predict the stockout risk index of each to-be-replenished goods label, and a replenishment instruction is generated according to the stockout risk index, so that intelligent warehouse stock prediction and accurate replenishment are realized, and the warehouse management efficiency and benefits are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based warehouse intelligent inventory forecasting and replenishment method and system. Background Art

[0002] Traditional warehouse inventory management relies heavily on manual experience or simple statistical analysis, which can't accurately predict inventory demand or effectively arrange replenishment. This can easily lead to inventory backlogs or stockouts, increasing warehousing costs and operational risks. As warehouses expand and operations become more complex, intelligent methods are urgently needed to achieve more efficient inventory forecasting and replenishment. Summary of the Invention

[0003] The purpose of the present invention is to provide an artificial intelligence-based warehouse intelligent inventory forecasting and replenishment method and system.

[0004] In a first aspect, an embodiment of the present invention provides an artificial intelligence-based intelligent warehouse inventory forecasting and replenishment method, comprising:

[0005] Acquire an inventory goods tag set, wherein the inventory goods tag set includes each inventory goods tag of the warehouse system;

[0006] Obtaining an inventory feature set and a warehousing operation feature set based on the inventory tag set, wherein the inventory feature set is obtained by extracting feature information of the goods attributes of each inventory tag, and the warehousing operation feature set is obtained by extracting feature information of the warehousing operation of each inventory tag, wherein the warehousing operation information is determined from the warehousing link to which the inventory tag belongs;

[0007] Integrating the inventory feature set and the warehousing operation feature set to obtain an integrated feature set, and extracting periodic correlation information of integrated features in the integrated feature set to obtain a periodic correlation feature set;

[0008] Extracting multiple link feature enhancement information corresponding to the periodic associated feature set from each preset link feature enhancement information, and refining the periodic associated feature set according to the multiple link feature enhancement information to obtain a target feature set;

[0009] The target feature set is used to predict and obtain the out-of-stock risk index corresponding to each of the to-be-replenished goods labels, and to generate a replenishment instruction based on the out-of-stock risk index corresponding to each of the to-be-replenished goods labels.

[0010] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is configured to execute the method described in the first aspect.

[0011] Compared to existing technologies, the present invention provides the following beneficial effects: Using the artificial intelligence-based intelligent warehouse inventory forecasting and replenishment method and system disclosed in the present invention, a set of inventory goods labels is obtained, and based on this, an inventory goods feature set and a warehouse operation feature set are extracted, and the two are integrated to obtain an integrated feature set, and then a periodic correlation feature set is extracted. Corresponding link feature enhancement information is extracted from preset link feature enhancement information, and the periodic correlation feature set is refined to obtain a target feature set. The target feature set is used to predict the out-of-stock risk index of each to-be-replenished goods label, and replenishment instructions are generated accordingly, achieving intelligent warehouse inventory forecasting and precise replenishment, thereby improving warehouse management efficiency and benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly describes the drawings required for use in the embodiments. It should be understood that the following drawings illustrate only certain embodiments of the present invention and should not be construed as limiting the scope of the present invention. Those skilled in the art can, without inventive effort, derive other relevant drawings from these drawings.

[0013] Figure 1 A schematic diagram of the steps of the artificial intelligence-based warehouse intelligent inventory forecasting and replenishment method provided in an embodiment of the present invention;

[0014] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the described embodiments are only a portion of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0016] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0017] In order to solve the technical problems in the above background technology, Figure 1 This is a flow chart of an artificial intelligence-based warehouse intelligent inventory forecasting and replenishment method provided in an embodiment of the present disclosure. The artificial intelligence-based warehouse intelligent inventory forecasting and replenishment method is introduced in detail below.

[0018] Step S201: Acquire an inventory goods tag set, wherein the inventory goods tag set includes each inventory goods tag in the warehouse system;

[0019] Step S202: Acquire an inventory feature set and a warehousing operation feature set based on the inventory tag set. The inventory feature set is obtained by extracting feature information of the goods attributes of each inventory tag. The warehousing operation feature set is obtained by extracting feature information of the warehousing operation of each inventory tag. The warehousing operation information is determined based on the warehousing link to which the inventory tag belongs.

[0020] Step S203: Integrate the inventory feature set and the warehousing operation feature set to obtain an integrated feature set, and extract periodic correlation information of the integrated features in the integrated feature set to obtain a periodic correlation feature set;

[0021] Step S204: extract multiple link feature enhancement information corresponding to the periodic associated feature set from each preset link feature enhancement information, and refine the periodic associated feature set based on the multiple link feature enhancement information to obtain a target feature set; the target feature set is used to predict the out-of-stock risk index corresponding to each to-be-replenished goods label, and replenishment instructions are generated based on the out-of-stock risk index corresponding to each to-be-replenished goods label.

[0022] In an embodiment of the present invention, illustratively, the server first needs to clarify all the cargo objects that need to be monitored and managed in the current warehousing system. The inventory cargo label is the unique identifier of the goods in the warehousing system, usually including the SKU (Stock Keeping Unit) number, barcode, RFID tag code, etc. For example, the e-commerce warehousing center manages 8,000 SKUs in 10 major categories such as household daily necessities, food and beverages, and 3C electronic products. Each SKU corresponds to a unique "inventory cargo label" (such as "SKU-20231015-001" represents the first category of daily necessities put into storage on October 15, 2023). The server interacts in real time with the interface of the warehouse management system (WMS) and automatically triggers the data pulling process at 0:00 am every day. Specifically, the server sends a query request to the WMS to obtain the label information of all current goods in stock, including goods that have not been fully shipped out, goods to be processed in the temporary storage area, and goods being sorted. For example, the inventory goods label set obtained in the early morning of a certain weekday may contain 7,500 valid labels (because some goods have completed the outbound process and have been removed). These labels constitute the basic data for subsequent analysis.

[0023] The server needs to extract two key features from inventory tags: the attributes of the goods themselves (the inventory feature set) and the behavioral characteristics of the warehousing operation process (the warehousing operation feature set). Each inventory tag corresponds to unique item attribute information, which is stored in the company's product database. Based on the SKU number on the inventory tag, the server retrieves the static attributes (such as category, volume, weight, shelf life, and procurement cost) and dynamic attributes (such as the average daily sales volume over the past 30 days, the sales volatility over the past 90 days, and the sales growth coefficient during historical promotional periods) of the goods from the product database. For example, for a 500mL bottle of household laundry detergent labeled "SKU-20231015-001," its static attributes include: Category = Daily Necessities, Volume = 0.5L, Weight = 0.6kg, Shelf Life = 24 Months, and Purchase Cost = 8 RMB / bottle. Dynamic attributes include: Average Daily Sales Volume over the Past 30 Days = 200 bottles, Sales Volatility over the Past 90 Days = 15% (i.e., sales fluctuations within ±15%), and Sales Growth Factor during the Historical Double 11 (Singles' Day) Promotional Period = 3 (i.e., sales during promotional periods are three times higher than normal). The server structures these attributes through feature engineering, such as one-hot encoding the category (daily necessities = 1000, food = 0100, etc.), normalizing the volume and weight to values ​​in the [0, 1] range, and ultimately forming the "inventory feature vector" for the product (e.g., [1, 0, 0, 0, 0.5, 0.6, 24, 8, 0.2, 0.15, 3]). After integrating the feature vectors corresponding to all inventory goods labels, we get the "inventory goods feature set". Warehousing operation information comes from the operation records of goods in the entire warehousing process, including warehousing, sorting, outbound, inventory and other links. The server extracts the operation data of each link for each inventory goods label from the WMS operation log. For example, the operation information of the laundry detergent with label "SKU-20231015-001" in the warehousing link includes: warehousing time = 2023-10-15-08:30, warehousing quantity = 1000 bottles, supplier arrival punctuality rate (nearly 3 months) = 98%, quality inspection pass rate = 100%; the operation information in the sorting link includes: sorting start time = 2023-10-15-09:00, sorting completion time = 2023- At 9:30 AM on October 15th, the sorting damage rate was 0%, and the sorting efficiency was 200 bottles / hour. Operational information for the outbound phase included: the average daily outbound orders for the past 30 days was 50, the on-time outbound rate was 95%, and the order fulfillment rate was 99% (i.e., the degree of matching between order requirements and actual outbound quantities). Operational information for the inventory phase included: the discrepancy rates (the difference between actual inventory and system inventory) for the past three counts were 0.2%, 0.1%, and 0%, respectively. The server needs to convert this multi-phase, multi-time point operational information into a structured feature vector.Specifically, for each inventory cargo tag, the server extracts time series features (such as the interval and frequency of operation time) and calculates statistical features (such as average value, standard deviation, and maximum value) for the operation data of each link. For example, the daily data of the "on-time arrival rate" of the warehousing link for the past three months is converted into the "weekly average on-time rate" and "on-time rate volatility", and the "sorting efficiency" of the sorting link is statistically analyzed hourly as "peak period efficiency" and "off-peak period efficiency". Ultimately, the warehousing operation information of each cargo tag is converted into a multi-dimensional "operation time series vector" (such as [0.98, 0, 50, 0.95, 0.99, 0.2, 200, 150]). After the operation time series vectors of all tags are integrated, the "warehousing operation feature set" is obtained.

[0024] The server needs to integrate inventory characteristics with warehouse operation characteristics to identify cyclical patterns (such as daily, weekly, and monthly sales cycles, promotional activity cycles, etc.), providing temporal correlation information for subsequent forecasting. The server concatenates or weightedly fuses the "inventory feature vector" corresponding to each inventory item tag with the "operation time series vector" to form an "integrated feature vector." For example, the inventory feature vector for laundry detergent is 11-dimensional, and the operation time series vector is 8-dimensional, resulting in a 19-dimensional integrated feature vector. The integrated feature vectors of all tags form an "integrated feature set," which combines product attributes, operation behavior, and time series information. The server uses time series analysis and machine learning models to identify cyclical patterns in the integrated features. The specific steps are as follows: Cycle Feature Mapping: The server performs Fourier transform or wavelet analysis on the time series data in the integrated features (such as average daily sales volume and number of outbound orders) to identify the main cyclical components. Fourier transform is preferred for stationary time series (such as long-term sales data); wavelet analysis is preferred for non-stationary time series (such as short-term data affected by promotional activities) to capture local cyclical characteristics. For example, average daily laundry detergent sales data shows that sales on Saturdays and Sundays are 1.5 times higher than on weekdays. Fourier transform results indicate that the 7-day cycle accounts for the highest energy share (85%), thus determining the primary cycle to be 7 days. Furthermore, sales begin to rise in the two weeks before major promotions like Double 11, forming a secondary cycle of approximately 30 days. The server converts this cycle information into a "target cycle correlation factor" (e.g., 7-day cycle factor = 0.85, 30-day cycle factor = 0.15), a "cycle feature index" (e.g., cycle length = 7 days, 30 days), and a "cycle feature base volume" (e.g., baseline value for average daily sales within the cycle = 200 bottles, baseline increment in the two weeks before the promotion = 100 bottles). Feature reorganization: The server calculates the correlation between the target cycle correlation factor and the cycle feature index (e.g., correlation coefficient between the 7-day cycle factor and the 7-day index = 0.92), and weights the cycle feature base volume based on the correlation. For example, the highly correlated 7-day cycle base volume is retained and amplified (weight 0.9), and the low-correlated 30-day cycle base volume is adjusted (weight 0.1), ultimately generating a "periodic correlation feature set" that reflects the core cycle laws (such as sales fluctuation patterns within a 7-day cycle, weekly changes in sorting efficiency, etc.).

[0025] The server needs to combine pre-configured feature enhancement information for each warehousing process (e.g., focusing on supplier stability in the incoming warehousing process and order fulfillment rate in the outgoing warehousing process) to filter and optimize periodicity-related features, remove redundant information, and retain key features highly correlated with stock-out risk. The server first assesses the importance of periodicity-related features, for example, using SHAP (SHapley Additive expressions) to analyze the impact of each feature on stock-out risk. Assuming the pre-configured warehousing process includes incoming warehousing, sorting, outgoing warehousing, and inventory counting, each process corresponds to a set of "process feature enhancement information" (e.g., enhanced information for incoming warehousing includes supplier on-time arrival rate and quality inspection pass rate; for outgoing warehousing, order fulfillment rate and outgoing on-time rate). SHAP analysis reveals that the outgoing warehousing process's "order fulfillment rate" has the highest weight on stock-out risk (0.4), followed by the incoming warehousing process's "supplier on-time arrival rate" (0.3). The sorting process's "sorting efficiency" (0.2) and the inventory counting process's "inventory discrepancy rate" (0.1) have relatively small impacts. Based on these weight coefficients, the server selects links (such as outbound and inbound) that are highly correlated with the cyclical correlation feature, retaining the corresponding enhanced information. The server uses this enhanced information to refine the cyclical correlation feature. For example, for the outbound link's "order fulfillment rate," the server extracts historical data (such as the daily fulfillment rate over the past three months) and cross-validates it against the "7-day sales cycle" in the cyclical correlation feature. If the fulfillment rate falls below 90% in a given week, the actual outbound volume may be lower than the system-recorded volume, resulting in inflated inventory and increased stock-out risk. Through this cross-analysis, the server adjusts the "sales forecast" in the cyclical correlation feature (for example, adjusting the original forecast of 200 bottles / day to 180 bottles / day to reflect the impact of insufficient fulfillment). Simultaneously, the server dynamically adjusts the weight coefficients of each link: the weights of the top-ranked outbound and inbound links remain unchanged (0.4, 0.3), while the weights of the lower-ranked sorting and inventory counting links are adjusted to the preset cyclical feature base value (such as 0.05). All weights are then normalized (summing to 1). Finally, the server assigns a target link weight coefficient to each link feature enhancement information (such as outbound = 0.45, inbound = 0.35, sorting = 0.1, inventory = 0.1). The server uses the corresponding target link weight to weight the refined features for each link feature enhancement information, such as multiplying the features of the outbound link by 0.45 and the features of the inbound link by 0.35. Finally, the weighted feature vectors of each link are integrated to form a "target feature set". This set focuses on the core features that have the greatest impact on out-of-stock risks (such as the correlation between outbound fulfillment rate and sales cycle, and the correlation between supplier on-time arrival rate and inventory replenishment speed), providing high-value input for subsequent predictions.

[0026] The server uses the target feature set as input and uses a trained warehouse intelligent decision-making model to predict the out-of-stock risk index of each to-be-replenished goods label, and generates replenishment instructions based on the risk level. The warehouse intelligent decision-making model is a machine learning model (such as XGBoost or a neural network) trained by the server based on historical data. Its input is the target feature set, and its output is the out-of-stock risk index (0-100 points, the higher the score, the greater the risk) of each to-be-replenished goods label (i.e., the candidate replenishment SKU). For example, the server inputs the target features of laundry detergent (such as an adjusted forecast value of 180 bottles / day within a 7-day sales cycle, a 95% outbound fulfillment rate, and a 98% supplier on-time arrival rate) into the model, and the model outputs its out-of-stock risk index for the next 7 days as 85 points (high risk). The server sets a risk threshold (such as 70 points) and filters out to-be-replenished goods labels whose risk index exceeds the threshold (such as 85 points for laundry detergent). For high-risk goods, the server further calculates the replenishment quantity: Based on the target feature's "7-day sales cycle forecast" (180 bottles / day), current inventory (500 bottles remaining), and the supplier's replenishment cycle (3 days), the server calculates the safety stock (180 × 3 = 540 bottles). Therefore, the required replenishment quantity = safety stock - current inventory = 540 - 500 = 40 bottles. Finally, a replenishment instruction is generated, including the product label "SKU-20231015-001", the replenishment quantity 40 bottles, supplier A, and the required arrival time before 2023-10-20 12:00. To ensure prediction accuracy, the server regularly verifies the match between the prediction results and the actual out-of-stock situation. For example, if the predicted out-of-stock risk index of a certain SKU is 85 points, but it is not actually out of stock (the risk index deviation exceeds 10%), the server will trigger the model iteration process: the target feature set of this time will be used as the integrated feature set again, and steps 3 (extracting periodic correlation features) and 4 (feature refinement) will be repeated, and the model parameters will be adjusted (such as increasing the weight of the inventory link) until the prediction accuracy meets the standard (such as the deviation is less than 5%).

[0027] Through the above steps, the server implements intelligent inventory management throughout the entire process, from cargo label acquisition, feature extraction, cycle correlation analysis, feature refinement, to risk prediction. Taking e-commerce warehousing centers as an example, this method can accurately identify goods at high risk of out-of-stock situations and dynamically generate replenishment instructions, effectively reducing inventory backlogs and out-of-stock rates, and improving warehouse operational efficiency. In practical applications, the server can adjust feature extraction strategies and model parameters based on different warehousing scenarios (such as the strict shelf life requirements of fresh food warehouses and the high sorting accuracy requirements of 3C warehouses), ensuring the universality and targeted nature of the method.

[0028] In the embodiment of the present invention, the acquisition of the inventory feature set and the warehousing operation feature set according to the inventory tag set can be implemented through the following examples.

[0029] According to each inventory tag in the inventory tag set, searching for the inventory feature corresponding to each inventory tag in each predetermined inventory feature to obtain the inventory feature set;

[0030] Acquire multiple storage operation information of each inventory cargo tag from multiple storage links to which each inventory cargo tag belongs;

[0031] For each inventory goods tag, multiple operation feature vectors of warehousing operation information are extracted to obtain multiple operation time sequence vectors, and the multiple operation time sequence vectors are integrated to obtain the warehousing operation feature set.

[0032] In an embodiment of the present invention, for example, taking a regional cold chain storage center of a fresh food e-commerce company as an example, the server, as the data processing core, needs to perform feature extraction operations on 2,000 SKUs (such as fresh vegetables, frozen meat, dairy products, etc.) in stock every day. The specific process is as follows: the server first retrieves the specific feature value corresponding to each inventory goods label from the product master database based on the pre-defined "inventory goods feature" field (such as static attributes: category, net content, shelf life, storage temperature; dynamic attributes: sales in the past 7 days, sales volatility in the past 30 days, sales growth in historical promotion periods). For example, for boxed fresh milk (250 mL / box) labeled "SKU-20240301-005", the server queries the product library through the SKU number "20240301-005" in the label and obtains its static attributes: category = dairy products, net content = 250 mL, shelf life = 7 days (calculated from the production date), storage temperature = 2-6°C; dynamic attributes: sales volume in the past 7 days = 800 boxes (an average of about 114 boxes per day), sales volatility in the past 30 days = 22% (fluctuations caused by increased household purchases on weekends), historical "618" promotion period sales growth rate = 150% (that is, the average daily sales during the promotion period is 2.5 times that of normal days). The server structures these attributes: The category is converted to [0, 1, 0, 0] using one-hot encoding (assuming dairy products are Category 2), the net content is normalized to 0.25 (based on 1L), the shelf life is converted to 7 days, and the storage temperature is normalized to 0.4 (within the range of 0-10°C). For dynamic attributes, the daily average sales volume over the past seven days is set to 114 boxes, the volatility is retained at 22%, and the promotional increase is retained at 150%. The resulting inventory feature vector for this tag is [0, 1, 0, 0, 0.25, 7, 0.4, 114, 0.22, 1.5]. The server iterates over all 2,000 inventory tags (e.g., fresh shiitake mushrooms "SKU-20240301-012" and frozen chicken breast "SKU-20240301-023") and stores the feature vector corresponding to each tag in the database, forming the "inventory feature set." The server extracts operational information from the warehouse management system (WMS) operation log for each inventory tag during the four core steps of warehousing, sorting, shipping, and inventory. Taking "SKU-20240301-005" fresh milk as an example: For the warehousing phase, operational information includes: warehousing time = 2024-03-01-06:00 (supplier delivers at 6:00 am daily), quantity received = 1,000 boxes, supplier's on-time delivery rate over the past 30 days = 92% (delayed twice due to traffic problems), quality inspection pass rate = 100% (no spoilage or packaging damage). For the sorting phase, operational information includes: sorting start time = 2024-03-01-06:30, sorting completion time = 2024-03-01-07:00 (taking 30 minutes), sorting efficiency = 200 boxes / 10 minutes (peak period), and sorting damage rate = 0% (no squeeze leakage).Outbound delivery: Operational information includes the average daily outbound delivery orders over the past seven days = 45 orders (corresponding to 45 online orders), outbound delivery punctuality rate = 90% (due to the concentration of orders at 8 a.m., there were three delays of 10 minutes), and order fulfillment rate = 98% (2% of orders were not fully met due to temporary shortages). Inventory: Operational information includes the discrepancy rate (deviation between system inventory and actual inventory) of the past three inventory counts: 0.5% (two boxes missing during the inventory count on March 1), 0% (February 25), and 0.3% (February 20). The server pulls the raw data of these links in real time through the interface, stores them by label, and forms a "multi-link operation information package" for each label. The server needs to convert the multi-link operation information of each label into a computable feature vector. The specific operations are as follows: (1) Operational feature vector extraction: for the warehousing link of "SKU-20240301-005", the server converts "arrival punctuality rate 92%" into a value of 0.92, and "warehousing quantity 1000 boxes" is normalized to 0.5 (with 2000 boxes as the maximum warehousing quantity); the sorting link converts "sorting time 30 minutes" into 0.5 (based on 60 minutes), and "sorting efficiency 200 boxes / 10 minutes" into 20 (boxes / minute); the outbound link converts "outbound punctuality rate 90%" into 0.9, and "order fulfillment rate 98%" into 0.98; the inventory link converts "the average of the difference rate of the last three times is 0.27%" into 0.0027. (2) Operation time series vector generation: The server calculates the intervals of the time series data of each link (such as storage time, sorting time, and outbound time). For example, the time interval from storage to sorting is 30 minutes (06:00 to 06:30), and the time interval from sorting to outbound is 7 hours and 30 minutes (07:00 to 14:30), forming the time series feature [0.5 hours, 7.5 hours]. Combined with the statistical features of each link (such as mean and volatility), the "operation time series vector" of this label is generated [0.92, 0.5, 0.5, 20, 0.9, 0.98, 0.0027, 0.5, 7.5]. (3) Integration to form a warehouse operation feature set: The server stores the operation time series vectors of 2,000 inventory goods labels into the feature library one by one. For example, the operation time series vector of fresh mushrooms "SKU-20240301-012" contains its unique sorting damage rate (0.1%), the regional distribution of outbound orders (such as 70% of orders from Area A), and other features. Finally, all vectors are integrated into the "warehouse operation feature set". Through the above process, the server completes the transformation from raw label data to a structured feature set, providing key input for subsequent cycle correlation analysis and replenishment prediction. For example, the inventory characteristics of fresh milk (short shelf life, large sales fluctuations) and warehouse operation characteristics (high sorting efficiency but low outbound on-time rate) will be combined for analysis to identify its potential out-of-stock risk points (such as excessive actual inventory consumption due to outbound delays).

[0033] In the embodiment of the present invention, the integration of the inventory feature set and the warehousing operation feature set to obtain an integrated feature set may be implemented through the following examples.

[0034] The inventory features in the inventory feature set are integrated with the corresponding warehousing operation features in the warehousing operation feature set to obtain the integrated feature set.

[0035] In an embodiment of the present invention, for example, taking the server processing of the cold chain warehousing center of a fresh food e-commerce company as an example, for boxed fresh milk (250mL / box) labeled "SKU-20240301-005", the server needs to integrate its inventory goods features with the warehousing operation features to form a complete feature representation that includes goods attributes and operation behaviors. The specific process is as follows: the server first uses the "inventory goods label" as a unique identifier to correspond one-to-one with the feature vectors in the inventory goods feature set and the warehousing operation feature set. For example, the inventory goods feature set stores a 10-dimensional feature vector of "SKU-20240301-005" (such as [category code, net content, shelf life, storage temperature, sales volume in the past 7 days, sales volatility, promotion growth rate]), and the warehousing operation feature set stores a 9-dimensional operation time series vector of the label (such as [arrival punctuality rate, warehousing quantity, sorting time, sorting efficiency, outbound punctuality rate, order fulfillment rate, inventory difference rate, inbound-sorting interval, sorting-outbound interval]). The server locates these two vectors using the tag ID "SKU-20240301-005" to ensure the accuracy of subsequent integration. The server merges the two features using a "feature concatenation" method, directly concatenating the inventory feature vector with the warehousing operation feature vector in sequence to form a higher-dimensional integrated feature vector. This method has the advantage of fully preserving the original information and avoiding information loss caused by weighted fusion. Taking "SKU-20240301-005" as an example: the inventory feature vector (10 dimensions): [0 (non-vegetable), 1 (dairy), 0 (non-meat), 0 (non-fruit), 0.25 (250mL normalized value), 7 (shelf life days), 0.4 (storage temperature normalized value), 114 (average daily sales), 0.22 (sales volatility), 1.5 (promotional growth)]. Warehousing operation feature vector (9 dimensions): [0.92 (on-time arrival rate), 0.5 (normalized incoming inventory quantity), 0.5 (normalized sorting time), 20 (sorting efficiency, boxes / minute), 0.9 (on-time outbound rate), 0.98 (order fulfillment rate), 0.0027 (mean inventory discrepancy rate), 0.5 (inbound-sorting interval, hours), 7.5 (sorting-outbound interval, hours)]. The server concatenates these two vectors in sequence to obtain a 19-dimensional integrated feature vector: [0, 1, 0, 0, 0.25, 7, 0.4, 114, 0.22, 1.5, 0.92, 0.5, 0.5, 20, 0.9, 0.98, 0.0027, 0.5, 7.5]. The server traverses all inventory tags (such as the 2,000 SKUs in the warehouse center), repeats the above "match by ID + feature splicing" operation for each tag, and finally stores all integrated feature vectors as an "integrated feature set".For example, fresh shiitake mushrooms with the label "SKU-20240301-012" have inventory characteristics that include attributes such as "shelf life 3 days" and "storage temperature 8°C." Warehouse operation characteristics include information such as "sorting damage rate 0.1%" and "regional distribution of outbound orders (70% from Area A)." After integration, these characteristics also form a 19-dimensional feature vector. This integrated feature set encompasses both the product's attributes (such as a short shelf life and large sales fluctuations) and warehouse operation behaviors (such as high sorting efficiency but low on-time delivery), providing multi-dimensional information support for subsequent analysis. For example, by analyzing the integrated features of "SKU-20240301-005," the server discovered a correlation between its "storage period of 7 days" and its "sorting-to-delivery interval of 7.5 hours." A shorter shelf life requires an efficient delivery process, and while the 7.5-hour interval is within the specified timeframe, combined with the fluctuations in the "90% on-time delivery rate," this could indicate the risk of delivery delays due to concentrated orders, which could in turn affect inventory depletion. This multi-dimensional correlation analysis is the core foundation for subsequently extracting cyclical correlation features and predicting out-of-stock risks. Through this process, the server completes the transformation from "single-dimensional features" to "multi-dimensional integrated features," providing more comprehensive and accurate data input for intelligent warehouse forecasting and replenishment decisions.

[0036] In the embodiment of the present invention, the extracting of periodic correlation information of the integrated features in the integrated feature set to obtain the periodic correlation feature set may be implemented through the following examples.

[0037] Performing periodic feature mapping on the integrated feature set to obtain a target periodic correlation factor set, a periodic feature index set, and a periodic feature base quantity set;

[0038] The correlation between the target period correlation factor set and the period feature index set is calculated to obtain a correlation degree set, and the period feature base quantity set is feature reorganized according to the correlation degree set to obtain the periodic correlation feature set.

[0039] In an embodiment of the present invention, for example, taking the server processing of a cold chain warehousing center of a fresh food e-commerce company as an example, for boxed fresh milk (250mL / box) labeled "SKU-20240301-005" in the integrated feature set, the server needs to extract the periodic correlation information in its integrated features. The specific process is as follows: The server first performs a periodic analysis on the time series data in the integrated features (such as average daily sales, number of outbound orders, sorting efficiency, etc.) to identify hidden time patterns. Taking "SKU-20240301-005" as an example, its integrated features include time series fields such as average daily sales (114 boxes / day), number of outbound orders (45 orders / day), and sorting efficiency (20 boxes / minute) in the past 90 days. The server performed frequency decomposition on these data through Fourier transform and found that: in the average daily sales data, the 7-day cycle energy accounted for the highest proportion (82%), which is manifested in that weekend (Saturday and Sunday) sales are 30% higher than weekdays (for example, 150 boxes on Saturday, 160 boxes on Sunday, and 100 boxes on weekdays); in the outbound order number data, the 7-day cycle energy accounted for 75%, which is synchronized with the sales cycle; in the sorting efficiency data, the 7-day cycle energy accounted for 68% (orders are concentrated on weekends, and the sorting efficiency is increased to 25 boxes / minute); in addition, sales increased slightly in the two weeks (about 14 days) before major promotional activities such as Double 11 and 618 (an increase of 15%), forming a secondary cycle (energy accounted for 15%). The server converts these cycle information into: target cycle correlation factor set: [0.82 (7-day cycle factor), 0.15 (14-day cycle factor)] (factor value is the energy proportion of each cycle); cycle feature index set: [7 (days), 14 (days)] (cycle length); cycle feature base quantity set: [100 (weekday sales benchmark), 155 (weekend sales benchmark), 40 (weekday order benchmark), 50 (weekend order benchmark), 20 (weekday sorting efficiency benchmark), 25 (weekend sorting efficiency benchmark)] (typical value within each cycle). The server needs to verify the actual correlation between the cycle correlation factor and the cycle index, and retain the feature base quantity that is strongly correlated with the core cycle. (1) Calculate the correlation degree set: The server analyzes the correlation between the cycle factor and the index through the Pearson correlation coefficient. For example, the correlation coefficient between the 7-day cycle factor (0.82) and the 7-day index is 0.95 (high correlation), and the correlation coefficient between the 14-day cycle factor (0.15) and the 14-day index is 0.62 (moderate correlation). The server stores these coefficients in the "correlation degree set": [0.95, 0.62]. (2) Feature reorganization: The server weights and adjusts the period feature base according to the correlation degree. The period base with high correlation is strengthened, while the period base with low correlation is weakened.For example, the base quantity of a 7-day cycle (e.g., weekday sales of 100 boxes and weekend sales of 155 boxes) is multiplied by a high correlation coefficient of 0.95, and retained as [100×0.95=95, 155×0.95=147.25]; the base quantity of a 14-day cycle (e.g., a 15% increase in sales before a promotion) is multiplied by a low correlation coefficient of 0.62, and adjusted to [15%×0.62=9.3%]. (3) Generate a set of periodic correlation features: The server integrates the adjusted base quantity with the period index and factor to form a feature set that reflects the core periodic law. For example, the periodic correlation features of "SKU-20240301-005" include: 7-day main cycle: weekday sales benchmark 95 boxes / day, weekend sales benchmark 147.25 boxes / day, corresponding to a periodic factor of 0.82; 14-day secondary cycle: pre-promotion sales benchmark 9.3%, corresponding to a periodic factor of 0.15. Through the above process, the server extracts core cyclical patterns (such as weekly sales fluctuations) from fresh milk inventory and operational characteristics, while de-emphasizing secondary cycles (such as small increases before major sales promotions). These characteristics provide a key basis for the time dimension in subsequent predictions of out-of-stock risks. For example, based on the 7-day weekend sales benchmark (147.25 boxes / day), combined with current inventory (assuming 500 boxes remaining) and the replenishment cycle (3 days), the server can predict potential weekend inventory gaps (147.25 × 3 = 441 boxes, with 500 boxes remaining to support 3 days, but safety stock must be reserved), thereby generating replenishment instructions in advance. Ultimately, the server integrates the periodic correlation features of all inventory tags into a "periodic correlation feature set," laying the foundation for feature refinement and risk prediction.

[0040] In an embodiment of the present invention, the extraction of multiple link feature enhancement information corresponding to the periodic associated feature set from each preset link feature enhancement information, and feature refinement of the periodic associated feature set based on the multiple link feature enhancement information to obtain the target feature set can be implemented through the following examples.

[0041] Performing feature importance evaluation based on the periodically associated feature set to obtain a link weight coefficient corresponding to each preset link feature enhancement information;

[0042] Filtering each preset link feature enhancement information according to the link weight coefficient corresponding to each preset link feature enhancement information to obtain a plurality of link feature enhancement information corresponding to the periodically associated feature set;

[0043] Refining the periodically associated feature sets according to the plurality of link feature enhancement information to obtain a plurality of current target feature sets;

[0044] The multiple current target feature sets are integrated according to the link weight coefficients corresponding to the multiple link feature reinforcement information to obtain the target feature set.

[0045] In an embodiment of the present invention, for example, taking the server processing of the cold chain warehousing center of a fresh food e-commerce company as an example, for boxed fresh milk (250mL / box) labeled "SKU-20240301-005", the server needs to combine the preset link feature enhancement information, and the preset link feature enhancement information includes: warehousing link (supplier arrival punctuality rate, quality inspection pass rate, warehousing quantity fluctuation rate); sorting link (sorting efficiency, sorting damage rate, sorting time fluctuation rate); outbound link (order fulfillment rate, outbound punctuality rate, average daily number of outbound orders); inventory link (inventory difference rate, inventory frequency), and refine its periodic correlation features. The specific process is as follows: The server pre-defines the "link feature enhancement information" of four warehousing links: warehousing link (focusing on supplier arrival punctuality rate, quality inspection pass rate), sorting link (focusing on sorting efficiency, damage rate), outbound link (focusing on order fulfillment rate, outbound punctuality rate), and inventory link (focusing on inventory difference rate). To evaluate the importance of these links to cyclically correlated features, the server uses the SHAP value analysis tool to calculate the impact weight of each link feature on the out-of-stock risk, for example, through TreeExplainer (applicable to tree models such as XGBoost) or KernelExplainer (applicable to any model), reflecting the marginal contribution of each feature to the model output. Taking the cyclical correlation characteristics of "SKU-20240301-005" (such as a 7-day sales cycle and increased sorting efficiency on weekends) as an example, SHAP analysis results show that the "order fulfillment rate" in the outbound stage has the highest weight (0.45) on stock-out risk, as a low fulfillment rate can lead to faster actual inventory consumption than recorded in the system. The "supplier on-time delivery rate" in the inbound stage has the second highest weight (0.35), as delayed arrivals directly affect replenishment speed. The "sorting efficiency" in the sorting stage has a lower weight (0.15), as fluctuations in efficiency in this stage have a minimal impact on inventory consumption. The "inventory discrepancy rate" in the inventory stage has the lowest weight (0.05), as the discrepancy rate is long-term stable (<0.5%) and has limited impact on prediction. These weight coefficients form the "stage weight coefficient set": [0.45 (outbound), 0.35 (inbound), 0.15 (sorting), 0.05 (inventory)]. Based on these stage weight coefficients, the server selects stages with high correlation to cyclical correlation characteristics. Usually, a threshold is set (such as 0.1), and links with weights ≥ the threshold are retained. In this example, the weights of outbound (0.45), inbound (0.35), and sorting (0.15) are all ≥ 0.1 and are retained; the inventory link (0.05) is excluded because its weight is too low. After screening, the "link feature enhancement information" is locked to the features of the three links of outbound, inbound, and sorting (such as the order fulfillment rate of the outbound link, the on-time arrival rate of the inbound link, and the sorting efficiency of the sorting link). For each filtered link, the server uses its feature enhancement information to refine the periodic correlation features and remove redundant or low-correlation information.(1) Refinement of the outbound link: The periodic correlation feature includes the "7-day sales cycle" (weekend sales benchmark 147.25 boxes / day). Combined with the historical data of the "order fulfillment rate" of the outbound link (the average value for the past 30 days is 98%, but there were 3 days below 95%), the server found that when the fulfillment rate is <95%, the actual outbound volume is 5% more than the system record (because urgent orders are shipped first). Therefore, the server adjusted the "weekend sales benchmark" to 147.25×1.05=154.61 boxes / day to reflect the impact of fulfillment rate fluctuations on inventory consumption. (2) Refinement of the inbound link: The periodic correlation feature includes the "supplier arrival on-time rate".

[0046] (92%). Combining the relationship between the "on-time arrival rate" of the warehousing link and the replenishment cycle (for every 1% decrease in the on-time rate, the replenishment is delayed by 0.5 days), the server calculates the adjustment value of the current replenishment cycle (3 days): if the on-time rate is 92% (8% lower than the benchmark 100%), the replenishment delay is 0.5×8=4 days, and the actual replenishment cycle = 3+4=7 days. Therefore, the "replenishment cycle" feature is adjusted from 3 days to 7 days to reflect the supplier delay risk. (3) Sorting link refinement: The periodic correlation feature includes "weekend sorting efficiency 25 boxes / minute". Combining the relationship between the "sorting efficiency" of the sorting link and the outbound speed (for every 5 boxes / minute increase in efficiency, the outbound time is shortened by 0.5 hours), the server finds that the current efficiency (25 boxes / minute) is 5 boxes / minute higher than that of weekdays (20 boxes / minute), which can shorten the outbound time by 0.5 hours and reduce the risk of inventory backlog. Therefore, this feature is retained and marked as a "positive optimization factor". After refining each stage, the server generates three "current target feature sets": refined features for the outbound stage: [weekend sales baseline of 154.61 boxes / day]; refined features for the inbound stage: [replenishment cycle of 7 days]; and refined features for the sorting stage: [sorting efficiency improvement factor of +0.5 hours]. The server then weights and integrates these three current target feature sets based on the weight coefficients for each stage (outbound 0.45, inbound 0.35, sorting 0.15). For example, outbound features (weight 0.45) contribute primarily to sales forecast adjustments; inbound features (weight 0.35) contribute critical replenishment cycle adjustments; and sorting features (weight 0.15) contribute auxiliary outbound time optimization. The final integrated "target feature set" is: [weekend sales baseline of 154.61 boxes / day (0.45), replenishment cycle of 7 days (0.35), sorting efficiency improvement factor of +0.5 hours (0.15)]. This collection focuses on the core features that most significantly impact out-of-stock risk (such as adjusted sales values ​​and extended replenishment cycles), providing precise input for subsequent prediction of the out-of-stock risk index for SKU-20240301-005. For example, the server combines the current inventory (500 boxes) with the adjusted weekend sales (154.61 boxes / day) to calculate the inventory consumption over three days as 154.61 x 3 = 463.83 boxes, leaving only 36.17 boxes (below the safety stock threshold of 50 boxes). Therefore, the server determines that the out-of-stock risk is high and requires an immediate replenishment order.

[0047] In an embodiment of the present invention, the multiple current target feature sets are integrated according to the link weight coefficients corresponding to the multiple link feature enhancement information to obtain the target feature set, which can be implemented through the following examples.

[0048] Acquiring characteristic dimensions of the plurality of link feature enhancement information, adjusting the link weight coefficient corresponding to each preset link feature enhancement information according to the characteristic dimensions, and obtaining an adjustment amount corresponding to each preset link feature enhancement information;

[0049] Normalizing the adjustment amount corresponding to each preset link feature enhancement information to obtain the target link weight coefficient corresponding to each preset link feature enhancement information;

[0050] Determining the target link weight coefficients corresponding to each of the plurality of link feature enhancement information from the target link weight coefficients corresponding to each of the preset link feature enhancement information;

[0051] The multiple current target feature sets are integrated according to the target link weight coefficients corresponding to the multiple link feature reinforcement information to obtain the target feature set.

[0052] In an embodiment of the present invention, for example, taking the server processing of a fresh food e-commerce cold chain storage center as an example, for boxed fresh milk (250 mL / box) labeled "SKU-20240301-005", the server needs to adjust the weight according to the dimensional differences of the link features and integrate the current target feature set. The specific process is as follows: the server first counts the "feature dimensions" of the feature enhancement information of each link (that is, the number or complexity of features contained in the link). Taking the three selected links of outbound, inbound, and sorting as an example: Outbound includes two features: "order fulfillment rate" and "outbound punctuality rate," with "order fulfillment rate" strongly correlated with the sales cycle (involving historical fluctuations), and a feature dimension of 2 (high complexity); Inbound includes two features: "supplier arrival punctuality rate" and "quality inspection pass rate," but "quality inspection pass rate" remains stable over the long term (100%), with a feature dimension of 1.5 (medium complexity); Sorting includes only one feature: "sorting efficiency" (weakly correlated with outbound time), with a feature dimension of 1 (low complexity). The server adjusts the original link weight coefficients (0.45 for outbound, 0.35 for inbound, and 0.15 for sorting) based on the feature dimensions: the higher the dimension, the more weight is needed to reflect its information value. The adjustment formula is: Adjustment amount = original weight × (feature dimension / total dimension). Total dimension = 2 + 1.5 + 1 = 4.5. Therefore, the adjustment for the outbound stage = 0.45 × (2 / 4.5) = 0.45 × 0.444 ≈ 0.2; the adjustment for the inbound stage = 0.35 × (1.5 / 4.5) = 0.35 × 0.333 ≈ 0.117; and the adjustment for the sorting stage = 0.15 × (1 / 4.5) = 0.15 × 0.222 ≈ 0.033. The server normalizes the adjusted stage weights to ensure that the sum of all weights is 1. The adjusted total weight = 0.2 + 0.117 + 0.033 = 0.35. Therefore, the target weights for outbound operations need to be scaled up accordingly: 0.2 / 0.35 ≈ 0.571; 0.117 / 0.35 ≈ 0.334; and 0.033 / 0.35 ≈ 0.095. Finally, the "target link weight coefficient" is: outbound 0.571, inbound 0.334, and sorting 0.095. The server only retains the target weights of the three selected links (outbound, inbound, and sorting), excluding the filtered inventory link (weight 0). Subsequently, the "current target feature set" of each link is weighted and integrated according to these weights. Taking "SKU-20240301-005" as an example, the current target feature set of each link is: outbound link: [weekend sales benchmark 154.61 boxes / day]; inbound link: [replenishment cycle 7 days]; sorting link: [sorting efficiency improvement factor + 0.5 hours].The server performs a weighted integration based on the target weights: outbound feature contribution: 154.61 boxes / day × 0.571 ≈ 88.3 boxes / day; inbound feature contribution: 7 days × 0.334 ≈ 2.34 days; sorting feature contribution: +0.5 hours × 0.095 ≈ +0.0475 hours. The resulting "target feature set" is: [weekend sales impact value: 88.3 boxes / day, replenishment cycle impact value: 2.34 days, sorting efficiency optimization value: +0.0475 hours]. Through dimensional adjustment and standardization, the server ensures that the impact of high-dimensional, high-complexity features (such as outbound) on the final target feature is appropriately amplified, while the influence of low-dimensional features (such as sorting) is appropriately weakened. For example, the adjusted outbound feature weight (0.571) is significantly higher than that of inbound (0.334) and sorting (0.095), reflecting the core influence of "order fulfillment rate" on inventory consumption. The resulting target feature set more accurately focuses on key risk factors (such as sales-adjusted values ​​and extended replenishment cycles), providing a reliable basis for subsequent predictions of the out-of-stock risk index for "SKU-20240301-005" (for example, based on the current inventory of 500 boxes, the consumption in three days is calculated to be 88.3 × 3 = 264.9 boxes, leaving 235.1 boxes, which is lower than the safety stock threshold of 300 boxes and is judged to be high risk).

[0053] In an embodiment of the present invention, adjusting the link weight coefficient corresponding to each preset link feature enhancement information according to the feature dimension to obtain the adjustment amount corresponding to each preset link feature enhancement information can be implemented through the following examples.

[0054] Determine, according to the feature dimension, from the link weight coefficients corresponding to each preset link feature enhancement information, the weight coefficients of each link that is ranked first and the weight coefficients of each link that is ranked last;

[0055] The weight coefficient of each link arranged at the front remains unchanged, and the weight coefficient of each link arranged at the back is adjusted to the preset period feature base amount, so as to obtain the adjustment amount corresponding to each preset link feature enhancement information.

[0056] In an embodiment of the present invention, for example, taking the server processing of the cold chain storage center of a fresh food e-commerce company as an example, for boxed fresh milk (250 mL / box) labeled "SKU-20240301-005", the server needs to adjust the weight coefficient of the preset link according to the feature dimension. The specific process is as follows: the server first defines the "feature dimension" as the complexity of the link feature enhancement information, which is specifically calculated through two indicators: the number of features and the correlation with periodic related features. For example: the outbound link: contains two features, "order fulfillment rate" (strongly correlated with the sales cycle, correlation coefficient 0.9) and "outbound punctuality rate" (related to inventory consumption speed, correlation coefficient 0.8), dimension = 2×(0.9+0.8) / 2=1.7 (high complexity); the inbound link: contains two features, "supplier arrival punctuality rate" (strongly correlated with the replenishment cycle, correlation coefficient 0.7) and "quality inspection pass rate" (long-term stability, correlation coefficient 0.2), dimension = 2×(0.7+0.2) / 2=0.9 (medium complexity); the sorting link: contains only one feature, "sorting efficiency" (weakly correlated with outbound time, correlation coefficient 0.5), dimension = 1×0.5=0.5 (low complexity); the inventory link: contains only one feature, "inventory difference rate" (weakly correlated with inventory accuracy, correlation coefficient 0.3), dimension = 1×0.3=0.3 (extremely low complexity). The server sorts the links from high to low according to the dimension value, and obtains the order of links: outbound (1.7), inbound (0.9), sorting (0.5), and inventory (0.3). Among them, the ones ranked at the top are outbound and inbound (dimension ≥ 0.9), and the ones ranked at the bottom are sorting and inventory (dimension < 0.9). The server pre-sets the "preset cycle feature base" to 0.1 (that is, the weight of the low-complexity link is as low as 0.1). For the links ranked at the top (outbound and inbound), their original link weight coefficients remain unchanged; for the links ranked at the bottom (sorting and inventory), their weights are adjusted to the preset base. The original link weight coefficients (from feature importance assessment) are: outbound: 0.45 (front column, remains unchanged); inbound: 0.35 (front column, remains unchanged); sorting: 0.15 (back column, adjusted to 0.1); inventory: 0.05 (back column, adjusted to 0.1). The adjusted values ​​are: outbound adjustment = 0.45; inbound adjustment = 0.35; sorting adjustment = 0.1; and inventory count adjustment = 0.1. This adjustment ensures that the weights of high-complexity steps (such as outbound and inbound) are not restricted by the dimensions, preserving their core impact on stock-out risk. The weights of low-complexity steps (such as sorting and inventory count) are uniformly increased to the preset base value, preventing feature information from being completely ignored due to low initial weights (such as 0.05 for inventory count).Taking "SKU-20240301-005" as an example, the adjusted link weights are more in line with the actual business logic: the 0.45 weight of the outbound link (high dimension) dominates the adjustment of sales forecasts (such as the weekend sales benchmark is adjusted from 147.25 boxes / day to 154.61 boxes / day), the 0.35 weight of the warehousing link (medium dimension) affects the calculation of the replenishment cycle (such as extending from 3 days to 7 days), and the 0.1 weight of the sorting link (low dimension) only assists in optimizing the outbound time (such as shortening it by 0.5 hours), and the 0.1 weight of the inventory link is used to verify inventory accuracy (such as not adjusting the forecast when the difference rate is <0.5%). This adjustment strategy balances the rationality of feature complexity and weight distribution, allowing the subsequent integrated target feature set to focus on core risk factors while retaining necessary auxiliary information, ultimately improving the accuracy of out-of-stock risk prediction (for example, based on the adjusted weight calculation, the server determines that the three-day inventory consumption of "SKU-20240301-005" is 154.61×3=463.83 boxes. Combined with the replenishment cycle of 7 days, 100 boxes need to be replenished immediately to avoid out-of-stock).

[0057] In an embodiment of the present invention, the target feature set is obtained by integrating the multiple current target feature sets according to the target link weight coefficients corresponding to the multiple link feature enhancement information, which can be implemented through the following examples.

[0058] For each link feature enhancement information, weight the corresponding current target feature set according to the target link weight coefficient to obtain each link weighted feature set;

[0059] The weighted feature set of each link is integrated to obtain the target feature set.

[0060] In an embodiment of the present invention, for example, taking the server processing of a fresh food e-commerce cold chain storage center as an example, for boxed fresh milk (250 mL / box) labeled "SKU-20240301-005", the server needs to weight the current target feature set of each link based on the target link weight coefficient, and integrate to generate the final target feature set. The specific process is as follows: the server first obtains the "target link weight coefficient" of each link (from the standardized result of step 6): 0.571 for the outbound link, 0.334 for the inbound link, and 0.095 for the sorting link. At the same time, the "current target feature set" of each link has been generated through feature refinement, specifically: Outbound link: Current target feature set = [weekend sales benchmark 154.61 boxes / day] (reflecting the sales adjustment value after the fulfillment rate fluctuation); Inbound link: Current target feature set = [replenishment cycle 7 days] (reflecting the replenishment cycle adjustment value after the supplier delay); Sorting link: Current target feature set = [sorting efficiency improvement factor + 0.5 hours] (reflecting the optimization value of sorting efficiency improvement on outbound time). The server performs a "link weighting" operation for each link, that is, using the target weight coefficient to weight the current feature, highlighting the core influence of the high-weight link. (1) Outbound link weighting: The server extracts the target weight of the outbound link 0.571, multiplies it with the current target feature "weekend sales benchmark 154.61 boxes / day", and calculates the weighted feature value: 154.61 boxes / day × 0.571 ≈ 88.3 boxes / day. This value represents the core impact of the outbound link on weekend sales (because it has the highest weight, it contributes the main sales forecast adjustment). (2) Weighting of the incoming link: The server extracts the target weight of the incoming link, 0.334, and multiplies it by the current target feature "replenishment cycle 7 days". The weighted feature value is calculated as follows: 7 days × 0.334 ≈ 2.34 days. This value represents the key impact of the incoming link on the replenishment cycle (because it has the second highest weight, it reflects the main contribution of supplier delays to replenishment time). (3) Weighting of the sorting link: The server extracts the target weight of the sorting link, 0.095, and multiplies it by the current target feature "sorting efficiency improvement factor + 0.5 hours". The weighted feature value is calculated as follows: +0.5 hours × 0.095 ≈ +0.0475 hours (about 2.85 minutes). This value represents the auxiliary optimization of the sorting link on the outbound time (because it has the lowest weight, it only contributes a small time adjustment). After completing the weighting process, the server generates three "link weighted feature sets": outbound weighted feature set = [88.3 boxes / day]; inbound weighted feature set = [2.34 days]; and sorting weighted feature set = [+0.0475 hours]. The server sequentially concatenates the weighted feature sets for each link to form the final "target feature set." This set combines the core influence of high-weighted links with auxiliary information from low-weighted links, comprehensively reflecting the key factors influencing out-of-stock risk.Taking "SKU-20240301-005" as an example, the integrated target feature set is: [88.3 boxes / day (weighted outbound sales volume), 2.34 days (weighted inbound replenishment cycle), +0.0475 hours (optimized sorting-weighted outbound time)]. This set is directly used as input to the intelligent warehouse decision-making model to predict the out-of-stock risk index. For example, the server combines the current inventory (500 boxes) and the target feature "weighted outbound sales volume 88.3 boxes / day" to calculate the inventory consumption over three days as 88.3 × 3 = 264.9 boxes, leaving a remaining inventory of 500 - 264.9 = 235.1 boxes. Furthermore, the "weighted inbound replenishment cycle 2.34 days" indicates that it takes approximately 2.34 days for the supplier to replenish the goods. If the safety stock threshold is 300 boxes, and the current inventory (235.1 boxes) is below the threshold, the server determines that the out-of-stock risk index for this tag is 85 points (high risk), and a replenishment instruction must be generated immediately (for example, replenishment quantity = 300 - 235.1 = 64.9 boxes, rounded up to 65 boxes). By empowering and integrating links, the server ensures that the core features of high-weighted links (such as outbound delivery) dominate the prediction results, while retaining auxiliary information from low-weighted links (such as sorting). This makes the prediction more relevant to actual business scenarios and ultimately improves the accuracy and timeliness of replenishment decisions.

[0061] In an embodiment of the present invention, after refining the periodically associated feature set according to the plurality of link feature enhancement information to obtain a target feature set, the following implementation manner is also provided.

[0062] The target feature set is used as the integrated feature set, and the steps of extracting periodic correlation information of integrated features in the integrated feature set to obtain a periodic correlation feature set are returned and executed until a replenishment forecast accuracy condition is met, thereby obtaining a precise target feature set;

[0063] The precise target feature set is used to predict and obtain the precise out-of-stock risk index corresponding to each to-be-replenished goods label, and generate a replenishment instruction based on the precise out-of-stock risk index corresponding to each to-be-replenished goods label.

[0064] In the embodiment of the present invention, for example, taking the server processing of the cold chain warehousing center of the fresh food e-commerce as an example, for the boxed fresh milk (250mL / box) labeled "SKU-20240301-005", after the server generates the target feature set for the first time, it needs to improve the prediction accuracy through iterative optimization. The specific process is as follows: After the server generates the "target feature set" for the first time, it uses it as the new "integrated feature set" and re-executes the "extract periodic correlation information-feature refinement" process to start the first round of iteration. (1) Initial prediction and deviation detection: The server uses the first round of target feature set (such as [88.3 boxes / day, 2.34 days, +0.0475 hours]) to input the warehouse intelligent decision model, predicts that the out-of-stock risk index of "SKU-20240301-005" in the next 7 days is 85 points (high risk), and generates a replenishment instruction (recommended replenishment of 65 boxes). However, in actual operation, only 480 boxes of this SKU were consumed within 7 days (with 20 boxes remaining), and no out-of-stock situation was triggered (the actual risk index should be 60 points). The server compared the predicted value (85 points) with the actual value (60 points) and calculated the deviation rate = (85-60) / 60≈41.7%, which far exceeded the preset accuracy threshold (5%), triggering iterative optimization. The server analyzed the cause of the deviation and found that the "outbound weighted sales volume of 88.3 boxes / day" in the first round of target features overestimated the actual weekend sales (actual weekend sales volume was 140 boxes / day, instead of the adjusted 154.61 boxes / day). This was mainly because the historical fluctuation data of the "order fulfillment rate" did not fully cover recent anomalies (for example, due to a promotion on a certain weekend, the fulfillment rate increased to 99%, instead of the historical average of 98%). The server used the first round of target feature set as the new integrated feature set and re-executed the periodic feature mapping. By supplementing the recent 30 days of fulfillment rate data (99%), recalculating the correlation factor for the 7-day cycle (increased from 0.82 to 0.85), and adjusting the base quantity of periodic features (the weekend sales benchmark was revised from 154.61 boxes / day to 140 boxes / day), a new "periodic correlation feature set" was generated. The server re-evaluated the link weight coefficients and found that the recent stability of the "order fulfillment rate" had improved (the volatility dropped from 22% to 15%). Therefore, the target weight of the outbound link was adjusted from 0.571 to 0.5 (reducing its excessive impact on sales forecasts), and the weight of the inbound link was adjusted to 0.4 (to improve the accuracy of the replenishment cycle). The periodic correlation features were refined with the new link weights to generate the second round of "current target feature set" (such as a weekend sales benchmark of 140 boxes / day and a replenishment cycle of 6 days). The server weights and integrates the second round of current target feature set according to the adjusted weights (0.5 for outbound delivery, 0.4 for inbound delivery, and 0.1 for sorting), and obtains the new target feature set: [140×0.5=70 boxes / day (weighted sales for outbound delivery), 6×0.4=2.4 days (weighted replenishment cycle for inbound delivery), +0.5×0.1=0.05 hours (sorting weight optimization)].The server re-predicts the out-of-stock risk index for "SKU-20240301-005" using the second-round target feature set, outputting a score of 65 (close to the actual value of 60). The deviation rate is (65-60) / 60≈8.3%, still above the 5% threshold, and further iteration is performed. The server further optimizes the periodic feature base (incorporating weather data to find that with no recent extreme weather, weekend sales are more stable) and adjusts the replenishment cycle for incoming goods (the supplier's recent on-time performance has increased to 95%, shortening the replenishment cycle from 6 days to 5 days). This generates the third-round target feature set: [135 × 0.55 = 74.25 boxes / day (weighted outbound sales), 5 × 0.35 = 1.75 days (weighted inbound replenishment cycle)]. The prediction results show a risk index of 62 (actual value 60) with a deviation rate of 3.3% ≤ 5%, meeting the accuracy requirement. The server then designates the third-round target feature set as the "precise target feature set." Based on a precise target feature set (e.g., weekend sales of 74.25 boxes / day and a replenishment cycle of 1.75 days), the server combines the current inventory (20 boxes remaining) and the safety stock threshold (50 boxes) to calculate the replenishment quantity = 50-20+(74.25×1.75)≈30+130=160 boxes (rounded up). This ultimately generates a precise replenishment instruction: "SKU-20240301-005, replenishment quantity 160 boxes, required to arrive before 2024-03-10-12:00," ensuring that inventory covers demand within the future replenishment cycle and avoiding stockouts. Through multiple rounds of iterative optimization, the server gradually converged from the initial high-deviation prediction to a precise target feature set. The resulting replenishment instruction fits actual demand, effectively reducing inventory waste and stockout risks.

[0065] In an embodiment of the present invention, after extracting multiple link feature enhancement information corresponding to the periodic associated feature set from each preset link feature enhancement information, and refining the periodic associated feature set based on the multiple link feature enhancement information to obtain the target feature set, the following implementation method is also provided.

[0066] Performing inventory parameter mapping on the target feature set according to each of the to-be-replenished goods labels to obtain inventory parameter mapping features corresponding to each of the to-be-replenished goods labels;

[0067] Converting the inventory parameter mapping features corresponding to each of the to-be-replenished goods labels into a stock-out risk index to obtain a stock-out risk index corresponding to each of the to-be-replenished goods labels;

[0068] screening each of the to-be-replenished goods labels according to the out-of-stock risk index corresponding to each of the to-be-replenished goods labels to obtain a target goods label;

[0069] Generate a replenishment instruction for the target goods label.

[0070] In an embodiment of the present invention, for example, taking the server processing of a fresh food e-commerce cold chain storage center as an example, for the "SKU-20240301-005" boxed fresh milk (250mL / box) for which a target feature set has been generated, the server needs to complete inventory parameter mapping, risk index conversion, target goods screening and replenishment instruction generation. The specific process is as follows: the server converts the abstract features in the target feature set into specific parameters for warehouse management (such as safety stock, replenishment points, average daily consumption, etc.), which can be directly used for inventory risk assessment. Take the target feature set of "SKU-20240301-005" (such as [74.25 boxes / day (outbound empowerment sales), 1.75 days (inbound empowerment replenishment cycle), +0.05 hours (sorting empowerment optimization)]) as an example: Average daily consumption: directly take the "outbound empowerment sales" of 74.25 boxes / day (reflecting the actual consumption rate after considering the fulfillment rate); Replenishment lead time: take the "inbound empowerment replenishment cycle" of 1.75 days (reflecting the replenishment time after the supplier arrives on time); Safety stock: calculated according to the historical fluctuation coefficient (such as 1.2), safety stock = average daily consumption × replenishment lead time × fluctuation coefficient = 74.25 × 1.75 × 1.2 ≈ 156 boxes; Current available inventory: obtain real-time data from WMS, the current inventory is 20 boxes (due to rapid consumption due to recent promotions). These parameters constitute the "inventory parameter mapping feature": [average daily consumption 74.25 boxes / day, replenishment lead time 1.75 days, safety stock 156 boxes, current inventory 20 boxes]. The server uses a preset risk assessment model (or formula) to convert the inventory parameter mapping feature into a stock-out risk index ranging from 0 to 100 (the higher the score, the greater the probability of stock-out). In this example, the server uses a linear regression model with the input [current inventory / safety stock, replenishment lead time] and outputs the risk index. The calculation formula is: Risk Index = 100 × (1 - current inventory / safety stock) + replenishment lead time × 10. Substitute the following data: Current inventory / Safety stock = 20 / 156 ≈ 0.128, 1 - 0.128 = 0.872; Replenishment lead time = 1.75 days, 1.75 × 10 = 17.5; Risk index = 100 × 0.872 + 17.5 = 87.2 + 17.5 = 104.7 (Because the current inventory is far below the safety stock, the risk index exceeds 100 and is marked as "very high risk"). The server sets a risk threshold (for example, 70) and filters out the tags for replenishment with a risk index ≥ 70. In this example, "SKU-20240301-005" has a risk index of 104.7, far exceeding the threshold, and is identified as a "target item tag." The server calculates the replenishment quantity based on the inventory parameter mapping features and generates a replenishment instruction with specific execution requirements. The replenishment quantity calculation formula is: Replenishment quantity = Safety stock - Current inventory + Average daily consumption × Replenishment lead time (covering consumption during the replenishment cycle).Substituting the data: Replenishment quantity = 156 - 20 + 74.25 × 1.75 ≈ 136 + 130 ≈ 266 boxes (rounded up to 270 boxes to ensure redundancy). The resulting replenishment instruction is: "Replenishment instruction [20240310-001]: Goods label SKU-20240301-005 (250mL boxed fresh milk), replenishment quantity 270 boxes, supplier A required to deliver before 2024-03-11 08:00, quality inspection pass rate must be ≥ 100%, and upon arrival, prioritized sorting to shelf 2 in the ambient temperature area." The server sends the replenishment instruction to the supplier's system and warehouse execution terminal, and the supplier delivers 270 boxes of fresh milk on time. Subsequent tracking revealed that this SKU consumed 74.25 x 1.75, which equals 130 boxes, within the replenishment cycle (1.75 days). Adding the original inventory of 20 boxes, the remaining quantity after consumption was 20 + 270 - 130 = 160 boxes, exceeding the safety stock of 156 boxes, successfully avoiding a stock-out. Through this process, the server converts abstract target features into executable replenishment instructions, ensuring a closed-loop inventory management system from prediction to execution, effectively improving the accuracy and efficiency of warehouse operations.

[0071] In the embodiments of the present invention, the following implementation modes are also provided.

[0072] Loading the inventory goods tag set into a warehouse intelligent decision model, the warehouse intelligent decision model including a feature encoding component, a periodic correlation feature encoding component, a feature refining component, and an instruction generation component;

[0073] Acquiring, by means of the feature coding component, an inventory feature set and a warehousing operation feature set corresponding to the inventory tag set, and integrating the inventory feature set with the warehousing operation feature set to obtain an integrated feature set;

[0074] Extracting periodic correlation information of the integrated features in the integrated feature set through the periodic correlation feature encoding component to obtain a periodic correlation feature set;

[0075] Extracting multiple link feature enhancement information corresponding to the periodic associated feature set from each preset link feature enhancement information through the feature refining component, and performing feature refinement on the periodic associated feature set based on the multiple link feature enhancement information to obtain a target feature set;

[0076] The instruction generation component predicts the out-of-stock risk index corresponding to each of the to-be-replenished goods labels based on the target feature set, and generates a replenishment instruction based on the out-of-stock risk index corresponding to each of the to-be-replenished goods labels.

[0077] In the embodiment of the present invention, for example, the server of the cold chain warehousing center of the fresh food e-commerce is used as the execution body. For the boxed fresh milk (250mL / box) labeled "SKU-20240301-005", the server completes the whole process from data input to replenishment instruction generation through the "warehouse intelligent decision model". The specific operation steps are as follows: At 0:00 am every day, the server obtains 2000 inventory goods labels (including "SKU-20240301-005") in the warehouse on that day from the warehouse management system (WMS), and inputs these labels into the "warehouse intelligent decision model" in batches. After the model is started, the "feature coding component" first takes over the data processing. (1) Extracting inventory goods features: The feature coding component retrieves its attribute information from the product master database based on the label "SKU-20240301-005": category = dairy products, shelf life = 7 days, sales volume in the past 7 days = 114 boxes / day, sales volatility = 22% (historical data). The component converts this information into a 10-dimensional feature vector: [0 (non-vegetable), 1 (dairy), 0 (non-meat), 0 (non-fruit), 0.25 (normalized to 250 mL), 7 (shelf life), 0.4 (normalized to storage temperature), 114 (average daily sales), 0.22 (volatility), 1.5 (promotional increase)]. (2) Extracting warehouse operation features: The component extracts the warehouse operation information of the tag from the WMS operation log: on-time rate of incoming goods = 92%, sorting efficiency = 20 boxes / minute, on-time rate of outgoing goods = 90%, inventory difference rate = 0.27% (average of the last three times). Converted into a 9-dimensional operation time series vector: [0.92 (punctuality), 0.5 (normalized incoming inventory), 0.5 (normalized sorting time), 20 (sorting efficiency), 0.9 (outgoing inventory punctuality), 0.98 (fulfillment rate), 0.0027 (difference rate), 0.5 (incoming-sorting interval), 7.5 (sorting-outgoing interval)]. (3) Integration features: The component concatenates the two types of vectors into a 19-dimensional integrated feature vector to form an “integrated feature set”. For example, the integrated vector of “SKU-20240301-005” is: [0, 1, 0, 0, 0.25, 7, 0.4, 114, 0.22, 1.5, 0.92, 0.5, 0.5, 20, 0.9, 0.98, 0.0027, 0.5, 7.5]. The component performs periodic analysis on the time series data (such as average daily sales and number of outbound orders) in the integrated features. A Fourier transform revealed that the sales of "SKU-20240301-005" have a 7-day primary cycle (accounting for 82% of the energy), manifested by weekend sales being 30% higher than weekdays (e.g., 150 boxes on Saturday and 160 boxes on Sunday). The component generates the following: target period correlation factor set: [0.82 (7-day factor), 0.15 (14-day promotion factor)]; period feature index set: [7 (days), 14 (days)]; period feature base quantity set: [100 (weekday sales baseline), 155 (weekend sales baseline)].The component then calculates the correlation between the factor and the index (a 7-day factor correlation coefficient of 0.95), restructures the base volume (the weekend sales benchmark is adjusted to 155 × 0.95 = 147.25 boxes / day), and generates a "periodic correlation feature set": [7-day primary cycle (weekend sales of 147.25 boxes / day), 14-day secondary cycle (9.3% increase due to the promotion)]. Based on the reinforcement information of the four preset links (warehousing, sorting, outbound, and inventory), the component evaluates the link weights using SHAP values ​​(0.45 for outbound, 0.35 for inbound, 0.15 for sorting, and 0.05 for inventory), filtering out the high-weighted outbound and inbound links. (1) Refinement of the outbound link: Based on the historical fluctuation of the "order fulfillment rate of 98%", the weekend sales benchmark was adjusted from 147.25 boxes / day to 154.61 boxes / day (considering an additional 5% consumption when the fulfillment rate is <95%); (2) Refinement of the inbound link: Based on the "supplier punctuality rate of 92%", the replenishment cycle was extended from 3 days to 7 days (a delay of 0.5 days for every 1% decrease in punctuality); and finally the "target feature set" was generated: [weekend sales of 154.61 boxes / day (outbound weighting 0.571), replenishment cycle of 7 days (inbound weighting 0.334)]. The component inputs the target features into the risk assessment model and calculates the out-of-stock risk index for SKU-20240301-005: the current inventory is 20 boxes, the safety stock is 156 boxes (74.25 boxes / day × 1.75 days × 1.2), and the risk index = 100 × (1-20 / 156) + 1.75 × 10 ≈ 104.7 points (extremely high risk). The component identifies the target product with this label, calculates the replenishment quantity (156-20+74.25×1.75≈270 boxes), and generates a replenishment instruction: "Replenishment instruction [20240310-001]: SKU-20240301-005 (250mL fresh milk), replenish 270 boxes, supplier A delivered before 2024-03-11-08:00, quality inspection must pass 100%." ​​Through the coordinated operation of the four major components of the model, the server completes the entire process from raw label data to accurate replenishment instructions, ensuring the accuracy of inventory forecasts and the timeliness of replenishment decisions.

[0078] In an embodiment of the present invention, the warehouse intelligent decision model is obtained in the following manner and can be implemented through the following examples.

[0079] Acquire a historical inventory goods label set, and determine a sample inventory goods label set and a target value of the sample goods label based on the historical inventory goods label set;

[0080] Loading the sample inventory goods label set into the original warehouse intelligent decision model to obtain the sample out-of-stock risk index corresponding to each generated to-be-replenished goods label;

[0081] Calculating the cost based on the sample out-of-stock risk index corresponding to each of the to-be-replenished goods labels and the target value of the sample goods labels to obtain a cost parameter;

[0082] The original warehousing intelligent decision-making model is trained according to the cost parameter until the warehousing intelligent decision-making model is obtained.

[0083] In an embodiment of the present invention, for example, the server of the cold chain warehousing center of the fresh food e-commerce is used as the execution body, and it is necessary to train the "warehouse intelligent decision-making model" through historical data to ensure that the model can accurately predict the risk of out-of-stock. The following takes the historical data of the center from January to December 2023 as an example to explain the model training process in detail: the server retrieves the historical data for the whole year of 2023 from the warehouse management system (WMS), order system and supplier system, including: a set of historical inventory goods labels: 3,000 SKU labels covering 12 months (such as "SKU-20230301-005" fresh milk, "SKU-20230515-012" fresh mushrooms, etc.), each label is associated with its inventory changes throughout the year, warehousing operation records (warehousing / sorting / outbound time, quantity) and sales data (average daily sales, promotional activities). Target value of sample goods label: "Whether out of stock actually occurs" is used as the target value (1 = out of stock, 0 = not out of stock), and combined with the severity of the out-of-stock (such as the number of days out of stock), it is converted into a continuous value of 0-100 points (for example, a SKU is out of stock for 3 days in June 2023, the target value = 85 points; the target value of the SKU that is not out of stock = 30 points). The server divides the training set (2400 labels) and the validation set (600 labels) in an 8:2 ratio to ensure that the samples cover different categories (dairy products, vegetables, meat) and seasons (such as high consumption in summer and low consumption in winter). The server initializes the "original warehousing intelligent decision model" (based on the XGBoo st framework, with random initial parameters), which includes four major components: feature encoding, period association encoding, feature refinement, and instruction generation. The 2400 sample labels in the training set are input into the model one by one. The model generates the “sample out-of-stock risk index” according to the following process: (1) Feature encoding component: For each label, inventory characteristics (such as shelf life, sales volatility) and warehouse operation characteristics (such as on-time storage rate, sorting efficiency) are extracted and integrated into a 19-dimensional feature vector (for example, the vector of “SKU-20230301-005” contains [0,1,0,0,0.25,7,0.4,114,0.22,1.5,0.92,0.5,0.5,20,0.9,0.98,0.0027,0.5,7.5]). (2) Cycle association encoding component: Identify the cyclical pattern in the features (such as a 7-day sales cycle) and generate cyclical association features (such as the weekend sales benchmark of 147.25 boxes / day). (3) Feature Refining Component: Filter key features (such as the order fulfillment rate in the outbound phase) based on the weights of the preset links (outbound, inbound, etc.), and generate target features after refinement (such as the adjusted weekend sales of 154.61 boxes / day). (4) Instruction Generation Component: Output the sample out-of-stock risk index (such as the initial predicted value of "SKU-20230301-005" = 70 points, while its actual target value = 85 points). The server uses "mean square error (MSE)" as the cost function to calculate the difference between the predicted value and the target value of each sample.For example: Sample 1 ("SKU-20230301-005"): predicted value 70 points, target value 85 points, error = (70-85)2 = 225; Sample 2 ("SKU-20230515-012" fresh mushrooms): predicted value 40 points, target value 30 points, error = (40-30). 2 =100; the overall MSE of the training set = (225+100+…) / 2400≈180 (the initial error is large). The server adjusts the model parameters (such as the normalization coefficient of the feature encoding component, the Fourier transform parameter of the period-related component, and the link weight of the feature refinement component) through the gradient descent algorithm to reduce the MSE. After each round of training, the generalization ability is evaluated using the validation set (600 samples). (1) First round of training: Adjust the "sales volatility" weight of the feature encoding component (from 0.22 to 0.3) to make the model pay more attention to SKUs with large fluctuations. The training set MSE is reduced to 150, and the validation set MSE = 160 (overfitting risk). (2) Second round of training: Add a "seasonal factor" (such as summer sales + 20%) to the feature refinement component to correct the period-related features. The training set MSE = 120, and the validation set MSE = 130 (error reduction). (3) The third round of training: Optimize the risk assessment formula of the instruction generation component (from linear regression to logistic regression), and introduce the "safety stock threshold" as a nonlinear factor. The training set MSE = 80, and the validation set MSE = 85 (close to the threshold of 100). When the validation set MSE ≤ 100, the model meets the standard. Finally, the server confirms the trained model as the "warehouse intelligent decision model". Its prediction value for "SKU-20230301-005" = 83 points (with an error of only 2 points from the target value of 85 points), and its prediction value for "SKU-20230515-012" = 32 points (with an error of 2 points from the target value of 30 points), meeting business requirements. Through the above training process, the server optimizes the original model into an intelligent decision model that can accurately capture inventory characteristics, cycle patterns and link influences, providing reliable algorithm support for subsequent real-time prediction and replenishment instruction generation.

[0084] In an embodiment of the present invention, the inventory goods tag set includes an inventory operation record identifier set, wherein the inventory operation record identifier set includes each inventory operation record identifier of the warehouse system, and each to-be-replenished goods tag includes each preset replenishment candidate identifier. The following implementation manner is also provided.

[0085] Obtaining a target feature set corresponding to the inventory operation record identification set, and predicting an out-of-stock risk index corresponding to each of the preset replenishment candidate identifications based on the target feature set corresponding to the inventory operation record identification set;

[0086] screening each of the preset replenishment candidate identifiers according to the out-of-stock risk index corresponding to each of the preset replenishment candidate identifiers to obtain a replenishment execution identifier;

[0087] The replenishment parameter package of the replenishment execution identifier triggers replenishment execution.

[0088] In an embodiment of the present invention, for example, the server of a cold chain warehousing center in North China for a fresh food e-commerce company is used as the execution entity. The "inventory operation record identification set" in its inventory system contains 1,500 operation records on March 10, 2024 (such as the incoming order ID "RK-20240310-001", the sorting task ID "FJ-20240310-005", the outgoing order ID "CK-20240310-012", etc.), and the "preset replenishment candidate identification" is the 100 high-frequency consumption SKUs preset by the system (such as "SKU-20240301-005" boxed fresh milk, "SKU-20240302-018" fresh spinach, etc.). The server needs to complete the replenishment execution through the following steps: The server first extracts the original data of the "inventory operation record identification set" from the warehouse management system (WMS). For example, the operation records associated with the incoming order ID "RK-20240310-001" include: product tag "SKU-20240301-005" (fresh milk), incoming time = 2024-03-10 06:00, incoming quantity = 200 boxes, and supplier A's on-time arrival rate (last 30 days) = 95%. The sorting task ID "FJ-20240310-005" is associated with: sorting start time = 06:30, sorting completion time = 07:00, sorting efficiency = 25 boxes / minute, and damage rate = 0%. The outgoing order ID "CK-20240310-012" is associated with: number of outgoing orders = 50, outgoing on-time rate = 98%, and order fulfillment rate = 99%. The server then associates these operation record identifiers with the previously generated "target feature set" (which contains refined features such as sales cycle, replenishment cycle, and sorting efficiency). For example, the target feature set for SKU-20240301-005 is: [weekend sales benchmark 140 boxes / day (after outbound weighting), replenishment cycle 5 days (after inbound weighting), sorting efficiency improvement factor + 0.5 hours]. The server inputs the target features corresponding to the "pre-set replenishment candidate identifiers" (e.g., 100 high-frequency SKUs) into the warehouse intelligent decision-making model and outputs a stock-out risk index (0-100) for each candidate identifier. Taking SKU-20240301-005 as an example, the model combines its target features (weekend sales 140 boxes / day, replenishment cycle 5 days) with the current inventory (80 boxes remaining) to calculate the risk index: Risk index = 100 × (1 - current inventory / safety stock) + replenishment cycle × 10. Safety stock = 140 boxes / day × 5 days × 1.2 (fluctuation coefficient) = 840 boxes. Current inventory / safety stock = 80 / 840 ≈ 0.095, 1 - 0.095 = 0.905. Risk index = 100 × 0.905 + 5 × 10 = 90.5 + 50 = 140.5 points (very high risk). Other candidate identifiers, such as "SKU-20240302-018" (spinach), have a risk index of 45 points (low risk).The server sets a risk threshold of 70 points and selects candidate indicators with a risk index ≥70 as "replenishment execution indicators." In this example, SKU-20240301-005, with a score of 140.5, far exceeds the threshold and is selected as a replenishment execution indicator. SKU-20240302-018, with a score of 45 below the threshold, is excluded. The server generates a "replenishment parameter package" for SKU-20240301-005, including the following: Replenishment quantity: Safety stock - Current inventory + Consumption during replenishment cycle = 840 - 80 + 140 × 5 = 760 + 700 = 1,460 boxes (rounded up to 1,500 boxes); Supplier: Supplier A (with a historical on-time performance of 95%); Arrival time: Required before 08:00, March 15, 2024 (replenishment cycle of 5 days); Quality inspection requirements: ≥ 5 days remaining before expiration date and ≤ 0.5% breakage rate. The server sends the replenishment parameter package to the supplier collaboration system and warehouse execution terminal via an API. Upon receipt, the supplier's system automatically generates a purchase order, and the warehouse terminal simultaneously updates the "pending replenishment" status. Ultimately, the supplier delivers 1,500 boxes of fresh milk at 7:30 AM on March 15, 2024. After passing quality inspection, the milk is put into storage, bringing the inventory back to 80 + 1,500 = 1,580 boxes, exceeding the safety stock of 840 boxes and successfully avoiding a stock-out. Through this process, the server, based on inventory operation records and pre-set replenishment candidate identifiers, implements a closed-loop management process from feature association and risk prediction to replenishment execution, ensuring timely replenishment of high-risk goods and reducing stock-out losses in warehouse operations.

[0089] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned artificial intelligence-based warehouse intelligent inventory forecasting and replenishment method. Figure 2 As shown, Figure 2 This is a block diagram of the structure of a computer device 100 provided in an embodiment of the present invention. Computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or exchange, memory 111, processor 112, and communication unit 113 are electrically connected to each other, directly or indirectly. For example, these components can be electrically connected via one or more communication buses or signal lines.

[0090] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in light of the above teachings. These embodiments have been selected and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the present disclosure and to utilize various embodiments with various modifications as appropriate for the specific application contemplated.

Claims

1. An artificial intelligence-based warehouse intelligent inventory forecasting and replenishment method, characterized by: include: Acquire an inventory goods tag set, wherein the inventory goods tag set includes each inventory goods tag of the warehouse system; Obtaining an inventory feature set and a warehousing operation feature set based on the inventory tag set, wherein the inventory feature set is obtained by extracting feature information of the goods attributes of each inventory tag, and the warehousing operation feature set is obtained by extracting feature information of the warehousing operation of each inventory tag, wherein the warehousing operation information is determined from the warehousing link to which the inventory tag belongs; Integrating the inventory feature set and the warehousing operation feature set to obtain an integrated feature set, and extracting periodic correlation information of integrated features in the integrated feature set to obtain a periodic correlation feature set; Extracting multiple link feature enhancement information corresponding to the periodic associated feature set from each preset link feature enhancement information, and refining the periodic associated feature set according to the multiple link feature enhancement information to obtain a target feature set; The target feature set is used to predict and obtain the out-of-stock risk index corresponding to each of the to-be-replenished goods labels, and to generate a replenishment instruction based on the out-of-stock risk index corresponding to each of the to-be-replenished goods labels.

2. The method according to claim 1, characterized in that The acquiring of the inventory feature set and the warehousing operation feature set according to the inventory tag set includes: According to each inventory tag in the inventory tag set, searching for the inventory feature corresponding to each inventory tag in each predetermined inventory feature to obtain the inventory feature set; Acquire multiple storage operation information of each inventory cargo tag from multiple storage links to which each inventory cargo tag belongs; For each inventory goods tag, multiple operation feature vectors of warehousing operation information are extracted to obtain multiple operation time sequence vectors, and the multiple operation time sequence vectors are integrated to obtain the warehousing operation feature set.

3. The method according to claim 1, characterized in that The extracting the periodic correlation information of the integrated features in the integrated feature set to obtain the periodic correlation feature set includes: Performing periodic feature mapping on the integrated feature set to obtain a target periodic correlation factor set, a periodic feature index set, and a periodic feature base quantity set; The correlation between the target period correlation factor set and the period feature index set is calculated to obtain a correlation degree set, and the period feature base quantity set is feature reorganized according to the correlation degree set to obtain the periodic correlation feature set.

4. The method according to claim 1, wherein The step of extracting multiple link feature enhancement information corresponding to the periodically associated feature set from each preset link feature enhancement information, and refining the periodically associated feature set according to the multiple link feature enhancement information to obtain a target feature set includes: Performing feature importance evaluation based on the periodically associated feature set to obtain a link weight coefficient corresponding to each preset link feature enhancement information; Filtering each preset link feature enhancement information according to the link weight coefficient corresponding to each preset link feature enhancement information to obtain a plurality of link feature enhancement information corresponding to the periodically associated feature set; Refining the periodically associated feature sets according to the plurality of link feature enhancement information to obtain a plurality of current target feature sets; Acquire characteristic dimensions of the plurality of link feature enhancement information, and determine, based on the characteristic dimensions, from the link weight coefficients corresponding to each of the preset link feature enhancement information, a weight coefficient of each link that is ranked first and a weight coefficient of each link that is ranked last; The weight coefficient of each link arranged at the front remains unchanged, and the weight coefficient of each link arranged at the back is adjusted to the preset period feature base amount, to obtain the adjustment amount corresponding to each preset link feature enhancement information; Normalizing the adjustment amount corresponding to each preset link feature enhancement information to obtain the target link weight coefficient corresponding to each preset link feature enhancement information; Determining the target link weight coefficients corresponding to each of the plurality of link feature enhancement information from the target link weight coefficients corresponding to each of the preset link feature enhancement information; For each link feature enhancement information, weight the corresponding current target feature set according to the target link weight coefficient to obtain each link weighted feature set; The weighted feature set of each link is integrated to obtain the target feature set.

5. The method according to claim 1, wherein After refining the periodically associated feature set according to the plurality of link feature enhancement information to obtain a target feature set, the method further includes: The target feature set is used as the integrated feature set, and the steps of extracting periodic correlation information of integrated features in the integrated feature set to obtain a periodic correlation feature set are returned and executed until a replenishment forecast accuracy condition is met, thereby obtaining a precise target feature set; The precise target feature set is used to predict and obtain the precise out-of-stock risk index corresponding to each to-be-replenished goods label, and generate a replenishment instruction based on the precise out-of-stock risk index corresponding to each to-be-replenished goods label.

6. The method according to claim 1, characterized in that After extracting multiple link feature enhancement information corresponding to the periodic associated feature set from each preset link feature enhancement information, and refining the periodic associated feature set based on the multiple link feature enhancement information to obtain a target feature set, the method further includes: Performing inventory parameter mapping on the target feature set according to each of the to-be-replenished goods labels to obtain inventory parameter mapping features corresponding to each of the to-be-replenished goods labels; Converting the inventory parameter mapping features corresponding to each of the to-be-replenished goods labels into a stock-out risk index to obtain a stock-out risk index corresponding to each of the to-be-replenished goods labels; screening each of the to-be-replenished goods labels according to the out-of-stock risk index corresponding to each of the to-be-replenished goods labels to obtain a target goods label; Generate a replenishment instruction for the target goods label.

7. The method according to claim 1, characterized in that The method further comprises: Loading the inventory goods tag set into a warehouse intelligent decision model, the warehouse intelligent decision model including a feature encoding component, a periodic correlation feature encoding component, a feature refining component, and an instruction generation component; Acquiring, by means of the feature coding component, an inventory feature set and a warehousing operation feature set corresponding to the inventory tag set, and integrating the inventory feature set with the warehousing operation feature set to obtain an integrated feature set; Extracting periodic correlation information of the integrated features in the integrated feature set through the periodic correlation feature encoding component to obtain a periodic correlation feature set; Extracting multiple link feature enhancement information corresponding to the periodic associated feature set from each preset link feature enhancement information through the feature refining component, and performing feature refinement on the periodic associated feature set based on the multiple link feature enhancement information to obtain a target feature set; The instruction generation component predicts the out-of-stock risk index corresponding to each of the to-be-replenished goods labels based on the target feature set, and generates a replenishment instruction based on the out-of-stock risk index corresponding to each of the to-be-replenished goods labels.

8. The method according to claim 7, characterized in that The warehousing intelligent decision model is obtained by the following methods, including: Acquire a historical inventory goods label set, and determine a sample inventory goods label set and a target value of the sample goods label based on the historical inventory goods label set; Loading the sample inventory goods label set into the original warehouse intelligent decision model to obtain the sample out-of-stock risk index corresponding to each generated to-be-replenished goods label; Calculating the cost based on the sample out-of-stock risk index corresponding to each of the to-be-replenished goods labels and the target value of the sample goods labels to obtain a cost parameter; The original warehousing intelligent decision-making model is trained according to the cost parameter until the warehousing intelligent decision-making model is obtained.

9. The method according to claim 1, characterized in that The inventory goods tag set includes an inventory operation record identifier set, the inventory operation record identifier set includes each inventory operation record identifier of the warehouse system, and each of the to-be-replenished goods tags includes each preset replenishment candidate identifier. The method further includes: Obtaining a target feature set corresponding to the inventory operation record identification set, and predicting an out-of-stock risk index corresponding to each of the preset replenishment candidate identifications based on the target feature set corresponding to the inventory operation record identification set; screening each of the preset replenishment candidate identifiers according to the out-of-stock risk index corresponding to each of the preset replenishment candidate identifiers to obtain a replenishment execution identifier; The replenishment parameter package of the replenishment execution identifier triggers replenishment execution.

10. A server system, characterized in that: The method comprises a server, wherein the server is configured to execute the method according to any one of claims 1 to 9.

Images