Inventory health assessment and optimization method based on LSTM prediction and state-driven mechanism

Abstract

The invention discloses an inventory health assessment and optimization method based on LSTM prediction and a state-driven mechanism, and belongs to the technical field of e-commerce inventory intelligent management. According to the method, product code SKU warehousing, sales and inventory data are collected, after cleaning and feature extraction, an LSTM model is utilized to dynamically predict an inventory trend and determine a stable inventory interval, four stages of a product life cycle are automatically divided, and an inventory health index (IHI) is calculated. Based on a state driving mechanism, IHI weights and model parameters are dynamically corrected, three-level early warning of three types of risks is achieved, and related information is displayed through a visual panel. According to the method, a'prediction-evaluation-optimization 'closed loop is constructed, the inventory prediction precision, the health evaluation comprehensiveness and the risk early warning timeliness are improved, the manual intervention cost is reduced, and enterprise data-driven refined inventory management is assisted.

Description

Technical Field

[0001] The invention relates to the field of intelligent e-commerce inventory management technology, and in particular to an inventory health assessment and optimization method based on LSTM prediction and state-driven mechanism. Background Technology

[0002] In recent years, the rapid development of the e-commerce industry has led to a dramatic increase in the number of product SKUs and transaction frequency, making inventory management a core aspect of business operations. Traditional e-commerce inventory management systems typically rely on manual experience or simple forecasting methods based on linear models. These methods often lack accuracy when dealing with complex market fluctuations and promotional activities, easily leading to frequent inventory backlogs or stockouts. With the diversification of sales channels and the increasing randomness of consumer behavior, inventory changes exhibit significant non-linear characteristics, making it difficult for traditional methods to effectively capture long-term dependencies and seasonal patterns.

[0003] While some companies have adopted statistical time series models or moving average methods for inventory forecasting, these methods are limited in their performance when faced with multi-dimensional characteristics (such as promotions, holidays, and shelf-ready times) and struggle to adapt to dynamic data changes. Furthermore, traditional systems often rely on a single inventory level as the core indicator, lacking a dynamic identification mechanism for the product lifecycle. This prevents timely assessment of different product stages (introduction, growth, maturity, or decline) based on sales trends, leading to delayed replenishment decisions or inappropriate inventory adjustments.

[0004] On the other hand, e-commerce inventory management also faces the problem of an imperfect risk warning system. Most platforms only judge inventory risk through static thresholds and lack dynamic early warning functions based on trend prediction, making it impossible to simultaneously identify multiple risk types such as slow-moving goods, stockouts, and excessive inventory. In addition, the information on the inventory management interface is fragmented and the data visualization is insufficient, making it difficult for managers to intuitively understand the patterns of SKU inventory changes and potential risks.

[0005] Therefore, there is an urgent need for an inventory health assessment and optimization method based on LSTM prediction and state-driven mechanism, which introduces the Long Short-Term Memory Neural Network (LSTM) model and lifecycle recognition mechanism to achieve intelligent prediction, visual analysis and refined management of e-commerce inventory. Summary of the Invention

[0006] This invention provides an inventory health assessment and optimization method based on LSTM prediction and state-driven mechanism. Its purpose is to achieve high-precision prediction of SKU-level inventory trends, automatic division of product life cycle, and real-time early warning of slow-moving / out-of-stock risks through deep learning time series prediction model and dynamic rule engine, thereby improving the efficiency and accuracy of inventory decision-making.

[0007] To achieve the above objectives, this invention provides an inventory health assessment and optimization method based on LSTM prediction and a state-driven mechanism, specifically including the following steps: S1. Collect data on SKU entry, sales and inventory for each product code, and store the data in a time series according to SKU granularity to obtain primary time series data; S2. Preprocess the primary time series data to obtain a standardized time series matrix; S3. Input the standardized time series matrix into the Long Short-Term Memory Neural Network Model (LSTM) to predict the inventory change trend and obtain the stable inventory range. S4. Automatically divide the product life cycle stages by combining inventory change trends and sales trends, including the introduction stage, growth stage, maturity stage, and decline stage. S5. Calculate the Inventory Health Index (IHI) and dynamically adjust IHI through a state-driven mechanism. S6. When the inventory status or IHI value changes abnormally, the LSTM model is retrained and the strategy is adjusted to achieve closed-loop optimization of "prediction - evaluation - status update - re-prediction". S7. Based on the stable inventory range and IHI value, set multiple thresholds to trigger multi-level early warning signals for three types of risks: stockout, excessive inventory, and slow-moving goods, and display them dynamically on the visualization interface. The risk level is determined based on the degree of deviation from the status and the IHI value, including three levels of response: potential risk, medium risk and high risk, and each level of response is automatically associated with the corresponding optimization action.

[0008] Preferably, the preprocessing described in S2 includes: data cleaning, feature extraction, and inventory status encoding generation; The data cleaning includes: missing value imputation, outlier detection, time alignment, and data smoothing; The features extracted include sales volatility, inventory change rate, in-transit inventory ratio, and lifecycle characteristics. The inventory status codes include inventory health codes, inventory backlog codes, and inventory shortage codes.

[0009] Preferably, the LSTM model described in S3 adopts a sliding time window dynamic update mechanism and automatically triggers the model retraining process each time a state transition or risk level changes. The prediction results of the LSTM model are used as inputs for state recognition and health assessment.

[0010] Preferably, the product life cycle stage division described in S4 is based on a combination of sales growth rate and inventory change rate. The period when inventory continues to rise while sales are low is considered the introduction period. When the sales growth rate continues to rise, it is considered to be in the growth stage; The mature stage is defined as when both inventory volatility and sales growth rate are in a stable range. A period of decline is defined as when sales decrease and inventory exceeds the upper limit of stable inventory. The system can dynamically adjust the stage boundaries based on real-time forecast results and changes in inventory health.

[0011] Preferably, the specific content of S5 includes: S51. Calculate the dimensions of the IHI value, including inventory turnover rate, inventory change rate, life cycle stage, sales trend, in-transit ratio, and status transition frequency. S52. Based on industry benchmarks and historical data, set the initial calculation weights for each dimension of indicators; S53. Using a multi-indicator fusion algorithm, the scores of each dimension indicator are weighted and summarized to obtain the initial value of IHI; S54. Through a state-driven mechanism, the inventory status is monitored in real time, and the calculation weights of each dimension indicator are dynamically adjusted according to the changes in the inventory status to recalculate the IHI value; at the same time, the IHI value is used as an auxiliary basis for determining the inventory status, realizing a two-way correlation between the two.

[0012] Preferably, the specific working content of the state-driven mechanism includes: S541. Real-time monitoring of inventory status and IHI value changes; S542. If the state changes significantly, adjust the LSTM model parameters and re-execute the prediction process. S543. Feed back the updated prediction results to S4 and S7 to achieve synchronous updates of prediction correction and risk level.

[0013] Preferably, in S542, the cases where the state change is significant include: S5421, Inventory Health Index (IHI) decreased by more than 10% compared to the previous calculation period; S5422. The inventory status is judged as "out of stock risk" or "overstock risk" for 3 consecutive days; S5423. After the system determines that it has entered the "decline period", the actual inventory is higher than 10% of the upper limit of the stable inventory range for two consecutive days.

[0014] Preferably, in S6, the output of the visualization interface includes historical inventory curves, predicted inventory curves, life cycle stage background areas, state transition trajectories, and inventory health index change graphs. The warning signals are categorized by different warning symbols and graded by color on the visual interface.

[0015] The inventory health assessment and optimization method based on LSTM prediction and state-driven mechanism provided by this invention can be used in a system, specifically including the following modules: Data acquisition module: Collects data on SKU inbound, sales and inventory, and stores it in time series according to SKU granularity; Data processing module: preprocesses the stored data, including cleaning, reconstruction, and feature extraction; Inventory forecasting module: Uses an LSTM model to perform time series forecasting of inventory and obtains a stable inventory range; Lifecycle segmentation module: Automatically segments the lifecycle based on inventory change rate and sales trend; Indicator calculation module: Calculates inventory turnover rate, capital ratio, sales revenue, and inventory health index (IHI); Status-driven optimization module: Real-time monitoring of inventory status and IHI value changes, with an operation status-driven mechanism; Risk warning module: Identifies three types of risks, including slow sales, stockouts, and excessive inventory; Visualization module: Displays forecast results, lifecycle range, inventory health status, and stable inventory range.

[0016] Preferably, the risk warning module works in conjunction with a state-driven mechanism, setting multiple threshold levels based on prediction results and IHI values; A stockout warning is triggered when inventory declines continuously and the predicted inventory falls below the stable inventory lower limit. An overstock warning is triggered when inventory rises continuously and exceeds the stable inventory limit. A slow-moving-stock warning is triggered when sales are below 50% of the historical average for three consecutive periods. The system generates three levels of response—potential risk, moderate risk, and high risk—based on the IHI value and the trend of status changes, and automatically associates corresponding optimization actions. The early warning results are displayed in the form of color-coded indicators, warning pop-ups, and strategy suggestions, and are updated to the visualization interface in real time.

[0017] Therefore, the inventory health assessment and optimization method based on LSTM prediction and state-driven mechanism adopted in this invention has the following advantages compared with the prior art: (1) By integrating LSTM neural network and state-driven bidirectional coupling mechanism, the prediction model incorporates multi-dimensional features such as inventory status encoding, promotional events, and in-transit quantity. Through dynamic training with a 30-day sliding time window and retraining triggered by state transition, the nonlinear change law of inventory is accurately captured. At the same time, the exponential moving average algorithm is combined to generate a stable inventory range. Compared with the traditional linear model, it effectively reduces the risk of stockouts or backlogs, and significantly improves the prediction accuracy and market adaptability. (2) On the one hand, by automatically dividing the product introduction period, growth period, maturity period and decline period, differentiated inventory strategies are formulated for different life cycle stages; on the other hand, a multi-dimensional inventory health index (IHI) is constructed, which comprehensively considers core indicators such as inventory turnover rate, sales trend and state transition frequency, and realizes quantitative assessment of inventory status through dynamic weight correction, breaking through the limitations of traditional single threshold judgment, and allowing inventory management to shift from experience-driven to data-driven. (3) Construct a three-level risk warning system based on stable inventory range and IHI value to identify three types of risks in advance: stockout, excessive inventory, and slow sales, and associate them with precise optimization actions (replenishment suggestions, promotion plans, and procurement suspension); at the same time, integrate multi-dimensional information such as forecast curves, life cycle stages, and state transition trajectories through a visual dashboard to achieve a closed loop of the whole process of "prediction-evaluation-early warning-optimization", which greatly reduces manual intervention and improves the efficiency of inventory decision-making and the timeliness of risk response.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a flowchart of an embodiment of the LSTM and state-driven inventory health management system of the present invention; Figure 2 This is a schematic diagram of the risk warning module in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] Example 1 like Figure 1 As shown, the LSTM-based and state-driven inventory health management system is architected and deployed according to the following modules: (1) Data acquisition module: Through API interface, it connects to the platform's ERP system, WMS warehouse management system, POS sales front-end system and promotion management system to collect in-warehouse records (including in-warehouse time, quantity and batch), sales records (including order time, sales volume and average order value), existing inventory records (including warehouse location and inventory quantity), in-transit inventory information (including logistics transportation status and estimated arrival time) and promotion event information (including promotion start and end time, discount and activity type) of each SKU. The data is serialized and stored in 1-day time unit at SKU granularity to form primary time series data to ensure the consistency of data time dimension.

[0022] (2) Data processing module: Performs preprocessing on the collected primary time-series data: Data cleaning includes steps such as supplementing missing sales / inventory data, removing outliers (e.g., negative sales caused by sudden large-scale returns, or abnormal inventory values ​​caused by data entry errors), aligning data from different systems by time, and periodically resampling.

[0023] Feature extraction includes sales volume change rate, inventory volatility, sales seasonality coefficient, inventory turnover rate, in-transit inventory ratio, lifecycle labels and inventory status codes, etc. After feature extraction, the system generates a standardized time series matrix, which is used as input data for the LSTM model.

[0024] Inventory status coding: Based on the state-driven mechanism, inventory health code, inventory backlog code, and inventory shortage code are generated and stored synchronously in the input feature set of the subsequent LSTM model.

[0025] (3) Inventory Forecasting Module: An LSTM neural network model is built, consisting of an input layer, two LSTM hidden layers, and a fully connected output layer. The ReLU activation function and Adam optimizer are used. A sliding window dynamic training and update mechanism is employed. Input features include sales volume, inbound volume, inventory volume, in-transit volume, promotional events (0 = no promotion, 1 = regular promotion, 2 = major promotion), and inventory status codes. The output is a 15-period (15-day) inventory forecast sequence. A stable inventory range is calculated using an exponential moving average smoothing function. The upper limit of the range is the predicted value plus one standard deviation, and the lower limit is the predicted value minus one standard deviation. The Adam optimizer (learning rate 0.001) automatically triggers a retraining process each time a state transition or risk level change occurs.

[0026] (4) Lifecycle Segmentation Module: Based on a combination of sales growth rate and inventory change rate, the module automatically segments the product lifecycle and dynamically adjusts the segment boundaries according to real-time forecast results and inventory health status. Introduction stage: Inventory continues to rise (inventory change rate > 0 for 7 consecutive days) and sales are low; Growth stage: Sales growth rate continues to rise; Maturity stage: Inventory volatility < 10% and sales growth rate fluctuation range < 5%; Decline stage: Sales decline (sales growth rate < 0 for 7 consecutive days) and inventory is higher than the stable inventory limit.

[0027] (5) Indicator Calculation Module: Real-time calculation of inventory turnover rate, capital ratio, sales volume and inventory health index (IHI) dimension indicators: inventory turnover rate, inventory change rate, life cycle stage, sales trend, in-transit ratio and status transition frequency; the initial weights are calibrated based on the e-commerce industry benchmark data and the platform's historical 3-year data, and are dynamically adjusted through the status-driven mechanism (e.g., when the inventory status changes from healthy to backlog, the weight of inventory change rate is increased and the weight of life cycle stage is decreased).

[0028] (6) Status-driven optimization module: Real-time monitoring of inventory status and IHI value changes. When a significant change in status is detected (transition from healthy to overstock, healthy to out-of-stock or decline stage), the LSTM model parameters are adjusted and the prediction process is re-executed. The updated prediction results are fed back to the life cycle division module and risk warning module to realize the synchronous update of prediction correction and risk level.

[0029] (7) Risk warning module: Works under a joint state-driven mechanism, setting multi-level thresholds based on prediction results and IHI values, such as... Figure 2 As shown.

[0030] Stockout warning: Inventory has declined for 5 consecutive days and the predicted inventory is below the lower limit of the stable inventory range. IHI < 60 (out of 100) triggers high risk, 60 ≤ IHI < 75 triggers medium risk, and 75 ≤ IHI < 85 triggers potential risk.

[0031] Overstock warning: Inventory increases for 7 consecutive days and the actual inventory is higher than the stable inventory limit. IHI < 60 triggers high risk, 60 ≤ IHI < 75 triggers medium risk, and 75 ≤ IHI < 85 triggers potential risk. Slow-moving inventory warning: Sales volume is lower than 50% of the historical average sales volume for 3 consecutive cycles (9 days). IHI < 60 triggers high risk, 60 ≤ IHI < 75 triggers medium risk, and 75 ≤ IHI < 85 triggers potential risk.

[0032] The system automatically associates and optimizes actions: high-risk actions trigger emergency replenishment / promotional clearance / purchase suspension orders, medium-risk actions push strategy suggestions, and potential risks generate monitoring alerts.

[0033] (8) Visualization module: Generates SKU-level inventory management visualization dashboard, outputting historical inventory curves, predicted inventory curves, life cycle stage background areas (different stages are marked with different colors), state transition trajectory (state change nodes), inventory health index change chart and key performance indicators (inventory turnover rate, capital ratio, sales); warning signals are graded by red (high risk), yellow (medium risk), and green (potential risk / no risk), and are displayed in real time with warning pop-ups and strategy suggestion panels.

[0034] Example 2 In this embodiment, a product is selected as the test object to apply the system.

[0035] Complete time-series data for nearly 180 days were collected, with the first 165 days used for model training and validation, and the last 15 days used to validate prediction accuracy and system response. This SKU went through three phases during the testing period: new product launch (introduction phase), sales growth (growth phase), and stable sales (maturity phase), including two regular promotions and one major sales event.

[0036] The data processing module cleans and extracts features from 180 days of primary time-series data, fills in missing data, removes outliers, generates a standardized time-series matrix, and assigns inventory status codes to daily data.

[0037] The LSTM model was trained based on the data from the previous 165 days to predict the inventory trend for the next 15 days. The prediction results showed that the inventory would drop to the lower limit of the stable range, then rebound and remain within the stable range. The actual inventory data verification showed that the average error between the predicted value and the actual value was small.

[0038] The system automatically identifies the product SKU's lifecycle stage, including the introduction, growth, maturity, and decline stages.

[0039] The system calculates the Inventory Health Index (IHI) using the following formula: ; in: For inventory turnover rate, This represents the sales trend coefficient. Weights for lifecycle stages This is the inventory volatility coefficient. For state stability, For dynamic weights.

[0040] When the system detects that the inventory of a SKU has decreased for three consecutive days and the IHI has dropped to 78, triggering a potential stockout risk, the state-driven mechanism automatically adjusts the LSTM model parameters (increasing the weight of the in-transit quantity feature), re-predicts and outputs the inventory forecast value, and feeds it back to the lifecycle segmentation module (maintaining the mature stage) and the risk warning module (upgrading to a moderate stockout risk).

[0041] When the system detects that the inventory of a SKU has decreased for 5 consecutive days and the predicted inventory is below the lower limit of the stable range (IHI=72), a moderate stockout risk is triggered. The system will automatically push replenishment suggestions (suggested replenishment amount = upper limit of stable inventory range - current inventory + predicted sales for the next 5 days) and remind inventory managers through a pop-up window in the visual interface.

[0042] The content that managers can intuitively view through the visual dashboard includes: Historical inventory curves show a slow rise during the introduction phase, rapid growth during the growth phase, and stable fluctuations during the maturity phase. Predict the downward and upward trends of the inventory curve; The state transition trajectory shows the transition between healthy and backlogged states; When the IHI change chart shows 60≤IHI<75, a warning sign is highlighted in yellow to indicate a moderate stockout risk, along with replenishment strategy suggestions.

[0043] Managers initiate replenishment based on system suggestions. When goods are received and inventory recovers, the system automatically triggers LSTM model retraining. After updating the prediction parameters, it predicts again that the inventory will remain within a stable range for the next 5 days, the IHI will rise to 85, the risk warning will be lifted, and the closed-loop optimization of "prediction - evaluation - status update - re-prediction" is achieved.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An inventory health assessment and optimization method based on LSTM prediction and state-driven mechanism, characterized in that, Specifically, it includes: S1. Collect data on SKU entry, sales and inventory for each product code, and store the data in a time series according to SKU granularity to obtain primary time series data; S2. Preprocess the primary time series data to obtain a standardized time series matrix; S3. Input the standardized time series matrix into the Long Short-Term Memory Neural Network Model (LSTM) to predict the inventory change trend and obtain the stable inventory range. S4. Automatically divide the product life cycle stages by combining inventory change trends and sales trends, including the introduction stage, growth stage, maturity stage, and decline stage. S5. Calculate the Inventory Health Index (IHI) and dynamically adjust IHI through a state-driven mechanism. S6. When the inventory status or IHI value changes abnormally, the LSTM model is retrained and the strategy is adjusted to achieve closed-loop optimization of "prediction - evaluation - status update - re-prediction". S7. Based on the stable inventory range and IHI value, set multiple thresholds to trigger multi-level early warning signals for three types of risks: stockout, excessive inventory, and slow-moving goods, and display them dynamically on the visualization interface. The risk level is determined based on the degree of deviation from the state and the IHI value, including three levels of response: potential risk, medium risk and high risk, and each level of response is automatically associated with the corresponding optimization action.

2. The inventory health assessment and optimization method based on LSTM prediction and state-driven mechanism according to claim 1, characterized in that, The specific content of the preprocessing described in S2 includes: data cleaning, feature extraction, and inventory status encoding generation; The data cleaning includes: missing value imputation, outlier detection, time alignment, and data smoothing; The features extracted include sales volatility, inventory change rate, in-transit inventory ratio, and lifecycle characteristics. The inventory status codes include inventory health codes, inventory backlog codes, and inventory shortage codes.

3. The inventory health assessment and optimization method based on LSTM prediction and state-driven mechanism according to claim 1, characterized in that, The LSTM model described in S3 employs a sliding time window dynamic update mechanism and automatically triggers the model retraining process each time a state transition or risk level changes. The prediction results of the LSTM model are simultaneously used as inputs for state recognition and health assessment.

4. The inventory health assessment and optimization method based on LSTM prediction and state-driven mechanism according to claim 1, characterized in that, The product life cycle stages described in S4 are based on a combination of sales growth rate and inventory change rate. The period when inventory continues to rise while sales are low is considered the introduction period. When the sales growth rate continues to rise, it is considered to be in the growth stage; The mature stage is defined as when both inventory volatility and sales growth rate are in a stable range. A period of decline is defined as when sales decrease and inventory exceeds the upper limit of stable inventory. The system can dynamically adjust the stage boundaries based on real-time forecast results and changes in inventory health.

5. The inventory health assessment and optimization method based on LSTM prediction and state-driven mechanism according to claim 1, characterized in that, The specific content of S5 includes: S51. Calculate the dimensions of the IHI value, including inventory turnover rate, inventory change rate, life cycle stage, sales trend, in-transit ratio, and status transition frequency. S52. Based on industry benchmarks and historical data, set the initial calculation weights for each dimension of indicators; S53. Using a multi-indicator fusion algorithm, the scores of each dimension indicator are weighted and summarized to obtain the initial value of IHI; S54. Through a state-driven mechanism, the inventory status is monitored in real time, and the calculation weights of each dimension indicator are dynamically adjusted according to the changes in the inventory status to recalculate the IHI value; at the same time, the IHI value is used as an auxiliary basis for determining the inventory status, realizing a two-way correlation between the two.

6. The inventory health assessment and optimization method based on LSTM prediction and state-driven mechanism according to claim 5, characterized in that, The specific working content of the state-driven mechanism includes: S541. Real-time monitoring of inventory status and IHI value changes; S542. If the state changes significantly, adjust the LSTM model parameters and re-execute the prediction process. S543. Feed back the updated prediction results to S4 and S7 to achieve synchronous updates of prediction correction and risk level.

7. The inventory health assessment and optimization method based on LSTM prediction and state-driven mechanism according to claim 6, characterized in that, In S542, the cases where the state change is significant include: S5421, Inventory Health Index (IHI) decreased by more than 10% compared to the previous calculation period; S5422. Inventory status is judged as "out of stock risk" or "overstock risk" for 3 consecutive days; S5423. After the system determines that it has entered the "decline period", the actual inventory is higher than 10% of the upper limit of the stable inventory range for two consecutive days.

8. The inventory health assessment and optimization method based on LSTM prediction and state-driven mechanism according to claim 5, characterized in that, S5 also includes setting multi-level thresholds based on prediction results and IHI values; A stockout warning is triggered when inventory declines continuously and the predicted inventory falls below the stable inventory lower limit. An overstock warning is triggered when inventory rises continuously and exceeds the stable inventory limit. A slow-moving-stock warning is triggered when sales are below 50% of the historical average for three consecutive periods. The system generates three levels of response—potential risk, moderate risk, and high risk—based on the IHI value and the trend of status changes, and automatically associates corresponding optimization actions. The early warning results are displayed in the form of color-coded indicators, warning pop-ups, and strategy suggestions, and are updated to the visualization interface in real time.

Images