Inventory prediction system based on multi-source data fusion

Abstract

The invention relates to the technical field of data fusion, and discloses an inventory prediction system based on multi-source data fusion. An inventory data acquisition module of the system is responsible for acquiring multi-dimensional dynamic data including a real-time sales fluctuation sequence, a storage turnover period record, a logistics transportation delay index and the like in a supply chain link; the feature fusion processing module extracts sales trend, storage efficiency and logistics stability features according to the multi-dimensional dynamic data, and generates a fusion feature matrix according to time axis alignment; the inventory state modeling module constructs a dynamic inventory state model capable of reflecting the incidence relation between the inventory consumption rate and the replenishment delay by using the fusion feature matrix; the anomaly detection module carries out deviation degree analysis on an inventory change curve output by the dynamic inventory state model, identifies abnormal fluctuation nodes and marks corresponding supply chain links; and the prediction result generation module outputs an inventory surplus early warning value and stockout risk probability distribution in a future period based on the corrected dynamic inventory state model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data fusion, in particular to an inventory prediction system based on multi-source data fusion. BACKGROUND

[0002] In the operation of the power industry, inventory management is directly related to power supply reliability and operation and maintenance efficiency, especially the inventory level of power materials (such as transformer spare parts, cables, switch cabinets, and repair tools) is crucial to cope with peak electricity consumption, equipment failure, and power grid upgrading scenarios. If the inventory is insufficient, it will cause delay in fault repair, affect normal electricity consumption of residents and enterprises, and even cause large-scale power outages. If the inventory is overstocked, it will occupy a large amount of funds, increase warehouse maintenance costs, and reduce asset turnover.

[0003] Traditional inventory prediction methods in the power industry have significant defects, relying too much on a single data source, focusing on historical inventory consumption data or simple electricity load statistics, and not fully integrating information from all links of the supply chain. For example, the influence of real-time electricity load fluctuations (such as the surge in air conditioning load in summer and the increase in heating load in winter) on spare parts demand is not considered, and key factors such as warehouse turnover cycle (such as the storage validity period of different materials, warehouse space limitations) and logistics transportation delays (such as transportation conditions in remote areas, and material transportation blocked by extreme weather) are ignored, resulting in one-sided prediction dimensions. The prediction logic lacks dynamic adaptability and cannot cope with the unique fluctuation scenarios in the power industry, such as sudden repair demand caused by extreme weather (typhoons, cold waves), and changes in material demand caused by temporary adjustments in power grid renovation projects. Traditional methods are difficult to quickly capture these dynamic information, resulting in a large deviation between predicted results and actual demand.

[0004] Taking a power supply company as an example, during the extremely hot weather in summer 2023, the company predicted the inventory of transformer spare parts based only on historical summer electricity data, without considering the real-time surge in electricity load and the risk of logistics transportation delay (high temperatures causing some roads to be restricted, extending the material transportation cycle by 30%). When a transformer in a certain area failed, the corresponding replacement spare parts inventory was insufficient, and the repair team had to wait for 3 days to obtain the spare parts, during which time 12,000 households in the area were cut off from electricity, resulting in a large number of user complaints and penalties for violating the power supply reliability assessment standards, with direct economic losses exceeding 500,000 yuan. At the same time, some low-voltage switch cabinets in the company had a 6-month overstocked inventory due to the delay in the power grid renovation project, occupying about 8 million yuan in funds, causing waste of funds and shortage of warehouse resources.

[0005] The root of such problems lies in that the traditional inventory forecasting method fails to realize effective fusion of multi-link data of a supply chain, cannot construct a forecasting model reflecting the correlation between power load, warehouse efficiency, logistics stability and inventory consumption, is difficult to accurately predict the real demand and supply risk of power materials, and urgently needs an inventory forecasting scheme capable of integrating multi-source data and dynamically adapting to the power industry scene to solve the above inventory management pain points. SUMMARY

[0006] The present application aims to provide an inventory forecasting system based on multi-source data fusion to solve the problems raised in the background art.

[0007] To achieve the above-mentioned purpose, the present application provides an inventory forecasting system based on multi-source data fusion, which comprises: An inventory data acquisition module acquires multi-dimensional dynamic data in the supply chain link, wherein the multi-dimensional dynamic data comprises real-time sales fluctuation sequence, warehouse turnover period record and logistics transportation delay index; A feature fusion processing module extracts sales trend features, warehouse efficiency features and logistics stability features based on the multi-dimensional dynamic data, and generates a fusion feature matrix after aligning the three types of features according to the time axis; An inventory state modeling module constructs a dynamic inventory state model reflecting the correlation between inventory consumption rate and replenishment delay according to the fusion feature matrix; An anomaly detection module analyzes the deviation degree of the inventory change curve output by the dynamic inventory state model, identifies abnormal fluctuation nodes and marks the corresponding supply chain link; A prediction result generation module outputs inventory surplus early warning values and out-of-stock risk probability distribution in future periods based on the corrected dynamic inventory state model.

[0008] Preferably, the inventory data acquisition module comprises: A sales data acquisition sub-module extracts hourly updated product category sales data from a terminal sales system, and calculates the fluctuation percentage of each category sales relative to the historical same period data; A warehouse data integration sub-module acquires the warehousing timestamp and warehousing time interval recorded in the warehouse management system, and calculates the average cargo residence time in a single warehouse period; A logistics data analysis sub-module receives GPS trajectory and planned arrival time deviation data of the transportation link, and generates a time-effect decay coefficient reflecting regional-level transportation delay.

[0009] Preferably, the feature fusion processing module comprises: A sales trend analysis sub-module performs seasonal decomposition on the real-time sales fluctuation sequence, extracts the linear growth component and residual component after removing the periodic influence; The warehouse efficiency calculation submodule calculates a warehouse throughput efficiency value per unit time according to the warehouse turnover period record and in combination with a goods category weight; The logistics feature mapping submodule clusters the logistics transportation delay indicators according to transportation routes to generate a route stability score and a sudden delay event frequency; The matrix generation submodule reorganizes the linear growth component, the warehouse throughput efficiency value, and the route stability score into a three-dimensional feature tensor according to a uniform time granularity.

[0010] Preferably, the inventory state modeling module comprises: The consumption rate calculation submodule derives a daily theoretical inventory consumption amount based on the sales trend feature in the fused feature matrix and in combination with a commodity shelf life constraint condition; The restocking delay compensation submodule corrects a restocking period promised by a supplier according to a logistics stability feature to generate a probability distribution of actual arrival time; The dynamic balance submodule establishes a time-varying differential equation between the consumption amount and the restocking amount, wherein the restocking delay probability serves as a damping coefficient of the equation.

[0011] Preferably, the anomaly detection module comprises: The baseline generation submodule calculates a mean and a variance of a historical inventory change curve by using a sliding window method, as a baseline range for current detection; The deviation calculation submodule compares a difference between a dynamic inventory state model output value and the baseline range in real time, and triggers an anomaly flag when a continuously over-threshold value duration reaches a preset period; The link tracing submodule associates an abnormal fluctuation node with an original data dimension in the feature fusion processing module to locate a sales, warehouse, or logistics link that causes the anomaly.

[0012] Preferably, the prediction result generation module comprises: The multi-scenario simulation submodule generates multiple simulation paths of inventory consumption and restocking based on different optimization hypotheses of supply chain links; The risk quantification submodule calculates a frequency of an occurrence of an inventory level lower than a safety threshold under each simulation path to output a stock-out risk heat map; The early warning generation submodule maps a region exceeding a critical value in the heat map to a specific commodity category and a time node to form an early warning event list.

[0013] Preferably, the multi-scenario simulation submodule comprises: A sales fluctuation scenario library stores feature parameters of historical extreme sales fluctuation events to generate test cases for sudden sales growth; A logistics interruption simulator constructs a cascading effect model of transportation delay according to regional weather data and traffic control records; The warehouse fault injector simulates the impact of equipment downtime and human operation errors on warehouse turnover efficiency.

[0014] Preferably, the predicted result generation module further comprises: a prediction confidence evaluation submodule for calculating the uncertainty interval of the out-of-stock risk probability distribution; The prediction confidence evaluation submodule assigns a confidence score to each prediction period based on the data integrity of the fusion feature matrix and the historical prediction accuracy of the dynamic inventory state model.

[0015] Preferably, the time-varying differential equation established by the dynamic balance submodule is further simulated and optimized through digital twinning technology: The system constructs a virtual inventory model corresponding to the physical supply chain; The virtual inventory model receives real-time updates of the multi-dimensional dynamic data and runs the time-varying differential equation to simulate changes in inventory state under different strategies, thereby verifying the effectiveness of the replenishment strategy before the physical inventory actually changes.

[0016] Preferably, the anomaly detection module further comprises: an adaptive learning submodule for dynamically adjusting the detection threshold used in the deviation degree calculation submodule; The adaptive learning submodule analyzes the subsequent impact of the labeled abnormal fluctuation nodes and the false positive records, and periodically optimizes the detection threshold based on the analysis results using a reinforcement learning algorithm to reduce false positives and improve sensitivity to real anomalies.

[0017] Compared with the prior art, the beneficial effects of the present application are: The inventory data acquisition module of the system can widely collect multi-dimensional dynamic data in various links of the supply chain. It covers real-time sales fluctuation sequences, which are like the "sales pulse" of the enterprise, allowing the enterprise to accurately grasp the immediate demand changes in the market. Through monitoring of real-time sales fluctuation sequences, the enterprise can quickly know when the sales peak will arrive and make preparations in advance. For example, during the promotion activities of e-commerce platforms, the system can capture the sharp increase in sales data in real time, and the enterprise can timely increase the inventory of popular goods to meet the purchasing needs of consumers and avoid missing sales opportunities due to lack of goods.

[0018] The warehouse turnover cycle record provides key information on warehouse operation efficiency for the enterprise. The enterprise can clearly understand the residence time of goods in the warehouse, the frequency of warehouse entry and exit, etc., so as to reasonably plan the warehouse space and optimize the warehouse layout. If it is found that the warehouse turnover cycle of some goods is too long, the enterprise can timely adjust the sales strategy, increase the promotion intensity, speed up the inventory turnover, and reduce the warehouse cost.

[0019] The logistics transportation delay indicator is also not to be ignored. It enables the enterprise to track the logistics transportation status in real time, and once there is a logistics delay, the enterprise can quickly respond. For example, when it is learned that a batch of goods may be delayed due to traffic congestion or other reasons, the enterprise can communicate with the customer in advance, explain the situation and provide a solution, while adjusting the production and sales plan to reduce the impact of logistics delay on the operation of the enterprise. Through the comprehensive collection of these multi-dimensional dynamic data, the enterprise can master the real situation of each link of the supply chain in real time, providing a solid data foundation for subsequent decision-making.

[0020] The feature fusion processing module and the inventory state modeling module cooperate with each other to provide strong support for accurate prediction of inventory changes. The feature fusion processing module extracts sales trend features, warehouse efficiency features, and logistics stability features based on multi-dimensional dynamic data. These features reflect the operation status of the supply chain from different angles, and by aligning them on the time axis to generate a fusion feature matrix, the dynamic changes of the supply chain can be more comprehensively and accurately described.

[0021] The inventory state modeling module constructs a dynamic inventory state model reflecting the correlation between inventory consumption rate and replenishment delay based on the fusion feature matrix. This model fully considers the influence of various factors on inventory and can more accurately predict the trend of inventory changes. For example, in the clothing industry, seasonal changes have a significant impact on sales. Through this model, the enterprise can accurately predict the inventory demand of different styles of clothing in different seasons by combining historical sales data, seasonal factors, and current market trends, and make replenishment plans in advance.

[0022] Accurate inventory prediction can help the enterprise plan inventory levels in advance, avoiding out-of-stock or overstock situations. When it is predicted that the demand for a certain product will increase significantly, the enterprise can increase inventory in advance to ensure market demand and improve customer satisfaction; when it is predicted that the demand for a certain product will decrease, the enterprise can reduce inventory in time to avoid overstock and reduce inventory costs. This not only helps the enterprise optimize inventory management, but also improves the efficiency of the use of funds, making the operation of the enterprise more stable. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 Comparison chart of multi-source data feature influence degree and correlation; Figure 2 Workflow diagram of the inventory data collection module; Figure 3 Workflow diagram of the feature fusion processing module; Figure 4 Multi-source data fusion feature matrix heat map. DETAILED DESCRIPTION

[0024] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described, obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0025] Please refer to Figure 1 The present application provides a kind of inventory prediction system based on multi-source data fusion, the system includes inventory data acquisition module is responsible for obtaining multidimensional dynamic data from supply chain link, these data include real-time sales fluctuation sequence, warehouse turnover period record and logistics transport delay index.Real-time sales fluctuation sequence reflects the immediate change situation of commodity sales, warehouse turnover period record captures the residence time of goods in warehouse, and logistics transport delay index measures the transport efficiency of goods from supplier to warehouse.Feature fusion processing module processes multidimensional dynamic data, extracts sales trend feature, warehouse efficiency feature and logistics stability feature.Sales trend feature identifies long-term growth or decline pattern from sales data, warehouse efficiency feature evaluates the throughput capacity of warehouse operation, and logistics stability feature analyzes the reliability of transport link.These features are aligned according to time axis, eliminate the problem of time stamp inconsistency, then combined into fusion feature matrix.Inventory state modeling module constructs dynamic inventory state model using fusion feature matrix, and dynamic inventory state model describes the correlation between inventory consumption rate and restocking delay through mathematical relationship.Dynamic inventory state model can simulate the change of inventory level with time, considering the comprehensive influence of sales fluctuation, warehouse efficiency and logistics delay.Abnormal detection module monitors the inventory change curve output by dynamic inventory state model, identifies abnormal fluctuation node through deviation analysis.Deviation analysis compares the difference between current inventory curve and historical benchmark, when the difference exceeds threshold, mark abnormal, and trace back to specific supply chain link.Prediction result generation module generates inventory prediction result in future period based on the corrected dynamic inventory state model.Prediction result includes inventory surplus early warning value and out-of-stock risk probability distribution, and inventory surplus early warning value indicates possible time point of inventory surplus, and out-of-stock risk probability distribution quantifies the possibility of inventory shortage.

[0026] Embodiment 1: refer to Figure 2The inventory data collection module is composed of a sales data collection submodule, a warehouse data integration submodule, and a logistics data analysis submodule. The function of the sales data collection submodule is to obtain real-time sales fluctuation sequence. The sales data collection submodule is programmatically connected with the sales system distributed in the retail terminal and online platform. The transaction records generated by these sales systems contain the key fields of product category identification, sales timestamp, and sales quantity. The sales data collection submodule actively pulls incremental data at a cycle of every hour through a pre-defined data interface. The data interface usually takes the form of an application programming interface, uses the Hypertext Transfer Protocol Secure protocol for data transmission, and the data format is Structured Data Markup Language or Lightweight Data Exchange Format. Incremental data refers to the sales records that are newly added or updated after the last collection cycle. The sales data collection submodule maintains a data buffer in memory for temporary storage of raw data. The calculation logic of the sales data collection submodule involves the derivation of fluctuation percentage, which is a quantitative indicator to measure the degree of deviation of current sales from the historical same period level. Historical same period data is selected from the sales records of the same time window in the past, for example, the sales data of the same hour on the same working day last week. The calculation process uses the current hour sales minus the historical same period sales to get a difference, divides the difference by the historical same period sales and multiplies by one hundred to convert it to percentage form. The calculation process needs to consider data integrity, and for the case of missing historical data, interpolation method or using earlier period data as a substitute. The sales data collection submodule independently performs calculation for each product category, generating a real-time sales fluctuation sequence sorted by time. The real-time sales fluctuation sequence is stored in a time series database, and each data point contains timestamp, category code, and fluctuation value attributes.

[0027] The warehouse data integration submodule is responsible for processing warehouse turnover period records. The input of the warehouse data integration submodule comes from the warehouse management system. The warehouse management system records the movement of goods through automatic identification technology, which includes radio frequency identification scanning and two-dimensional code scanning. The warehouse data integration submodule accesses the database view of the warehouse management system to obtain the in-warehouse timestamp and out-warehouse timestamp. The in-warehouse timestamp records the time when the goods are accepted into the warehouse storage area, and the out-warehouse timestamp records the time when the goods are picked and leave the warehouse. The accuracy of the timestamp reaches the minute level. The warehouse data integration submodule calculates the residence duration of a single good by subtracting the in-warehouse timestamp from the out-warehouse timestamp. The calculation result is expressed in hours or days. The warehouse data integration submodule calculates the average residence duration of all goods that have completed the out-warehouse in a statistical period. The statistical basis for the average residence duration is the complete list of goods in a single warehouse period, which is defined as a fixed time interval, such as twenty-four hours. The warehouse data integration submodule aggregates all the residence durations of the goods and calculates the arithmetic mean. For special goods such as large items or hazardous goods, a weight coefficient can be set. The warehouse turnover period record also contains goods category information. The warehouse data integration submodule can group the average residence duration by goods category. The warehouse data integration submodule organizes the processing results into structured records, each record containing the statistical period start time, statistical period end time, goods category, and average residence duration fields. These records are persisted to the designated table of the relational database for subsequent calling by the feature fusion processing module.

[0028] The logistics data analysis submodule focuses on processing logistics transportation delay indicators. The logistics data analysis submodule establishes a data channel with the information system of the logistics service provider. The data channel transmits global positioning system trajectory data and planned arrival time deviation data. The global positioning system trajectory data comes from the vehicle terminal of the transportation vehicle, which reports latitude and longitude coordinates and speed information at a fixed frequency. The logistics data analysis submodule analyzes the trajectory data to reconstruct the transportation path, compares the actual path with the planned path to identify deviations. The planned arrival time deviation data is directly obtained from the status update of the logistics tracking system. The logistics data analysis submodule calculates the difference between the actual arrival time and the contract time. The deviation data is classified by transportation route, and each transportation route has a unique route code identifier. The logistics data analysis submodule generates a time decay coefficient, which is a standardized value reflecting the degree of transportation delay in a specific area. Calculating the time decay coefficient requires aggregating deviation data from multiple transportation routes. The logistics data analysis submodule calculates the average deviation time by grouping by destination area. Comparing the average deviation time with the standard transportation time gives the delay ratio, and the time decay coefficient is a normalized representation of the delay ratio, with a range of zero to one. A time decay coefficient close to one indicates severe delays, while a value close to zero indicates on-time transportation. The logistics data analysis submodule updates the time decay coefficient regularly, with an update frequency consistent with the logistics data reporting frequency. The time decay coefficient is stored in association with route information and time information, forming a time series of logistics transportation delay indicators.

[0029] The inventory data acquisition module consists of three sub-modules that operate independently yet collaboratively. The workflow of the sales data acquisition sub-module, the warehouse data integration sub-module, and the logistics data parsing sub-module is coordinated by a unified scheduler. The unified scheduler triggers data acquisition tasks based on the system clock, ensuring that data from different sources remain synchronized in the time reference system. Data quality checks are implemented during data acquisition, including checks for null values, ranges, and logic. When data anomalies are detected, an exception handling process is triggered, which includes logging errors, attempting to re-acquire data, and using default values ​​as a replacement. The inventory data acquisition module publishes the processed multi-dimensional dynamic data to a message middleware, which uses a publish-subscribe model to decouple the modules. The multi-dimensional dynamic data is encapsulated into standard data packets, each containing a metadata header and a payload. The metadata header records the data source, generation time, and data format version, while the payload contains specific values ​​for real-time sales fluctuation sequences, warehouse turnover cycle records, and logistics transportation delay indicators. The feature fusion processing module, acting as a data consumer, subscribes to these data packets from the message middleware and initiates subsequent feature extraction and fusion processes. The entire inventory data acquisition module is deployed on a server cluster with a highly available architecture. The server cluster is configured with load balancing and failover mechanisms to ensure service continuity. Network communication uses a VPN channel for encrypted transmission, and data is anonymized before being written to disk to protect business privacy. The module's operating status is monitored in real time, with monitoring metrics including data acquisition success rate, processing latency, and system resource utilization. Monitoring data is displayed on a unified visual operation and maintenance dashboard.

[0030] Example 2: See Figure 3 The feature fusion processing module receives multi-dimensional dynamic data from the inventory data acquisition module. This data includes real-time sales fluctuation sequences, warehouse turnover cycle records, and logistics delay indicators. The first component of this module is the sales trend analysis submodule, which specifically processes the real-time sales fluctuation sequences to extract sales trend features. This submodule uses a time series decomposition algorithm to separate the real-time sales fluctuation sequences into three components: a long-term trend component, a seasonal component, and an irregular residual component. The decomposition process is based on a sliding window model, with the window size set to a fixed number of days according to the business cycle. The submodule uses a locally weighted regression method to smooth the sequence data, which can adapt to non-linear changes in the sequence. The extracted linear growth component represents the long-term direction of sales data change, while the residual component contains random fluctuations and noise information. The linear growth component is used to predict future sales trends, and the residual component can be used for anomaly detection. The sales trend analysis submodule outputs the processing results to the matrix generation submodule for further integration.

[0031] The warehouse efficiency calculation submodule is responsible for processing warehouse turnover period records and generating warehouse efficiency features. The warehouse efficiency calculation submodule reads the average cargo dwell time data from the warehouse turnover period records and calculates the warehouse throughput efficiency value by combining the pre-set cargo category weight coefficients. The cargo category weight coefficients are based on cargo characteristics, with higher weight coefficients for perishable goods and high-value goods. The calculation process uses a weighted average method, multiplying the average dwell time of each type of cargo by its weight coefficient and then summing the results, and then dividing by the total weight coefficient. The warehouse throughput efficiency value is a quantitative indicator of the warehouse's ability to handle cargo per unit of time, and the warehouse efficiency calculation submodule recalculates this value at fixed time intervals. The warehouse efficiency calculation submodule also considers warehouse operation capacity limitations and introduces a capacity utilization factor for correction in the calculation. The calculation results are formatted into a standard data structure, containing timestamp, warehouse identifier, and efficiency value fields.

[0032] The logistics feature mapping submodule focuses on analyzing logistics transportation delay indicators and generating logistics stability features. The logistics feature mapping submodule receives the time decay coefficient data from the logistics transportation delay indicators and performs cluster analysis on the transportation routes based on the time decay coefficient data. The cluster analysis uses a density-based spatial clustering algorithm to group routes with similar delay patterns into the same category. The logistics feature mapping submodule calculates two core indicators for each route category: route stability score and sudden delay event frequency. The route stability score is calculated based on the variance of historical delay data for the route, with smaller variance indicating more stable routes. The sudden delay event frequency counts the number of abnormal events that exceed the threshold value per unit of time. The logistics feature mapping submodule organizes the analysis results into a route feature table, which contains route number, stability score, and event frequency fields. The route feature table is provided to the matrix generation submodule for feature fusion.

[0033] The matrix generation submodule is the core component of the feature fusion processing module, and is responsible for integrating features from different sources into a unified fusion feature matrix. The matrix generation submodule first establishes a unified time axis, with the smallest granularity consistent with the data collection frequency. The matrix generation submodule performs time alignment operations on feature data from different sources, using interpolation methods to handle data points with incomplete timestamp matching. The aligned feature data is reorganized into a three-dimensional feature tensor, with the three dimensions corresponding to the time dimension, feature type dimension, and feature value dimension. The time dimension is arranged in chronological order, and the feature type dimension includes sales trend features, warehouse efficiency features, and logistics stability features. The matrix generation submodule performs standardization processing on the feature values to eliminate the influence of different feature dimensions. The standardization processing uses the Z-score method to convert the feature values to a standard distribution with a mean of zero and a variance of one. The final generated fusion feature matrix is stored in the shared memory area for the inventory state modeling module to call.

[0034] The inventory state modeling module constructs a dynamic inventory state model based on the fusion feature matrix. The consumption rate calculation submodule of the inventory state modeling module first extracts the sales trend feature from the fusion feature matrix, which contains linear growth component information. The consumption rate calculation submodule derives the daily theoretical inventory consumption amount in combination with the commodity shelf life constraint condition, which comes from the commodity master data table. The calculation process considers the impact of the remaining shelf life on the sales speed, and the consumption rate of the commodity approaching the shelf life will be adjusted accordingly. The consumption rate calculation submodule uses an exponential decay model to simulate the shelf life effect, and the decay coefficient in the exponential decay model is inversely proportional to the remaining days of the shelf life. The daily theoretical inventory consumption amount is an important input parameter of the dynamic inventory state model, and the consumption rate calculation submodule calculates this value separately for each commodity category.

[0035] The replenishment delay compensation submodule processes the logistics stability feature in the fusion feature matrix, which contains route stability score data. The replenishment delay compensation submodule corrects the replenishment cycle promised by the supplier according to the route stability score, and the correction method is to increase the delay buffer time on the basis of the nominal replenishment cycle. The delay buffer time is negatively related to the route stability score, and the lower the stability score, the longer the buffer time. The replenishment delay compensation submodule uses historical delay data to fit the actual arrival time probability distribution, and the probability distribution uses a gamma distribution model to describe the random characteristics of the delay time. The replenishment delay compensation submodule establishes an independent probability distribution model for each supplier route, and the model parameters are dynamically updated as new data arrives. The actual arrival time probability distribution is input into the dynamic inventory state model as a random variable. The dynamic balance submodule is the final output link of the inventory state modeling module, and the dynamic balance submodule establishes a time-varying differential equation between the consumption amount and the replenishment amount. The time-varying differential equation takes the inventory level as the state variable, the inventory consumption rate as the outflow item, and the replenishment arrival rate as the inflow item. The replenishment delay probability in the differential equation is used as the equation damping coefficient, which adjusts the influence strength of the replenishment delay on the dynamic response of the system. The dynamic balance submodule solves the differential equation using a numerical integration method, and the numerical integration method uses the fourth-order Runge-Kutta method to ensure the calculation accuracy. The solving process advances with a fixed time step, and the parameter values are recalculated in each time step. The inventory change curve output by the dynamic balance submodule reflects the dynamic evolution process of the inventory level under the influence of multiple factors. The inventory state modeling module provides the dynamic inventory state model and its real-time calculation results to the subsequent anomaly detection module and prediction result generation module for use, completing the complete conversion process from multi-source data to inventory state prediction.

[0036] Referring to Figure 4The figure is a direct representation of the core function of the feature fusion processing module, focusing on the fusion feature values of the sales trend, warehouse efficiency, and logistics stability of three types of features in the 12-hour time dimension. The vertical axis feature type corresponds to: sales trend feature, warehouse efficiency feature, and logistics stability feature. The horizontal axis time is a unified time granularity, ensuring that the three types of features are aligned along the time axis, achieving synchronization of the time dimension of multi-source data. The color gradient quantifies the standardization degree of the fusion feature value, and the numerical annotation clearly presents the specific value of each feature at each time point: the sales trend shows a continuous growth, reflecting the linear growth trend of sales; the warehouse efficiency falls after reaching the peak, reflecting the periodical fluctuation of warehouse throughput; the logistics stability maintains a high level in the early stage and gradually decreases, reflecting the time decay of transportation route stability. This figure visualizes the three-dimensional feature tensor of time-feature type-feature value through the heat map, which is the function landing of the matrix generation submodule in the feature fusion processing module. The linear growth component of the sales trend, the warehouse throughput efficiency value, and the route stability score are reorganized into a fusion feature matrix according to the unified time granularity, providing a structured and visual multi-source fusion data basis for subsequent inventory state modeling, effectively supporting the transformation and verification from multi-source heterogeneous data to a unified feature space.

[0037] In an embodiment, the abnormality detection module analyzes the dynamic inventory state model results output by the inventory state modeling module to identify abnormal fluctuations. The abnormality detection module includes a baseline generation submodule, a deviation calculation submodule, and a link tracing submodule, which work in a pipeline manner. The baseline generation submodule processes the historical inventory change curve using a sliding window method. The historical inventory change curve is the output sequence of the dynamic inventory state model in the past period of time. The sliding window method defines a fixed-length window, which slides along the time axis and extracts the inventory data segment within the window. The baseline generation submodule calculates statistical features for the data in each window, including the arithmetic mean and the standard deviation. The arithmetic mean reflects the central tendency of the inventory level, and the standard deviation measures the dispersion of inventory fluctuations. The baseline generation submodule stores the calculation results of each window as a reference range, which is composed of the mean value plus or minus several times the standard deviation. The reference range is dynamically updated over time as new data enters the window and the oldest data is removed, maintaining the recency of the data in the window.

[0038] The deviation calculation sub-module monitors the current output value of the dynamic inventory state model in real time, and the current output value of the dynamic inventory state model is the inventory level prediction value at the latest time point. The deviation calculation sub-module compares the current output value with the benchmark range provided by the baseline generation sub-module, and the comparison operation calculates the degree to which the current output value deviates from the boundary of the benchmark range. The deviation calculation sub-module defines a continuous over-threshold time length judgment rule, and triggers an abnormality mark when the inventory value exceeds the benchmark range at a plurality of time points in succession. The setting of the continuous over-threshold time length needs to consider the business scenario, and too short is easy to produce false positives and too long may lead to false negatives. The deviation calculation sub-module sets different threshold parameters for different commodity categories, and the threshold of high-value commodities or key materials is more sensitive. The output of the deviation calculation sub-module is an abnormal state flag and a deviation quantization value, and the abnormal state flag indicates whether an anomaly is detected, and the deviation quantization value represents the severity of the anomaly.

[0039] The link tracing sub-module starts the analysis process after the abnormality mark is triggered, and the link tracing sub-module associates the abnormal fluctuation node with the original data dimensions in the feature fusion processing module. The abnormal fluctuation node is the abnormal time point identified by the deviation calculation sub-module, and the link tracing sub-module establishes a reverse link from the abnormal node to the multi-source data through a data bloodline tracking technology. The data bloodline tracking records the source and transformation history of each data in the feature fusion processing process, and the link tracing sub-module queries the bloodline information to locate the root cause link causing the anomaly. The link tracing sub-module analyzes the original input data of the sales data acquisition sub-module, the warehouse data integration sub-module and the logistics data analysis sub-module, and identifies which link's data pattern is most related to the abnormal fluctuation. The link tracing sub-module generates an abnormality analysis report, and the report includes the time of the anomaly, the affected commodity category, the suspected link positioning and the confidence evaluation.

[0040] The core of the anomaly detection module is the quantification method of the deviation, and the deviation calculation sub-module calculates the real-time deviation index using the following formula:

[0041] Wherein: represents the time point of the abnormal index, and the larger the value is, the more serious the deviation from the benchmark is. is the size of the backtracking time window, which determines the number of points considering historical data. i represents the time index offset. is the time decay weight coefficient, which decreases with the increase of , and gives higher importance to recent data. is the actual inventory observation value or model output value at time point . The character is the baseline average value corresponding to time point . The character is the actual inventory observation value or model output value at time point The corresponding baseline standard deviation. The summation operation accumulates the normalized deviation of each time point within the window.

[0042] The implementation of the anomaly detection module needs to consider the computational efficiency and real-time requirements. The anomaly detection module uses a distributed stream processing framework to handle high-speed data streams. The sliding window calculation maintains the data within the window through a state management mechanism, which ensures that the calculation state can be restored in the event of system failure. The anomaly detection module is integrated with the alarm system, which automatically sends alarm notifications to relevant personnel when an anomaly is confirmed. The alarm notification contains key contextual information, such as the duration of the anomaly, the scope of the impact, and possible causes. The anomaly detection module also provides a historical anomaly query interface, supporting the retrieval of historical anomaly events by time range, product category, and anomaly type. The analysis of historical anomaly events helps to optimize detection parameters and improve the accuracy of baseline generation and deviation calculation. In the implementation of the baseline generation submodule of the anomaly detection module, the size of the sliding window is an important parameter. The window size needs to balance sensitivity and stability, as a too small window is sensitive to noise and a too large window is slow to respond. The baseline generation submodule experiments the detection effect of different window sizes and determines the optimal window length in combination with the business cycle. The baseline generation submodule also considers seasonal factors and introduces seasonal adjustment for inventory data with strong seasonality. The seasonal adjustment uses a classic decomposition method to separate the seasonal component, and calculates the baseline for the data after removing the seasonal component. This method improves the representativeness of the baseline and reduces false positives caused by seasonal fluctuations.

[0043] The continuous over-threshold detection logic of the deviation calculation submodule needs to handle boundary conditions such as data missing or transient spikes. The deviation calculation submodule implements a data interpolation mechanism to handle occasional missing points, which uses the average of the valid data before and after to fill the gaps. For transient spikes, the deviation calculation submodule uses a combination of duration and deviation amplitude to make a comprehensive judgment and avoid false positives triggered by short-term interference. The deviation calculation submodule maintains a state machine that tracks the current number of consecutive over-threshold points. The state machine switches between normal and warning states, and only confirms an anomaly when the set number is reached. The data lineage tracking of the link tracing submodule is based on the metadata management system, which records the source, transformation, and dependency relationships of the data. The link tracing submodule queries the metadata to obtain the lineage map of the abnormal data points, which is a directed acyclic graph showing the data flow path. The link tracing submodule traces back along the lineage map and calculates the contribution of each source data to the anomaly. The contribution calculation uses a sensitivity analysis method to analyze the impact of input data perturbation on the anomaly indicator. The link tracing submodule marks the source link with the highest contribution as the suspected root cause, providing a clear direction for manual intervention.

[0044] The deployment architecture of the anomaly detection module considers scalability and fault tolerance. The anomaly detection module is deployed in the form of microservices on a container platform. Each submodule runs as an independent service, and services communicate asynchronously through a message queue. This architecture allows individual submodules to be horizontally scaled to handle increasing data volume. The anomaly detection module configures a health check endpoint for each service, and the monitoring system periodically checks the status of the services. The implementation code of the anomaly detection module includes unit tests and integration tests, and the test cases simulate various abnormal scenarios to verify the correctness of the detection logic. Automated testing ensures that code modifications do not break existing functionality and maintain system stability.

[0045] In embodiment 4, the prediction result generation module outputs inventory prediction results for future periods based on a dynamic inventory status model. The prediction result generation module includes three core components: a multi-scenario simulation submodule, a risk quantification submodule, and a warning generation submodule. The multi-scenario simulation submodule is responsible for generating multiple possible inventory consumption and replenishment simulation paths. The multi-scenario simulation submodule constructs a test environment based on different supply chain link optimization assumptions, including sales promotion activities, logistics network efficiency improvement, warehouse operation process optimization, and other scenarios. The multi-scenario simulation submodule uses the Monte Carlo simulation method to randomly generate parameter combinations, each of which represents a possible future state. The generation of simulation paths considers the continuity in the time dimension, and the state changes at different time points conform to the business logic constraints. The multi-scenario simulation submodule runs the simulation process independently for each product category, which is classified according to sales characteristics, shelf life, and supply stability. The number of simulation paths is determined according to the computing resources and accuracy requirements, usually requiring thousands of simulations to cover possible situations.

[0046] The risk quantification submodule statistically analyzes the simulation paths output by the multi-scenario simulation submodule. The risk quantification submodule calculates the frequency of inventory levels falling below a safety threshold for each simulation path. The safety threshold is a pre-set minimum inventory level below which a stockout risk may occur. The risk quantification submodule converts the frequency into a probability value, which represents the likelihood of a stockout occurring at a specific future time point. The output of the risk quantification submodule is a stockout risk heat map, which organizes data in a two-dimensional matrix format. The row dimension of the matrix represents time points, and the column dimension represents product categories. The matrix element values are stockout probabilities. The risk quantification submodule visualizes the probability size using color coding, with dark areas representing high risk and light areas representing low risk. The heat map generation uses image rendering technology, supporting interactive zooming and detailed querying. The early warning generation submodule converts the stockout risk heat map generated by the risk quantification submodule into actionable early warning information. The early warning generation submodule sets a critical probability value as the early warning trigger condition. The critical probability value is set based on the enterprise's risk tolerance, and different product categories can have different critical values. The early warning generation submodule scans each matrix element in the heat map to identify risk points that exceed the critical value. The early warning generation submodule maps the risk points to specific product categories and time nodes, forming a structured early warning event list. The early warning event list includes fields such as event number, product code, warning time, risk probability level, and recommended measures. The early warning generation submodule sends early warning events to relevant management personnel through a message push interface.

[0047] The implementation of the multi-scenario simulation submodule relies on three specialized components: the sales fluctuation scenario library, the logistics disruption simulator, and the warehouse failure injector. The sales fluctuation scenario library stores the characteristic parameters of historical extreme sales fluctuation events, including fluctuation amplitude, duration, and impact range. The sales fluctuation scenario library uses a database table to manage scenario data, with each scenario having a unique identifier and detailed description. The sales fluctuation scenario library supports scenario retrieval and matching functions, recommending similar historical scenarios based on current sales trends. The logistics disruption simulator builds a transportation delay model based on external data sources, including regional weather forecasts, traffic control announcements, and supplier operation status. The logistics disruption simulator uses an influence propagation algorithm to calculate the cascading effects of delays, taking into account the topology of the transportation network. The warehouse failure injector simulates the impact of equipment failures and human errors on warehouse efficiency, with the failure model based on historical maintenance records and operation log analysis. The warehouse failure injector defines a library of fault types, each with a corresponding occurrence probability and impact intensity.

[0048] The prediction result generation module further includes a prediction confidence evaluation submodule that evaluates the reliability of the out-of-stock risk probability distribution. The prediction confidence evaluation submodule analyzes the data integrity of the fused feature matrix, which is measured by the missing value ratio and data freshness indicators. The prediction confidence evaluation submodule calculates the historical prediction accuracy of the dynamic inventory state model, which is calculated using the mean absolute percentage error method. The prediction confidence evaluation submodule generates a confidence score by combining the data integrity and historical accuracy, with the confidence score ranging from zero to one hundred percent. The prediction confidence evaluation submodule assigns an independent confidence score to each prediction period, and the score result is attached to the prediction result for user reference. The simulation path generation process of the multi-scenario simulation submodule requires management of a large amount of computing resources, and the multi-scenario simulation submodule uses a distributed computing framework to execute simulation tasks in parallel. The distributed computing framework decomposes simulation tasks into multiple subtasks, which are assigned to different computing nodes for simultaneous execution. The multi-scenario simulation submodule designs a task scheduling algorithm to optimize computing efficiency, taking into account node load and data locality. The storage of simulation results uses a columnar database, which provides efficient data compression and fast query capabilities. The multi-scenario simulation submodule implements the repeatability of the simulation process, ensuring that the same parameters produce the same results by setting a random seed.

[0049] The statistical analysis method of the risk quantification submodule needs to handle the tail risk of the probability distribution, and the risk quantification submodule uses the extreme value theory to estimate the occurrence probability of small probability high impact events. The extreme value theory models the extreme values of the simulation results, fitting a generalized Pareto distribution to describe the tail characteristics. The risk quantification submodule calculates the value at risk and the value at risk indicators, with the value at risk representing the maximum possible loss at a given confidence level, and the value at risk measuring the expected loss exceeding the value at risk. The risk quantification submodule generates quantile information of the probability distribution, which helps users understand the overall picture of the risk distribution. The warning event list of the early warning generation submodule needs to support priority ranking, and the early warning generation submodule designs a multi-factor priority scoring algorithm. The priority scoring algorithm considers four dimensions: risk probability, impact degree, processing urgency, and commodity importance, each with a weight coefficient. The early warning generation submodule implements automatic classification and routing of early warning events, with different categories of early warning sent to different processing teams. The early warning generation submodule provides a warning confirmation and feedback mechanism, and management personnel can record processing measures and effects after handling the warning. The early warning generation submodule maintains the life cycle state of the early warning event, including pending, in processing, solved, and closed.

[0050] The scoring model of the prediction confidence assessment submodule needs to be updated regularly to maintain accuracy. A model retraining cycle is set. Model retraining updates scoring parameters using the latest historical data to adapt to changes in data characteristics. The prediction confidence assessment submodule provides explanatory information for confidence scores. The explanatory information helps users understand the influencing factors behind the scores. The prediction confidence assessment submodule is integrated with a user feedback system. User evaluations of prediction accuracy serve as input for improving the scoring model. See Table 1.

[0051] Table 1: Multi-scenario simulation parameter configuration table Parameter Category Parameter Name Parameter Description Value Method Sales Parameter Fluctuation Amplitude Relative Proportion of Sales Volume Change Historical Quantile Sales Parameter Growth Trend Long-term Change Direction of Sales Linear Regression Slope Logistics Parameter Delay Probability Possibility of Delay Occurring in Transportation Historical Frequency Statistics Logistics Parameter Delay Duration Duration of Delay Events Distribution Fitting Warehouse Parameter Failure Rate Number of Equipment Failures per Unit Time Poisson Distribution Parameter Warehouse Parameter Recovery Time Average Time Required for Failure Repair Exponential Distribution Mean The parameter configuration of the multi-scenario simulation submodule uses the structured scheme shown in the table. The parameter configuration table defines the meaning and value method of each simulation parameter. The parameter configuration table is stored in the configuration management database, supporting version control and change audit. The multi-scenario simulation submodule provides a graphical interface for parameter adjustment. Users can modify parameter values through sliders and drop-down menus. Parameter modifications take effect immediately without the need to restart the simulation process. The multi-scenario simulation submodule records the parameter version used in each simulation, ensuring the traceability of the results. The parameter configuration table also contains parameter constraint rules that prevent unreasonable parameter combinations. The design of the parameter configuration table supports extensibility. New parameters can be added to the table without modifying the program code.

[0052] In Example 5, the time-varying differential equation established by the dynamic balance submodule of the inventory state modeling module is simulated and optimized through digital twinning technology. The system constructs a virtual inventory model that fully corresponds to the physical supply chain. The virtual inventory model is a high-fidelity mapping of the physical entity in the digital space. The virtual inventory model receives real-time multi-dimensional dynamic data updates from the inventory data acquisition module. The multi-dimensional dynamic data includes real-time sales fluctuation sequences, warehouse turnover period records, and logistics transportation delay indicators. These data continuously flow into the virtual inventory model through the data bus. The virtual inventory model internally runs the time-varying differential equation established by the dynamic balance submodule. The time-varying differential equation takes inventory level as the state variable and replenishment delay probability as the damping coefficient. The simulation environment of the virtual inventory model allows testing various replenishment strategies under zero risk conditions. For example, it can simulate the impact of shortening a supplier's replenishment period from seven days to five days on overall inventory levels. The virtual inventory model outputs simulation results including inventory level change curves, out-of-stock probability curves, and cost change curves. Management personnel can select the optimal solution by comparing the simulation results of different strategies.

[0053] The implementation of digital twin technology requires establishing an accurate correspondence between physical entities and virtual models, with each warehouse, transportation route, and sales channel having a corresponding digital entity in the virtual world. The digital entity is associated with the physical entity through a unique identifier, and the attribute values of the digital entity are updated in real time through a sensor network and business systems. The virtual inventory model runs using an event-driven mechanism that triggers model recalculation when receiving data update events. The time advancement mechanism of the virtual inventory model supports both real-time simulation and accelerated simulation modes, with real-time simulation keeping pace with the physical world and accelerated simulation allowing rapid simulation of long-term effects. The virtual inventory model provides a rich visualization interface that displays the supply chain network structure in three dimensions and highlights bottleneck links.

[0054] The adaptive learning submodule of the anomaly detection module implements dynamic optimization of the detection threshold. The adaptive learning submodule analyzes the subsequent impact of labeled abnormal fluctuation nodes and false positive records. The subsequent impact of abnormal fluctuation nodes is quantified by tracking inventory changes after the occurrence of anomalies, such as recording the amount of stockout or the degree of inventory accumulation caused by anomalies. False positive records are events that the system labels as anomalies but do not actually have a negative impact on business, and are obtained through manual feedback and reverse verification of business results. The adaptive learning submodule uses a reinforcement learning algorithm to process these historical data, which models the threshold adjustment problem as a Markov decision process. The state space of the reinforcement learning algorithm includes the current threshold setting, historical anomaly statistics, and business environment characteristics, the action space is the magnitude of threshold increase or decrease, and the reward function is designed based on the weighted combination of false positive rate and false negative rate. The reinforcement learning algorithm of the adaptive learning submodule is implemented using the Q-learning algorithm, which finds the optimal policy by iteratively updating the action value function. The learning process of the Q-learning algorithm uses the experience replay technique to improve sample utilization, and the experience replay technique stores historical state transition data for repeated learning. The adaptive learning submodule sets a balance mechanism between exploration and utilization, encouraging diverse threshold settings in the early stages of the algorithm and gradually favoring known effective settings as experience accumulates. The learning period of the adaptive learning submodule aligns with the business cycle, such as adjusting the threshold weekly or monthly, to avoid excessive frequent changes that affect system stability.

[0055] Digital twin technology and adaptive learning mechanisms work together in practice, with the virtual inventory model providing a rich simulation environment for adaptive learning. The adaptive learning submodule can test the effects of different threshold settings in the virtual inventory model, quickly accumulating learning experience through a large number of simulations without affecting actual business. This collaborative mechanism significantly shortens the convergence time of the adaptive learning submodule and improves the safety and efficiency of threshold optimization. The virtual inventory model can also simulate extreme scenarios to provide the adaptive learning submodule with critical training data that is less likely to occur in reality.

[0056] The implementation case of the digital twin system is in the inventory management optimization of a large retail enterprise, which has fifty warehouses and three hundred transportation routes distributed across the country. The virtual inventory model construction phase first completes the digital modeling of physical assets, creating a detailed model for each warehouse that includes capacity, processing power, and spatial layout. The transportation route model includes distance, average speed, and historical delay data, and the sales channel model includes demand patterns and seasonal characteristics. The virtual inventory model connects the enterprise's enterprise resource planning system, warehouse management system, and transportation management system through a data interface, obtaining real-time inventory levels, order status, and vehicle location information. In this specific case, the virtual inventory model is used to evaluate a new regional distribution center site selection scheme, and management personnel test three candidate sites in the virtual environment. The virtual inventory model simulates the impact of each site on transportation timeliness, inventory turnover, and emergency response capability, with a simulation period of six months. The multi-dimensional evaluation report generated by the virtual inventory model shows that locating near a transportation hub can reduce average replenishment time by eighteen hours and reduce safety stock level by fifteen percent. Based on the simulation results, the enterprise makes a final site selection decision, and actual operational data later verifies the accuracy of the simulation.

[0057] The application case of the adaptive learning submodule occurs in the optimization of the same enterprise's anomaly detection system. During the initial operation phase, the system has a high false positive rate due to conservative threshold settings. The adaptive learning submodule collects data from the first month of operation, including two hundred and forty anomaly labels and subsequent verification results. After analyzing this data, the reinforcement learning algorithm finds that the threshold sensitivity can be appropriately increased for high turnover rate items and dynamically adjusted for seasonal items. The adaptive learning submodule implements the optimized threshold strategy starting in the second month, reducing the false positive rate from thirty-five percent to twelve percent while maintaining a high capture rate for true anomalies. Data synchronization between the virtual inventory model and the physical system is achieved through a distributed message queue, which ensures data transmission reliability and order. The virtual inventory model is designed using a microservices architecture, with each functional module deployed as an independent service and communicating through application programming interfaces. The simulation engine of the virtual inventory model is based on discrete event simulation technology, which efficiently handles concurrent activities in the supply chain. The virtual inventory model supports version management functionality, which allows saving model states at different points in time for easy backtracking and result comparison. The adaptive learning submodule is implemented using a combination of online learning and offline training, with online learning gradually adjusting threshold parameters and offline training performing large-scale historical data mining. The reward function design of the adaptive learning submodule incorporates business cost factors, such as considering stockout costs and inventory holding costs. The adaptive learning submodule provides an explainable report on threshold adjustments, explaining the basis and expected effects of each adjustment, enhancing user trust in system decisions.

[0058] The long-term operation of the digital twin system generates a large amount of simulation data, which is stored in a time-series database for model optimization and knowledge discovery. The virtual inventory model continuously calibrates model parameters by comparing simulation data with actual operational data, and the calibration process uses a maximum likelihood estimation method to improve model accuracy. The virtual inventory model also supports a hypothesis analysis function, which allows users to modify model parameters and observe changes in system behavior, providing support for strategic decision-making. The reinforcement learning algorithm of the adaptive learning submodule faces the problem of reward sparsity, which is particularly evident when there are few abnormal events. To address the problem of reward sparsity, the adaptive learning submodule uses a hierarchical reinforcement learning framework, which decomposes the threshold optimization task into multiple subtasks. The adaptive learning submodule also uses imitation learning techniques to initialize the policy network, which learns the pattern of manual threshold adjustment and accelerates algorithm convergence. The performance evaluation of the adaptive learning submodule uses a cross-validation method, which ensures that the learned policy has good generalization ability. The combined action of the virtual inventory model and the adaptive learning submodule creates a continuously evolving inventory management system that predicts future states through digital twin technology and optimizes decision parameters through adaptive learning. This combined approach enables the inventory management system to adapt to dynamic business environments and maintain robustness and efficiency in complex supply chain networks.

[0059] It should be noted that the relational terms herein, such as first and second, are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any such actual relationship or order between or among the entities or actions. Moreover, the terms "comprises," "comprising," or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0060] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A multi-source data fusion based inventory forecasting system, characterized in that, The application relates to a supply chain inventory state anomaly detection and prediction method, which comprises the following steps: An inventory data acquisition module acquires multi-dimensional dynamic data in a supply chain link, wherein the multi-dimensional dynamic data comprises a real-time sales fluctuation sequence, warehouse turnover period records and a logistics transportation delay index; A feature fusion processing module extracts a sales trend feature, a warehouse efficiency feature and a logistics stability feature based on the multi-dimensional dynamic data, and generates a fusion feature matrix after aligning the three types of features according to a time axis; An inventory state modeling module constructs a dynamic inventory state model reflecting the correlation between inventory consumption rate and replenishment delay according to the fusion feature matrix; An anomaly detection module performs deviation analysis on an inventory change curve output by the dynamic inventory state model, identifies an abnormal fluctuation node and marks the corresponding supply chain link; A prediction result generation module outputs an inventory surplus early warning value and a stockout risk probability distribution in a future period based on the corrected dynamic inventory state model. 2.The inventory forecasting system based on multi-source data fusion according to claim 1, wherein, The inventory data acquisition module comprises a sales data acquisition submodule, a warehouse data integration submodule and a logistics data analysis submodule. 3.The inventory forecasting system based on multi-source data fusion of claim 1, wherein, The feature fusion processing module comprises a sales trend analysis submodule, a warehouse efficiency calculation submodule and a logistics feature mapping submodule. 4.The inventory forecasting system based on multi-source data fusion of claim 1, wherein, The inventory state modeling module comprises a consumption rate calculation submodule, a replenishment delay compensation submodule and a dynamic balance submodule. 5.The inventory forecasting system based on multi-source data fusion of claim 1, wherein, The anomaly detection module comprises a baseline generation submodule, a deviation calculation submodule and a link tracing submodule. 6.The inventory forecasting system based on multi-source data fusion of claim 1, wherein, The prediction result generation module comprises: a multi-scenario simulation submodule generates multiple simulated paths of inventory consumption and replenishment based on optimization assumptions of different supply chain links; a risk quantification submodule counts the frequency of occurrence of inventory levels below a safety threshold under each simulated path and outputs a stockout risk heat map; and a warning generation submodule maps areas in the heat map that exceed a critical value to specific product categories and time nodes to form a warning event list.

7. The inventory forecasting system based on multi-source data fusion according to claim 6, characterized in that, The multi-scenario simulation submodule comprises: a sales fluctuation scenario library stores feature parameters of extreme sales fluctuation events in history to generate test cases for sudden sales growth; a logistics interruption simulator constructs a cascading effect model of transportation delays based on regional weather data and traffic control records; and a warehouse failure injector simulates the impact of equipment downtime and human operation errors on warehouse turnover efficiency. 8.The inventory forecasting system based on multi-source data fusion of claim 1, wherein, The prediction result generation module further comprises: a prediction confidence evaluation submodule for calculating the uncertainty interval of the stockout risk probability distribution; and the prediction confidence evaluation submodule assigns a confidence score to each prediction period based on the data integrity of the fusion feature matrix and the historical prediction accuracy of the dynamic inventory state model. 9.The inventory forecasting system based on multi-source data fusion of claim 4, wherein, The time-varying differential equation established by the dynamic balance submodule is further simulated and optimized through digital twinning technology: a virtual inventory model corresponding to the physical supply chain is constructed; the virtual inventory model receives real-time updates of the multi-dimensional dynamic data and runs the time-varying differential equation to simulate changes in inventory status under different strategies, thereby verifying the effectiveness of replenishment strategies before actual changes in physical inventory occur.

10. The inventory forecasting system based on multi-source data fusion of claim 5, wherein, The anomaly detection module further comprises: an adaptive learning submodule for dynamically adjusting the detection threshold used in the deviation calculation submodule; the adaptive learning submodule analyzes the subsequent impact of labeled abnormal fluctuation nodes and false positive records, and periodically optimizes the detection threshold using a reinforcement learning algorithm based on the analysis results to reduce false positives and improve sensitivity to real anomalies.

Images