An ERP inventory management system dynamic optimization management method and system based on artificial intelligence

Abstract

The application discloses an ERP (Enterprise Resource Planning) purchase, sale and storage system dynamic optimization management method and system based on artificial intelligence, relates to the cross technical field of artificial intelligence and supply chain management, and comprises the following steps: collecting ERP purchase, sale and storage business data, performing standardized processing to form material dynamic state basic data set, extracting running state features based on the material dynamic state basic data set, and constructing a material dynamic state vector; an enhanced context sequence is constructed by fusing a historical demand sequence and a future business event marker, a multi-expert gate demand prediction model is input in combination with the material dynamic state vector, and a demand prediction result is generated. Through the prediction model fusing business knowledge and the risk grading mechanism driven by data, the application realizes accurate quantification of demand disturbance and dynamic adjustment of service level, and significantly improves the adaptability and inventory efficiency of the ERP purchase, sale and storage system dynamic optimization management.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and supply chain management, and in particular to a dynamic optimization management method and system for ERP inventory management based on artificial intelligence. Background Technology

[0002] In recent years, with the acceleration of enterprise digital transformation, artificial intelligence technology has been gradually introduced into the inventory management of ERP systems to improve the intelligence level of forecasting and inventory decision-making. Early methods mainly relied on static safety stock models (such as the (ROP,Q) strategy), whose safety factor and service level were usually set by human experience, making it difficult to adapt to sudden changes in demand or supply fluctuations. Subsequently, time series models (such as ARIMA and exponential smoothing) and machine learning methods (such as XGBoost and LSTM) were used for demand forecasting, but most of them only used historical sales data, ignoring the structural impact of known future business events (such as promotions and holidays) on demand. In recent years, some studies have attempted to integrate external knowledge to build "enhanced" forecasting models and introduce multi-expert architectures to improve adaptability to different demand patterns, while evaluating forecast reliability through uncertainty quantification (such as quantile regression and Monte Carlo Dropout).

[0003] However, existing technologies still have shortcomings. First, while traditional AI prediction models can output point predictions or probability distributions, they fail to separate and calibrate data uncertainty from model uncertainty, leading to distorted safety stock calculations—especially in scenarios involving new products, low-frequency or highly volatile materials, easily resulting in overstocking or frequent stockouts. Second, current ERP systems often use a globally uniform service level (e.g., 95%), without differentiating settings based on the current supply and demand risk level of materials. This makes it impossible to achieve refined control of "high risk, high protection; low risk, low redundancy." Even when some systems introduce risk scoring, its weights and thresholds are often determined based on experience, lacking a data-driven calibration process based on historical stockout attribution and business cost structure, resulting in rigid strategies and difficulties in optimization. Therefore, current ERP inventory management systems still suffer from problems such as the inability to effectively translate predictive uncertainty into dynamic inventory parameters and a disconnect between risk assessment and replenishment strategies. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an AI-based dynamic optimization management method and system for ERP inventory management systems to solve the problems of predictive uncertainty not being effectively converted into dynamic inventory parameters and the disconnect between risk assessment and replenishment strategies.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a dynamic optimization management method for an ERP inventory management system based on artificial intelligence, comprising: Collect ERP sales and inventory data and standardize it to form a basic dataset of material dynamic status. Extract operational status features based on the basic dataset of material dynamic status and construct material dynamic status vectors. By constructing an enhanced context sequence that integrates historical demand sequences and future business event markers, and combining the material dynamic state vector input to a multi-expert gated demand forecasting model, demand forecasting results are generated. Based on expert scale parameters and disturbance reconstruction inference, data uncertainty and model uncertainty are calculated and calibrated respectively. Through historical normalized residual quantile calibration, comprehensive demand disturbance parameters are obtained. Based on the comprehensive demand disturbance parameters, a segmented mapping risk index is constructed, the target service level and the corresponding safety factor are dynamically determined, and the optimization control parameters are calculated to form a dynamic optimization control parameter set. The set of dynamically optimized control parameters is written into the ERP control table to drive the generation of inventory warnings and procurement tasks, and the results of parameter execution feedback are collected for analysis and optimization management.

[0007] As a preferred embodiment of the AI-based dynamic optimization management method for ERP inventory management systems described in this invention, the collection of ERP inventory business data refers to collecting original business records related to the target material within a preset statistical period from the ERP system already deployed by the enterprise.

[0008] As a preferred embodiment of the AI-based ERP inventory management system dynamic optimization method described in this invention, the standardization process to form a material dynamic status basic dataset refers to performing standardized processing operations such as material coding unification, unit of measurement unification, timestamp alignment, and data cleaning on the original business records to form a material dynamic status basic dataset.

[0009] As a preferred embodiment of the AI-based ERP inventory management system dynamic optimization method described in this invention, the step of extracting operational status features from the material dynamic status dataset and constructing a material dynamic status vector refers to calculating key operational status features for each material based on the material dynamic status dataset to form a material dynamic status vector.

[0010] As a preferred embodiment of the AI-based dynamic optimization management method for ERP inventory management system described in this invention, the method of obtaining comprehensive demand disturbance parameters refers to extracting historical actual daily demand from the ERP database to construct a historical sequence, and combining it with future promotions, holidays and peak season markers to form an enhanced context sequence along the time axis that includes historical observations and future known business knowledge. The enhanced context sequence and the material dynamic state vector are input into the multi-expert gated demand forecasting model. The joint hidden state is extracted by forward and reverse GRU and concatenated with the material dynamic state vector to form a joint feature vector. The gating network calculates the weight of each expert based on this vector, and the weighted fusion of the forecasts output by each expert generates the future demand forecast. Using the actual future demand within the historical window as the supervision signal, a Gaussian mixture prediction distribution weighted by gating weights is constructed. The negative log-likelihood loss function is adopted, and the multi-expert gating demand prediction model is jointly optimized by backpropagation and gradient descent. Based on the output of the multi-expert gated demand forecasting model, the data uncertainty and model uncertainty are calculated separately, and then summed to obtain the total uncertainty. The total uncertainty is then calculated using historical normalized residuals. The quantile determines the calibration coefficient, the total uncertainty is calibrated and averaged to generate the comprehensive demand disturbance parameter, and the cumulative and average predicted demand within the prediction period are calculated simultaneously.

[0011] As a preferred embodiment of the AI-based ERP inventory management system dynamic optimization method described in this invention, the formation of the dynamic optimization control parameter set refers to the weighted calculation based on the comprehensive demand disturbance parameters and the standard deviation of the procurement lead time, combined with the average predicted demand and the lead time length, to obtain a risk index. The target service level is determined by segmenting and mapping according to the risk index through a preset threshold, and the corresponding dynamic safety coefficient is obtained by calling the standard normal distribution quantile function. Based on the dynamic safety factor, the dynamic safety stock, dynamic replenishment point, replenishment trigger value, target inventory limit, and suggested replenishment quantity after correction by the maximum replenishment quantity constraint are calculated in sequence to form a dynamic optimization control parameter set.

[0012] As a preferred embodiment of the AI-based ERP inventory management system dynamic optimization method described in this invention, the step of writing the dynamic optimization control parameter set into the ERP control table to drive inventory warning generation and procurement task generation refers to writing the dynamic optimization control parameter set into the ERP material inventory control master data table, generating procurement application records based on replenishment trigger parameters, and writing warning or adjustment prompt records when the inventory age exceeds the limit or the inventory exceeds the limit.

[0013] As a preferred embodiment of the AI-based dynamic optimization management method for ERP inventory management systems described in this invention, the parameter execution feedback result refers to the actual execution feedback data collected from the ERP database to generate business records.

[0014] As a preferred embodiment of the AI-based dynamic optimization management method for ERP inventory management systems described in this invention, the analysis and optimization management refers to calculating prediction deviation indicators, stockout deviation indicators, and inventory deviation indicators based on actual execution feedback data, weighting and fusing the deviation indicators, calculating a comprehensive update trigger value, and setting an update threshold for optimization management judgment.

[0015] Secondly, the present invention provides an AI-based dynamic optimization management system for ERP inventory management systems, comprising: The data acquisition standardization module is used to collect various types of original business records from the ERP system, and after unification of coding, unit conversion, time alignment and cleaning, form a basic dataset. The material status vectorization module is used to calculate a seven-dimensional feature vector based on standardized data, including average daily demand, demand standard deviation, inventory coverage days, lead time and its fluctuation, stockout frequency, and inventory aging, to construct a feature vector that represents the real-time operating status of materials. The enhanced context modeling module is used to integrate historical daily demand with event markers such as future promotions / holidays to construct a joint historical-future input sequence with masked placeholders; The multi-expert gating prediction module is used to generate probabilistic demand predictions with scale parameters by dynamically weighting the outputs of multiple experts through a gating network using BiGRU encoding of joint sequences. The risk-driven control parameter generation module is used to construct a segmented mapping risk index based on demand disturbances and procurement lead time fluctuations, dynamically determine the safety factor, and calculate a complete set of inventory control parameters.

[0016] The beneficial effects of this invention are as follows: The multi-expert gating demand forecasting constructed in this invention not only integrates historical demand but also explicitly embeds known future business events (such as promotions and holidays) in the ERP system. It adaptively matches different demand patterns through a multi-expert gating mechanism, while separating and calibrating data uncertainty and model uncertainty to generate interpretable demand disturbance parameters that have been corrected for historical error quantiles. Furthermore, the demand disturbance parameters and the fluctuation of procurement lead time together constitute a comprehensive risk index, driving the dynamic grading of the target service level. This abandons fixed thresholds and instead automatically divides low, medium, and high risk intervals based on historical risk distribution data, mapping them to increasing service levels. The results are then transformed into dynamic safety coefficients through a standard normal quantile function, achieving a closed-loop optimization from "intelligent forecasting" to "adaptive inventory control". Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a dynamic optimization management method for an ERP inventory management system based on artificial intelligence.

[0019] Figure 2 This is a schematic diagram of a dynamic optimization management system for an ERP inventory management system based on artificial intelligence.

[0020] Figure 3 Flowchart for dynamically optimizing control parameters calculation.

[0021] Figure 4 Flowchart for determining dynamic safety factor. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a dynamic optimization management method for an ERP inventory management system based on artificial intelligence, including the following steps: S1. Collect ERP sales and inventory business data, and standardize the data to form a basic dataset of material dynamic status. Extract the operating status features based on the basic dataset of material dynamic status and construct the material dynamic status vector. S1.1: Collect data from the enterprise's existing ERP system for a preset statistical period. Internal original business records related to the target material.

[0026] Specifically, from the enterprise's existing ERP system, collect original business records related to the target material within a preset statistical period, including: Purchase order records (including order placement time, purchase quantity, and supplier information); Purchase and delivery records (including actual delivery time and quantity received); Sales order records (including customer order time and order quantity); Sales outbound records (including actual delivery time and outbound quantity); Inventory ledger records (including daily book inventory); Lock inventory records (including inventory frozen due to sales orders or production reservations); Purchase records in transit (including the quantity of orders placed but not yet delivered and the estimated delivery time); Material master data records (including material code, inventory unit, purchasing unit and unit conversion relationship).

[0027] It should be noted that, The period is 30 calendar days, which covers a complete business month. This effectively balances data stability and timeliness, and conforms to the monthly operational analysis practices of enterprises and the default settings of mainstream ERP systems.

[0028] S1.2: Perform material coding unification, unit of measurement unification, timestamp alignment, and data cleaning and standardization on the original business records to form a basic dataset of material dynamic status.

[0029] Specifically, the original business records are standardized and key inventory status indicators are calculated to characterize the real-time availability of materials, forming a basic dataset of material dynamic status organized in a two-dimensional structure of "material identifier - statistical time", which serves as the sole input source for subsequent feature extraction and parameter calculation.

[0030] The standardized processing operations include: Unified material coding: Mapping different codes (such as old codes and aliases) for the same material from different sources to a unique standard material code in the ERP system; Unified units of measurement: Converting quantities from non-inventory units such as purchasing and sales into unified inventory units set by the ERP system; Timestamp alignment: Aggregate all events (such as order placement, arrival, and shipment) by natural day granularity, with 24:00 as the statistical cutoff time for each day; Data cleaning: Remove duplicate records, correct outliers (such as negative inventory or excessively large purchases) using the average of the nearest days, and fill in missing values ​​using linear interpolation.

[0031] The key inventory status indicators include book inventory and available inventory, and the calculation formulas are as follows:

[0032] In the formula, It is a material Book inventory at the end of day t It is a material Available inventory at the end of day t. It is a material Book inventory at the end of day t-1 It is a material Total inbound quantity at the end of day t (from purchases received, returns, etc.) It is a material Sales outbound volume at the end of day t It is a material Other outbound quantities at the end of day t (such as scrapped or transferred items). It is a material The amount of inventory locked by sales orders or production plans at the end of day t (from locked inventory records). It is a material The quantity of purchases that have not yet arrived but are in transit by the end of day t (from the in-transit purchase record).

[0033] The conversion to the unified inventory unit rule set in the ERP is as follows:

[0034] In the formula, This refers to the volume of business after standardizing the inventory unit. It is the quantity in the original record. It is a material The unit conversion factor is read from the material master data in the ERP system, and its value is determined by the preset conversion relationship between the material's purchasing unit and inventory unit.

[0035] It should be noted that each row of the material dynamic status basic dataset contains the business volume after unifying the inventory unit. Book inventory and available inventory .

[0036] S1.3: Based on the material dynamic state dataset, calculate the key operating state characteristics of each material to form a material dynamic state vector.

[0037] Specifically, based on the material dynamic status dataset, the average daily demand, demand standard deviation, inventory coverage days, average procurement lead time, procurement lead time standard deviation, stockout frequency index, and inventory aging index are calculated for each material to form a material dynamic status vector.

[0038] The formula for calculating the average daily demand is as follows:

[0039] In the formula, It is a material The average daily demand on day t It is a material In the Actual daily demand It is a preset statistical period.

[0040] The formula for calculating the standard deviation of demand is:

[0041] In the formula, It is the standard deviation of demand.

[0042] The formula for calculating the inventory coverage days is:

[0043] In the formula, It is the number of days the inventory covers. It is a material Available inventory at the end of day t. It is a very small positive number to prevent the denominator from being zero.

[0044] The average procurement lead time is calculated by extracting completed material purchase order records from the original business records, calculating the number of days from order placement to actual delivery for each order, and then averaging the results. The calculation formula is as follows:

[0045] In the formula, It is the average procurement lead time. is the actual lead time (delivery date - order date) for the m-th purchase, and N is the fixed number of samplings.

[0046] The formula for calculating the standard deviation of the procurement lead time is:

[0047] In the formula, It is the standard deviation of procurement lead time.

[0048] The out-of-stock frequency indicator is calculated by statistically analyzing data within a preset statistical period. Available inventory The number of days when demand is less than the daily demand (i.e., the number of days when there is a stockout) is calculated by dividing by a preset statistical period. The formula is as follows:

[0049] In the formula, It is an indicator of the frequency of stockouts. It represents the number of days that the stockout occurred.

[0050] The inventory aging index is calculated by obtaining the receipt date of each batch of material inventory from the ERP inventory ledger, dividing the inventory into three intervals according to age, and calculating the inventory aging index based on the inventory level in each interval:

[0051] In the formula, It is an indicator of inventory aging. It is the first Inventory levels for each age range It is the first The representative age of each age range (taking the midpoint of the range). It is a very small positive number to prevent the denominator from being zero.

[0052] The material dynamic state vector is expressed as: .

[0053] It should be noted that the inventory age is divided into three intervals, namely, , , , This is the upper limit of the first interval. In the ERP system, after the goods are purchased, they are usually sold or consumed within 30 days. Inventory that has not been used for more than 30 days begins to face the risk of slow sales. In combination with this implementation scenario (high-frequency business of purchasing, sales and inventory, such as electronic components / fast-moving consumer goods), 30 days can accurately distinguish between "active inventory" and "potential backlog". This is the upper limit of the second range. Corporate financial systems usually classify "inventory that has not been turned over for more than 90 days" as obsolete materials that require special analysis. At the same time, mainstream ERP (such as Oracle SCM) inventory health reports default to 90 days as the critical point.

[0054] S2. By constructing an enhanced context sequence that integrates historical demand sequences and future business event markers, and combining the material dynamic state vector input to a multi-expert gated demand forecasting model, demand forecasting results are generated. Based on expert scale parameters and disturbance reconstruction inference, data uncertainty and model uncertainty are calculated and calibrated respectively. Through historical normalized residual quantile calibration, comprehensive demand disturbance parameters are obtained. S2.1: Extract historical actual daily demand from the ERP database to construct a historical sequence, and combine it with future promotions, holidays and peak season markers to form an enhanced contextual sequence along the timeline that includes historical observations and future known business knowledge.

[0055] Specifically, the actual daily demand within a preset statistical period W is extracted from the Enterprise Resource Planning (ERP) database to construct a basic historical demand sequence. At the same time, the ERP system automatically generates a known business event tagging sequence within a future forecast period H based on the business calendar and activity configuration, and constructs a corresponding business knowledge vector for any future date. In order to enable the model to utilize both historical demand information and future known business knowledge, the basic historical demand sequence and the known business event tagging sequence are constructed into an enhanced context sequence along the time axis.

[0056] The business event markers include at least promotion markers, holiday markers, and peak season markers, which are determined by the ERP system based on existing business plans, holiday calendars, and seasonal configurations. These are future known business knowledge that can be obtained in advance.

[0057] The enhanced context sequence is specifically as follows: For historical time steps The input vector is defined as:

[0058] In the formula, It is the input vector. It is a material In the Actual daily demand It is a material In the Daily promotional tags, It is a material In the Holiday markings for the day, It is a material In the The peak season marker for the day.

[0059] For future time steps Since the actual demand has not yet occurred at this moment, a preset mask placeholder is used to replace the demand input, and it is concatenated with the corresponding business knowledge vector to obtain:

[0060] In the formula, It is a preset mask placeholder.

[0061] The enhanced context sequence is formed as follows:

[0062] In the formula, It is an enhanced context sequence. It is a material The input vector in the future t+H days.

[0063] It should be noted that H, based on an in-depth analysis of typical procurement lead times in Enterprise Resource Planning (ERP) systems, extracted a large number of completed procurement records from the company's historical purchase order data. For each record, its "procurement lead time" (the number of days from order placement to delivery) was calculated. It was found that the procurement lead time for the vast majority of materials (covering more than 90% of conventional materials) was concentrated in the range of 5 to 10 days. In order to ensure that the forecast range can fully cover the longest common lead time and leave a certain margin to cope with minor delays, the forecast period was set to 14 days. This can ensure the effectiveness of the forecast while avoiding unnecessary errors introduced by an excessively long forecast period.

[0064] S2.2: The enhanced context sequence and the material dynamic state vector are input into the multi-expert gated demand forecasting model. The joint hidden state is extracted by forward and reverse GRU and concatenated with the material dynamic state vector to form a joint feature vector. The gating network calculates the weights of each expert based on this vector, and the weighted fusion of the forecasts output by each expert generates the future demand forecast.

[0065] Specifically, enhanced context sequences and material dynamic state vector A multi-expert gated demand forecasting model is jointly input to generate demand forecasting results for future forecasting periods. The multi-expert gated demand prediction model includes a time-series coding unit, an expert prediction unit, and a gated fusion unit; wherein, the time-series coding unit employs a bidirectional gated recurrent unit network to enhance the context sequence. Encoding is performed to extract time-series features resulting from the combined effects of historical demand information and known future business knowledge. The expert prediction unit includes multiple prediction experts configured in parallel, which learn the prediction mapping relationship under different demand patterns and output the predicted mean and prediction scale parameters of future demand. The gating fusion unit calculates the gating weights of each prediction expert based on the joint feature vector, so that different demand patterns are processed by more suitable prediction experts, thereby improving the model's adaptability to stable demand, seasonal demand, activity-driven demand, and fluctuating demand, and generating future demand forecasts based on the output results of each prediction expert.

[0066] Based on enhanced context sequence The input vector is encoded by a bidirectional gated recurrent unit network along both the forward and reverse directions to obtain the joint hidden state. The joint hidden state corresponding to the last time step of the enhanced context sequence is taken as the temporal feature representation and combined with the material dynamic state vector. The features are concatenated to form a joint feature vector. The joint feature vector is then simultaneously input into K parallel prediction experts and gating fusion units. The prediction experts output the predicted mean and prediction scale parameters of future demand. The gating fusion units calculate the gating weights of each prediction expert based on the joint feature vector. Based on the gating weights of each prediction expert and the predicted mean of future demand, the predicted amount of future demand is calculated.

[0067] The formula for calculating the joint hidden state is: In the formula, It is time A step The positive hidden state. It is a time step The reverse hidden state, It is a time step The joint hidden state, It is a forward-gated loop unit. It is a reverse-gated loop unit. It is a vector concatenation operation.

[0068] The joint feature vector expression is:

[0069] In the formula, It is a joint eigenvector. It is the joint hidden state corresponding to the last time step of the enhanced context sequence.

[0070] The gating weights of each prediction expert are calculated using the following formula:

[0071] In the formula, It is the first The gating weight of each expert, It is the k-th expert score output by the gating network, which is the output of the A-gated neural network with the cumulative predicted demand, available inventory, procurement lead time deviation, and external disturbance factors of the material at time t as input. The first output of the gating network Each expert rating value It is the temperature parameter, and K is the total number of experts.

[0072] The formula for calculating the future demand forecast is as follows:

[0073] In the formula, It is a material In the future Daily demand The Kth expert's prediction for the future The average forecast of daily demand.

[0074] It should be noted that the forecast scale parameter does not directly participate in the weighted summation calculation of future demand forecasts, but is used to construct subsequent demand disturbance parameters or forecast confidence parameters, so that the forecast model can output scale information related to demand fluctuations in addition to the output point forecast results.

[0075] K is configured according to the demand pattern clustering, with a typical value range of [3,8]. The preferred value is 5, based on the fact that if it is less than 3, the expert diversity is insufficient and it is difficult to cover the demand pattern (such as trend type, periodic type, sudden type); if it is more than 8, the computational cost will increase and it is easy to overfit (when K>6 in actual tests, the validation set error no longer decreases significantly).

[0076] Determined by the training / tuning process, the typical value range is (0.1, 0.8). The optimal value is found by back-searching to minimize the prediction error (such as MAE / MSE) or inventory service rate index on the validation set. The preferred value is 0.35. In this implementation scenario, material demand fluctuates drastically and the division of labor among experts is clear, requiring strong gating discrimination. When the value is 0.35, the average gating entropy is 1.2, which avoids excessive concentration (single expert monopoly) while ensuring the dominance of key experts.

[0077] S2.3: Using the actual future demand within the historical window as the supervision signal, a Gaussian mixture prediction distribution weighted by gating weights is constructed. The negative log-likelihood loss function is adopted, and the multi-expert gating demand prediction model is jointly optimized by backpropagation and gradient descent.

[0078] Specifically, the actual future demand within a historical window is used as a monitoring signal to jointly train the time-series coding unit, expert prediction unit, and gating fusion unit; let the actual future demand corresponding to the training samples be denoted as . Then the mean prediction and prediction scale parameter output by each expert are: The model training loss function is defined as negative log-likelihood loss based on the predicted distribution of future demand. The model parameters in the temporal coding unit, expert prediction unit and gating fusion unit are backpropagated and differentiated based on the loss function. The model parameters are iteratively updated using gradient descent or its variant until the loss function converges or reaches the preset number of training rounds, thereby obtaining the trained multi-expert gating demand prediction model.

[0079] The formula for calculating the predicted distribution of future demand is:

[0080] In the formula, It is a predicted distribution of future demand. It is a Gaussian probability density function.

[0081] The formula for calculating the negative log-likelihood loss is as follows:

[0082] In the formula, It is the negative log-likelihood loss, i.e., the loss function. This is the number of training samples.

[0083] S2.4: Based on the output of the multi-expert gated demand forecasting model, calculate the data uncertainty and model uncertainty separately, and add them together to obtain the total uncertainty. Then, utilize the historical normalized residuals... The quantile determines the calibration coefficient, the total uncertainty is calibrated and averaged to generate the comprehensive demand disturbance parameter, and the cumulative and average predicted demand within the prediction period are calculated simultaneously.

[0084] Specifically, to measure the reliability of the forecast results, data uncertainty and model uncertainty are calculated separately for the future forecast demand results. The data uncertainty and model uncertainty are then added together to obtain the total uncertainty. To further improve the robustness of the uncertainty measurement, a normalized residual is defined for any date within the historical window. All normalized residuals within the historical window are sorted in ascending order, and the lowest value is selected. The calibration coefficients are obtained from the quantiles. Multiplying these calibration coefficients by the total uncertainty yields the future calibrated demand disturbance parameters, which are then used to obtain the comprehensive demand disturbance parameters for the entire forecast period. The comprehensive demand disturbance parameter is obtained by averaging the future calibrated demand disturbance parameters; and based on the future predicted demand, the cumulative predicted demand and the average predicted demand within the prediction period are calculated.

[0085] The data uncertainty is used to characterize the degree of random fluctuation in the demand data itself, and is obtained by weighting the scaling parameters output by each expert:

[0086] In the formula, Is the future number The uncertainty of the data for the day It is the prediction scale parameter output by the Kth expert.

[0087] The model uncertainty is used to characterize the model's cognitive stability to the current input state. During the inference phase, the joint feature vector is perturbed and reconstructed to obtain the predicted point values ​​for each set of perturbed reconstructions. The model uncertainty is then calculated based on the sample standard deviation of each set of predicted point values.

[0088] In the formula, Is the future number The uncertainty of the model, It is the point prediction value obtained from the m-th perturbation inference. It is the average of the predicted values ​​of M points, where M is the total number of disturbances.

[0089] The formula for calculating the normalized residual is as follows:

[0090] In the formula, It is the normalized residual. It is a material In the The actual daily demand It is a material In the Daily retrospective forecast of demand. It is the first The total uncertainty corresponding to the day, It is a very small positive number that prevents the denominator from being zero.

[0091] The formula for calculating the calibration coefficient is:

[0092] In the formula, It is the calibration coefficient. yes Quantile function, It is a preset significance level.

[0093] The formula for calculating the cumulative forecasted demand is as follows:

[0094] In the formula, It is the cumulative forecast demand. It is for materials In the future The forecast value for the demand per day, where H is the total number of forecast days.

[0095] The formula for calculating the average predicted demand is as follows:

[0096] In the formula, It is the average forecast demand.

[0097] It should be noted that, The threshold calibration method based on business KPI back-engineering is set with a typical value range of [0.01, 0.1], and the preferred value is 0.04. It is based on the dual objectives of historical out-of-stock rate constraint (≤3%) and minimizing inventory cost increment. The results are determined by evaluating the quantile calibration effect on the validation set. It is suitable for prediction calibration scenarios in multi-category manufacturing enterprises that take into account both service reliability and inventory efficiency.

[0098] S3. Construct a segmented mapping risk index based on comprehensive demand disturbance parameters, dynamically determine the target service level and the corresponding safety factor, and calculate the optimization control parameters to form a dynamic optimization control parameter set. S3.1: Based on the comprehensive demand disturbance parameters and the standard deviation of procurement lead time, the risk index is obtained by weighted calculation in combination with the average predicted demand and lead time length. The target service level is determined by segmented mapping through preset thresholds according to the risk index, and the corresponding dynamic safety coefficient is obtained by calling the standard normal distribution quantile function.

[0099] Specifically, to enable inventory control intensity to adapt to demand disturbances and supply fluctuations, based on comprehensive demand disturbance parameters... and standard deviation of procurement lead time Construct a current risk index for the material, and determine the target service level based on the risk index using a segmented mapping method. Then, find the standard normal distribution quantile value based on the target service level to obtain the dynamic safety factor.

[0100] The formula for calculating the risk index is as follows:

[0101] In the formula, It is a material The comprehensive risk index at time t, It is a risk weighting coefficient. It is a very small positive number, used to prevent the denominator from being zero. Used to characterize the relative degree of disturbance on the demand side. Both are used to characterize the relative volatility of the supply side, and together they reflect the level of stockout risk of materials under the current business conditions.

[0102] The formula for calculating the target service level is:

[0103] In the formula, It is a material The target service level at time t It is the risk classification threshold, and These correspond to target service levels under low-risk, medium-risk, and high-risk conditions, respectively, to meet the following requirements. .

[0104] The formula for calculating the dynamic safety factor is as follows:

[0105] In the formula, It is a dynamic safety factor. It is the quantile function of the standard normal distribution.

[0106] It should be noted that, The weights are determined based on cost sensitivity or the Analytic Hierarchy Process (AHP) to ensure that the sum of the weights equals 1. The preferred values ​​are 0.55 and 0.45, respectively, which are derived from the statistical analysis of the contribution rates of the demand and supply sides in historical shortage events.

[0107] Based on the quantile segmentation method of historical risk distribution, the typical value ranges are [0.3, 0.6] and [0.7, 1.2], and the preferred values ​​are 0.42 and 0.78, respectively. This is based on the fact that actual test data has verified that it can effectively distinguish three types of risk levels and match business resource allocation strategies.

[0108] S3.2: Based on the dynamic safety factor, calculate the dynamic safety stock, dynamic replenishment point, replenishment trigger value, target inventory limit, and suggested replenishment quantity after correction by the maximum replenishment quantity constraint in sequence to form a dynamic optimization control parameter set.

[0109] Specifically, based on dynamic safety factor Based on cumulative forecast demand Average forecast demand and comprehensive demand disturbance parameters Combined with average procurement lead time Standard deviation of procurement lead time Available inventory In addition to the in-transit inventory, calculate the dynamic safety stock parameters, dynamic replenishment point parameters, replenishment trigger parameters, and target inventory limit parameters, and calculate the basic suggested replenishment quantity. Based on the preset maximum replenishment processing quantity parameter, the final output is the constraint-corrected replenishment quantity parameter. The dynamic safety stock parameters, dynamic replenishment point parameters, replenishment trigger parameters, target inventory limit parameters, and constraint-corrected replenishment quantity parameters are used as dynamic optimization control parameters.

[0110] The dynamic safety stock parameter is used to address the stockout risk caused by demand fluctuations and uncertainties in supply lead times. The calculation formula is as follows:

[0111] In the formula, It is a dynamic safety stock parameter. It is the safety factor corresponding to the service level.

[0112] The dynamic replenishment point parameter represents the inventory threshold that triggers a purchasing action, and the calculation formula is as follows:

[0113] In the formula, These are dynamic replenishment point parameters. It is the expected demand consumption within the average procurement lead time, plus safety stock, which constitutes the complete replenishment trigger benchmark.

[0114] The replenishment trigger parameter is used to determine in real time whether the replenishment process should be initiated, and the calculation formula is as follows:

[0115] In the formula, It is a replenishment trigger parameter. This refers to inventory in transit.

[0116] The target inventory ceiling parameter is used to prevent inventory backlog caused by over-purchasing, and the calculation formula is as follows:

[0117] In the formula, This is the target inventory limit parameter.

[0118] The basic recommended replenishment quantity is calculated using the following formula:

[0119] In the formula, This is the basic recommended replenishment quantity.

[0120] The formula for calculating the replenishment quantity parameter after constraint correction is as follows:

[0121] In the formula, This is the replenishment quantity parameter after constraint correction. This is the preset maximum replenishment processing capacity parameter.

[0122] It should be noted that, The system reads the following in real time from the ERP master data: the maximum quantity for a single purchase (defined by the "maximum order quantity" field in the material master data); the maximum storage capacity (calculated from the "current available storage capacity" in the warehouse master data and the material volume / weight); and the maximum budget mapping limit (dynamically calculated by the finance module based on the current purchase budget balance and the material unit price). It then compares these limits and automatically selects the minimum value as the maximum quantity. .

[0123] S4. Write the dynamic optimization control parameter set into the ERP control table to drive the generation of inventory early warning and procurement tasks, and collect the parameter execution feedback results for analysis and optimization management.

[0124] S4.1: Write the set of dynamic optimization control parameters into the ERP material inventory control master data table, generate purchase requisition records based on replenishment trigger parameters, and write early warning or adjustment prompt records when the inventory age exceeds the limit or the inventory exceeds the limit.

[0125] Specifically, the dynamic optimization control parameters are written into the preset "Material Inventory Control Master Data Table" in the ERP system. When the replenishment trigger parameter in the dynamic optimization control parameters is triggered... When the value is greater than 0, it indicates that the sum of the current available inventory and the inventory in transit is lower than the dynamic replenishment point, meeting the replenishment trigger condition. The ERP system's procurement task processing module (such as the MRP engine) automatically generates a purchase requisition record, which is stored in the "Purchase Requisition Pending Processing List" as the input source for subsequent purchase order creation. For any material, if the inventory age of any batch exceeds the preset aging threshold, the ERP inventory warning module automatically generates a high inventory age warning record and writes it to the "Inventory Anomaly Warning Log Table". When the available inventory... Greater than the target inventory limit parameter When this happens, the ERP system will automatically generate an inventory adjustment prompt record, suggesting transfers, promotions, or suspension of purchases. The record will be stored in the "Inventory Optimization Suggestion Table" for planners to refer to, but it is not mandatory.

[0126] The purchase requisition record includes the material code, the required date, and the replenishment quantity parameters after constraint correction. .

[0127] It should be noted that the default "Material Inventory Control Master Data Table" is a standard ERP functional component (such as the MRP view in SAP or the Item Sourcing table in Oracle), used to store the inventory strategy parameters for each material.

[0128] The generated purchase requisition records, inventory warning records, and inventory adjustment prompt records are all data objects or transaction records within the ERP system. Their generation is entirely driven by system rules and does not include any restrictions on manual approval processes or business strategies.

[0129] The preset aging threshold is set by the inventory age-consumption probability regression analysis method, with a typical value range of [160~190] days and a preferred value of 180 days. This is based on the use of a typical sample library of Chinese manufacturing industry (covering machinery, electronics and chemical industries) to conduct a global inventory age-consumption rate regression analysis to obtain the average consumption probability curve. On this curve, the inventory age point at which the consumption probability first stably falls below 20% is identified as 182 days. In order to balance the simplicity of system implementation and business interpretability, this value is rounded down to the nearest whole number of ten, i.e., 180 days.

[0130] S4.2: Collect actual execution feedback data from the ERP database to generate business records.

[0131] Specifically, actual execution feedback data for generating business records is collected from the ERP database, including: the final delivery quantity and date of the purchase requisition, the daily actual sales and outbound quantity of materials, the available inventory at the end of each day, and the number of stockouts recorded by the ERP stockout log module.

[0132] S4.3: Based on actual execution feedback data, calculate prediction deviation index, stockout deviation index, and inventory deviation index, weight and merge the deviation indexes, calculate the comprehensive update trigger value, and set update thresholds for optimization management judgment.

[0133] Specifically, based on actual execution feedback data, prediction deviation, stockout deviation, and inventory deviation indicators are calculated. These deviation indicators are then weighted and fused to calculate a comprehensive update trigger value. An update threshold is set. When the comprehensive update trigger value exceeds the update threshold, it is determined that the current model and parameters are no longer effectively adapted to the current business situation and must be updated, automatically triggering the following operations: Initiate an incremental training process for the artificial intelligence demand forecasting model, utilizing newly generated historical sales data, procurement arrival data, and market activity data since the last model update to perform online learning and parameter fine-tuning of the artificial intelligence model used to generate average forecasted demand, cumulative forecasted demand, and demand disturbance parameters; After the update is complete, the new model parameter set and inventory control parameter set will recalibrate the statistical parameters in the formula used to calculate dynamic safety stock, and update the calculation benchmarks for average procurement lead time and standard deviation of procurement lead time based on the latest procurement lead time performance data.

[0134] The comprehensive update trigger value is calculated using a variant of the Mean Absolute Percentage Error (MAPE) commonly used in the supply chain field to measure the accuracy of demand forecasting.

[0135] In the formula, It is a comprehensive update trigger value. It is for materials In the future The forecast value made based on the daily demand. It is a material In the future The actual demand for the day It is a very small positive number to prevent the denominator from being zero.

[0136] The stockout deviation index is used to measure the effectiveness of inventory strategies in preventing stockouts, and the calculation formula is as follows:

[0137] In the formula, It is a stockout deviation indicator. It represents the number of days that the stockout occurred.

[0138] The inventory deviation index is used to measure the degree of deviation between the actual inventory level and the target control level, and the calculation formula is as follows:

[0139] In the formula, It is an inventory deviation indicator. This is the actual inventory level (from the ERP inventory ledger). It is a material In the future Daily target inventory limit parameter This is the evaluation completion period.

[0140] The formula for calculating the comprehensive update trigger value is as follows:

[0141] In the formula, It is a comprehensive update trigger value. It is the weighting coefficient.

[0142] It should be noted that the fixed value setting method for the update threshold typically ranges from [0.2, 0.3], with a preferred value of 0.25. This is based on the fact that 0.25 is a widely adopted rule of thumb in the industry. For example, in mainstream supply chain planning software such as SAP IBP (Integrated Business Planning), when the prediction accuracy (1-MAPE) is below 75% (i.e., MAPE>25%), the system will suggest or automatically trigger a model review process.

[0143] By setting fixed values, typical value ranges are [0.3, 0.5], [0.3, 0.5], and [0.1, 0.3], satisfying the constraint that the sum is 1. The preferred values ​​are 0.4, 0.4, and 0.2, respectively. This is based on the summary of the SCOR (Supply Chain Operations Reference) model and a large number of manufacturing ERP implementation cases. In order to balance the two core objectives of "cost" and "service", a "service-first" weight allocation strategy is generally adopted. That is, the two indicators that directly affect the service level, forecasting and stockouts, are given a total weight of 80%, while the weight of inventory cost control is set at 20%.

[0144] This embodiment also provides an AI-based dynamic optimization management system for ERP inventory management, including: The data acquisition standardization module is used to collect various types of original business records from the ERP system, and after unification of coding, unit conversion, time alignment and cleaning, form a basic dataset. The material status vectorization module is used to calculate a seven-dimensional feature vector based on standardized data, including average daily demand, demand standard deviation, inventory coverage days, lead time and its fluctuation, stockout frequency, and inventory aging, to construct a feature vector that represents the real-time operating status of materials. The enhanced context modeling module is used to integrate historical daily demand with event markers such as future promotions / holidays to construct a joint historical-future input sequence with masked placeholders; The multi-expert gating prediction module is used to generate probabilistic demand predictions with scale parameters by dynamically weighting the outputs of multiple experts through a gating network using BiGRU encoding of joint sequences. The risk-driven control parameter generation module is used to construct a segmented mapping risk index based on demand disturbances and procurement lead time fluctuations, dynamically determine the safety factor, and calculate a complete set of inventory control parameters.

[0145] In summary, this invention significantly improves the adaptive control capabilities and inventory management efficiency of the ERP inventory management system in complex supply and demand environments by: constructing a dynamic state vector of materials by integrating multi-source business data from ERP systems; generating demand disturbance parameters calibrated for dual-source uncertainty by combining enhanced context sequences of future business events with a multi-expert gating prediction model; constructing a risk index based on these parameters and procurement lead time fluctuations to drive the dynamic grading of target service levels and safety factors; and calculating safety stock, replenishment points, and recommended replenishment quantities in real time. Finally, the control parameters are written into the ERP master data table to automatically trigger procurement and early warnings. Simultaneously, closed-loop iterative optimization of the model and strategy is achieved based on feedback from actual execution deviations.

[0146] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. 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 be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A dynamic optimization management method for an ERP inventory management system based on artificial intelligence, characterized in that, include: Collect ERP sales and inventory data and standardize it to form a basic dataset of material dynamic status. Extract operational status features based on the basic dataset of material dynamic status and construct material dynamic status vectors. By constructing an enhanced context sequence that integrates historical demand sequences and future business event markers, and combining the material dynamic state vector input to a multi-expert gated demand forecasting model, demand forecasting results are generated. Based on expert scale parameters and disturbance reconstruction inference, data uncertainty and model uncertainty are calculated and calibrated respectively. Through historical normalized residual quantile calibration, comprehensive demand disturbance parameters are obtained. Based on the comprehensive demand disturbance parameters, a segmented mapping risk index is constructed, the target service level and the corresponding safety factor are dynamically determined, and the optimization control parameters are calculated to form a dynamic optimization control parameter set. The set of dynamically optimized control parameters is written into the ERP control table to drive the generation of inventory warnings and procurement tasks, and the results of parameter execution feedback are collected for analysis and optimization management.

2. The method for dynamic optimization management of an ERP inventory management system based on artificial intelligence as described in claim 1, characterized in that: The collection of ERP inventory management data refers to the collection of original business records related to the target material within a preset statistical period from the ERP system already deployed by the enterprise.

3. The method for dynamic optimization management of an ERP inventory management system based on artificial intelligence as described in claim 2, characterized in that: The standardization process to form a basic dataset of material dynamic status refers to the standardization process of unifying material codes, unifying units of measurement, aligning timestamps, and cleaning data from the original business records to form a basic dataset of material dynamic status.

4. The dynamic optimization management method for an ERP inventory management system based on artificial intelligence as described in claim 3, characterized in that: The process of extracting operational status features based on the material dynamic status dataset and constructing a material dynamic status vector refers to calculating key operational status features for each material based on the material dynamic status dataset to form a material dynamic status vector.

5. The dynamic optimization management method for an ERP inventory management system based on artificial intelligence as described in claim 4, characterized in that: The acquisition of comprehensive demand disturbance parameters refers to extracting historical actual daily demand from the ERP database to construct a historical sequence, and combining it with future promotions, holidays and peak season markers to form an enhanced context sequence along the time axis that includes historical observations and future known business knowledge. The enhanced context sequence and the material dynamic state vector are input into the multi-expert gated demand forecasting model. The joint hidden state is extracted by forward and reverse GRU and concatenated with the material dynamic state vector to form a joint feature vector. The gating network calculates the weight of each expert based on the joint feature vector, and the weighted fusion of the forecasts output by each expert generates the future demand forecast. Using the actual future demand within the historical window as the supervision signal, a Gaussian mixture prediction distribution weighted by gating weights is constructed. The negative log-likelihood loss function is adopted, and the multi-expert gating demand prediction model is jointly optimized by backpropagation and gradient descent. Based on the output of the multi-expert gated demand forecasting model, the data uncertainty and model uncertainty are calculated separately, and then summed to obtain the total uncertainty. The total uncertainty is then calculated using historical normalized residuals. The quantile determines the calibration coefficient, the total uncertainty is calibrated and averaged to generate the comprehensive demand disturbance parameter, and the cumulative and average predicted demand within the prediction period are calculated simultaneously.

6. The dynamic optimization management method for an ERP inventory management system based on artificial intelligence as described in claim 5, characterized in that: The formation of the dynamic optimization control parameter set refers to the weighted calculation based on the comprehensive demand disturbance parameters and the standard deviation of the procurement lead time, combined with the average predicted demand and the lead time length, to obtain the risk index. The target service level is determined by segmenting and mapping according to the risk index through a preset threshold, and the corresponding dynamic safety coefficient is obtained by calling the standard normal distribution quantile function. Based on the dynamic safety factor, the dynamic safety stock, dynamic replenishment point, replenishment trigger value, target inventory limit, and suggested replenishment quantity after correction by the maximum replenishment quantity constraint are calculated in sequence to form a dynamic optimization control parameter set.

7. The dynamic optimization management method for an ERP inventory management system based on artificial intelligence as described in claim 6, characterized in that: The step of writing the dynamic optimization control parameter set into the ERP control table to drive the generation of inventory warnings and procurement tasks refers to writing the dynamic optimization control parameter set into the ERP material inventory control master data table, generating procurement application records based on replenishment trigger parameters, and writing warning or adjustment prompt records when the inventory age exceeds the limit or the inventory exceeds the limit.

8. The method for dynamic optimization management of an ERP inventory management system based on artificial intelligence as described in claim 7, characterized in that: The execution feedback result of the collected parameters refers to the actual execution feedback data collected from the ERP database to generate business records.

9. The dynamic optimization management method for an ERP inventory management system based on artificial intelligence as described in claim 8, characterized in that: The aforementioned analysis and optimization management refers to calculating prediction deviation indicators, stockout deviation indicators, and inventory deviation indicators based on actual execution feedback data, weighting and fusing the deviation indicators, calculating a comprehensive update trigger value, and setting an update threshold for optimization management judgment.

10. A dynamic optimization management system for an ERP inventory management system based on artificial intelligence, based on the dynamic optimization management method for an ERP inventory management system based on artificial intelligence as described in any one of claims 1 to 9, characterized in that, include: The data acquisition standardization module is used to collect various types of original business records from the ERP system, and after unification of coding, unit conversion, time alignment and cleaning, form a basic dataset. The material status vectorization module is used to calculate a seven-dimensional feature vector based on standardized data, including average daily demand, demand standard deviation, inventory coverage days, lead time and its fluctuation, stockout frequency, and inventory aging, to construct a feature vector that represents the real-time operating status of materials. The enhanced context modeling module is used to integrate historical daily demand with event markers such as future promotions / holidays to construct a joint historical-future input sequence with masked placeholders; The multi-expert gating prediction module is used to generate probabilistic demand predictions with scale parameters by dynamically weighting the outputs of multiple experts through a gating network using BiGRU encoding of joint sequences. The risk-driven control parameter generation module is used to construct a segmented mapping risk index based on demand disturbances and procurement lead time fluctuations, dynamically determine the safety factor, and calculate a complete set of inventory control parameters.

Images