Mobile phone accessory automatic replenishment method and system based on dynamic inventory

Abstract

The application relates to the technical field of supply chain inventory management, and discloses a mobile phone accessory automatic replenishment method and system based on dynamic inventory, which comprises the following steps: acquiring mobile phone accessory in-out record sequences, seasonal rotation characteristics and life cycle characteristics; performing time sequence memory extraction on the in-out record sequences to obtain an initial demand trend; performing clustering division and correction according to the initial demand trend, the seasonal rotation characteristics and the life cycle characteristics to obtain a target consumption rhythm; comparing the target consumption rhythm with a preset threshold to generate a replenishment cycle; and generating a replenishment triggering instruction after integrated decision evaluation. The method can realize dynamic self-adaptive control of inventory allocation, and can satisfy the reliability of supply-demand matching under complex demand changes.

Description

Technical Field

[0001] This invention relates to the field of supply chain inventory management technology, and in particular to a method and system for automatic replenishment of mobile phone accessories based on dynamic inventory. Background Technology

[0002] Currently, in the field of supply chain inventory management technology, how to achieve dynamic inventory control of replenishment timing through industrial data processing has become a key research topic for preventing stockouts and overstocking.

[0003] In one existing technology, a fixed-period inventory replenishment method is mainly used to ensure the stability of inventory turnover. This method relies primarily on historical total business volume or a single replenishment rule for evaluation. For example, a preset safety stock line mode is used to execute automatic replenishment tasks, and the demand for goods is judged by the accumulation of simple shipment volume. However, this approach is clearly insufficient in complex sales environments. Fixed rules cannot adapt to the dynamic changes in the product life cycle, and the evaluation of a single indicator is easily affected by seasonal changes and new product updates. It is also affected by the large number of accessory models and the increased randomness of demand fluctuations. Especially in complex scenarios where demand alternates between high and low, it is difficult to avoid replenishment decision errors caused by the continuous drift of historical shipment volume values, resulting in low overall operational efficiency.

[0004] In summary, existing technologies are insufficient to achieve dynamic and adaptive control of mobile phone accessory inventory allocation, and cannot meet the core requirement of supply chain inventory management to maintain reliable supply-demand matching when facing complex demand changes. Summary of the Invention

[0005] This invention provides a method and system for automatic replenishment of mobile phone accessories based on dynamic inventory, so as to realize dynamic adaptive control of the mobile phone accessory inventory allocation process and meet the core requirement of supply chain inventory management to maintain the reliability of supply and demand matching when facing complex demand changes.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for automatic replenishment of mobile phone accessories based on dynamic inventory, comprising: Obtain the entry and exit record sequence of mobile phone accessories and perform time-series memory retrieval processing to obtain the initial demand trend; Based on the initial demand trend, pre-extracted seasonal rotation features, and life cycle features, clustering calculations are performed to obtain the initial consumption rhythm; Calculate the fluctuation range of the initial consumption rhythm; if the fluctuation range does not meet the preset stability determination condition, perform smoothing correction processing to obtain the target consumption rhythm; if the stability determination condition is met, use the initial consumption rhythm as the target consumption rhythm. If the target consumption rate is greater than the preset out-of-stock determination threshold, a first replenishment cycle in the out-of-stock state is generated; if the target consumption rate is less than the preset backlog determination threshold, a second replenishment cycle in the backlog state is generated; if the target consumption rate is within the normal range, the current state is maintained. An integrated decision evaluation is performed on the first replenishment cycle and the second replenishment cycle to obtain a target replenishment decision; the target cumulative value is extracted based on the target replenishment decision and the actual consumption rhythm is verified. If the verification is successful, a replenishment trigger instruction is generated to execute inventory allocation.

[0007] Secondly, the present invention provides an automatic replenishment system for mobile phone accessories based on dynamic inventory, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the automatic replenishment method for mobile phone accessories based on dynamic inventory as described above.

[0008] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the automatic replenishment method for mobile phone accessories based on dynamic inventory as described above.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention obtains the entry and exit record sequence of mobile phone accessories, performs sliding interception and implicit gating update calculation to obtain the hidden layer time sequence vector, and obtains the initial demand trend through nonlinear dimension mapping. Then, the initial demand trend is concatenated with the pre-extracted seasonal rotation features and life cycle features to construct a multi-dimensional fusion feature matrix. After clustering and iterative calculation, convergent classification clusters are obtained and the initial consumption rhythm is mapped. This invention breaks through the limitations of traditional fixed rules that rely on historical total shipment volume or single replenishment rules, which cannot adapt to the dynamic changes of seasonal rotation and product life cycle. It integrates the temporal memory extraction capability of long short-term memory network and multi-source business feature clustering to accurately depict the dynamic consumption pattern of different models of accessories under the dual influence of season and life cycle. It solves the problem that the historical value of shipment volume continuously drifts with the season and product life cycle, causing the fixed replenishment judgment to fail and the demand prediction to be distorted. It provides an accurate demand trend basis for the stability correction of consumption rhythm and replenishment decision.

[0010] (2) This invention calculates the fluctuation range of the initial consumption rhythm and performs smoothing correction when the preset stability judgment condition is not met to obtain the target consumption rhythm. If the target consumption rhythm is greater than the shortage judgment threshold, the actual available inventory is calculated by combining the physical inventory balance and the replenishment quantity in transit and the estimated depletion time to generate the replenishment cycle for the shortage state. If it is less than the backlog judgment threshold, the inventory turnover index is deduced by combining the stagnant inventory and the warehouse occupancy status and the inventory consumption rate is simulated to generate the second replenishment cycle for the backlog state. This invention breaks through the limitations of the traditional fixed safety stock line, which cannot distinguish between the shortage and backlog states and the rigid replenishment cycle. Based on the premise of fluctuation range stability correction, it adaptively generates a targeted replenishment cycle according to the differentiated consumption rhythm threshold, accurately distinguishes the two different operational risks of shortage and backlog and formulates differentiated replenishment strategies to solve the problems of shortage during strong demand and backlog after demand weakens, and the misalignment between replenishment actions and the actual consumption rhythm.

[0011] (3) This invention calculates the difference between the first and second replenishment cycles and constructs a decision feature vector. It inputs the integrated decision model to perform node split evaluation and leaf node weighted voting statistics to obtain the target replenishment decision. Then, it extracts the target cumulative value and verifies the actual consumption rhythm. After verification, it generates a replenishment trigger instruction to execute inventory allocation. It constructs a closed-loop system of dynamic inventory allocation with a full link, from multi-dimensional time series feature extraction, cluster consumption rhythm division, fluctuation stabilization correction, shortage and backlog status judgment to integrated decision verification and trigger execution. This effectively improves the stability and timeliness of automatic replenishment of mobile phone accessories under diverse and heterogeneous sales data, and meets the core requirements of supply chain inventory management for dynamic inventory allocation and supply and demand matching reliability. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of a method for automatic replenishment of mobile phone accessories based on dynamic inventory provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the fluctuation amplitude numerical smoothing correction provided in the first embodiment of the present invention; Figure 3 This is a schematic diagram of the target consumption rhythm threshold comparison provided in the first embodiment of the present invention; Figure 4 This is a schematic diagram of the target consumption rhythm threshold comparison provided in the first embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Reference Figure 1 The first embodiment of the present invention provides a method for automatic replenishment of mobile phone accessories based on dynamic inventory, including the following steps: S11: Obtain the entry and exit record sequence of mobile phone accessories and perform time-series memory retrieval processing to obtain the initial demand trend; S12, perform clustering calculations based on the initial demand trend, pre-extracted seasonal rotation features and life cycle features to obtain the initial consumption rhythm; S13, calculate the fluctuation amplitude value of the initial consumption rhythm; if the fluctuation amplitude value does not meet the preset stability judgment condition, perform smoothing correction processing to obtain the target consumption rhythm; if the stability judgment condition is met, use the initial consumption rhythm as the target consumption rhythm. S14, if the target consumption rate is greater than the preset out-of-stock determination threshold, a first replenishment cycle for the out-of-stock state is generated; if the target consumption rate is less than the preset backlog determination threshold, a second replenishment cycle for the backlog state is generated; if the target consumption rate is within the normal range, the current state is maintained. S15, perform integrated decision evaluation on the first replenishment cycle and the second replenishment cycle to obtain the target replenishment decision; extract the target cumulative value according to the target replenishment decision and verify the actual consumption rhythm. If the verification is successful, generate a replenishment trigger instruction to execute inventory allocation.

[0015] In step S11, obtaining the entry and exit record sequence of mobile phone accessories includes: Based on the pre-acquired inventory records and transaction vouchers, perform data matching processing to generate initial transaction records; If the business type of the initial transaction record is inbound or outbound, then the aggregation and compression process is performed according to the transaction time to obtain the daily change record; The daily change records are subjected to abnormal fluctuation filtering to obtain smoothed change quantities; Based on the smoothed change quantity, a time span summary data is constructed, and the inflow and outflow record sequences are extracted from the summary data.

[0016] It should be noted that, firstly, the inventory records originate from the warehouse management database, and the transaction vouchers originate from the purchasing and sales voucher database. Inventory records are synchronized once daily after receiving, issuing, and inventory counting, while transaction vouchers are synchronized once after each purchase receipt or sales issue. The inputs for this step are inventory records and transaction vouchers. Inventory records include fields such as material code, material name, specifications, warehouse code, and physical inventory balance. Transaction vouchers include fields such as voucher number, material code, transaction time, business type, and quantity change. During data matching, the material code is used as the primary key. Inner joins are performed between the inventory records and transaction vouchers, retaining records with the same material code and a business type of receiving or issuing. Duplicates are then removed by voucher number to generate initial transaction records. Initial transaction records contain four fields: material code, transaction time, business type, and quantity change. The quantity change is in units of pieces, and the transaction time is accurate to the day.

[0017] Next, aggregation and compression are performed on records with the business type of "inbound" or "outbound" in the initial transaction log. Since multiple inbound and outbound transactions of the same model of part may occur within the same calendar day, directly using the transaction log can easily lead to fragmented data interfering with subsequent consumption trends. During aggregation, material codes and transaction dates are used as grouping criteria. Outbound changes are recorded as positive numbers, and inbound changes as negative numbers. The algebraic sum of changes within the same calendar day is calculated to obtain the net change quantity for that day. Dates with a net change quantity of 0 are still retained to ensure the continuity of the time series. After aggregation, the daily change record is output, containing two fields: transaction date and net change quantity for that day, with the quantity unit being pieces.

[0018] Subsequently, abnormal fluctuation filtering is performed on the daily change records. Sudden large-scale transfers or inventory adjustments often occur during daily inventory turnover; these irregular fluctuations can interfere with the true consumption trend. The filtering process uses a moving average algorithm with a window length of 7 days. For each calendar day, the arithmetic mean of the daily net change quantity of the three days before and after it (7 days total) is calculated. For boundary dates less than 7 days, the average is calculated based on the actual number of available days. The anomaly judgment threshold is set based on the daily net change quantity of the same type of mobile phone accessories within the same warehouse, up to the judgment date and preceding 30 consecutive calendar days. The difference between the 5th percentile and the 95th percentile of this data set is calculated as the normal fluctuation bandwidth. When the daily net change quantity deviates from its moving average by more than twice the normal fluctuation bandwidth, that day is judged as an anomaly, and the original value is replaced with the corresponding moving average for that day. The baseline value for the deviation multiple is 2.0. To reduce misjudgment, it can be adjusted between 1.5 and 2.5. Below 1.5, normal peaks are easily misjudged as anomalies; above 2.5, minor anomalies are easily missed.

[0019] Then, a time-span summary data is constructed based on the smoothed change quantities. The smoothed change quantities, after filtering out abnormal fluctuations, are arranged in chronological order by transaction date to form a continuous daily sequence. The most recent 90 days are extracted from this sequence as the inflow / outflow record sequence. The 90-day period is chosen because mobile phone accessory sales typically follow a quarterly seasonal cycle; this duration covers a complete seasonal change while avoiding interference from excessively long lifecycle phases. The final inflow / outflow record sequence consists of smoothed net outflow quantities for 90 consecutive natural days, with each time step being a scalar unit in pieces, used for subsequent time-series memory retrieval processing.

[0020] It is worth noting that the field names and quantity units in the above steps are consistent, the time unit is uniformly set to day, and the quantity unit is uniformly set to piece, without the need for additional normalization processing.

[0021] In one implementation, the execution of the time-series memory retrieval process to obtain the initial demand trend includes: Perform sliding truncation processing on the ingress and egress record sequence to construct multiple consecutive sequence segments; The sequence fragment is input into a preset temporal extraction network to perform implicit gating update calculation, thereby obtaining the hidden layer temporal vector; A nonlinear dimension mapping calculation is performed on the hidden layer time series vector to obtain the initial demand trend.

[0022] It should be noted that a sliding truncation process is first performed on the inbound / outbound record sequence. The inbound / outbound record sequence consists of smoothed net outbound volumes for 90 consecutive natural days, with each time step being a scalar unit in pieces. The sliding window length is set to 7 days, and the step size is set to 1 day. The 7-day window length can cover the consumption cycle fluctuations of mobile phone accessories within a week, avoiding interference from single-day noise. Starting from the beginning of the inbound / outbound record sequence, a continuous 7-day segment is extracted every one day forward, resulting in 84 sequence segments for the 90-day sequence, each with a dimension of 7×1.

[0023] Before inputting each sequence segment into the preset temporal extraction network, min-max normalization is performed on the seven values ​​within the segment. The lower bound of the normalization is the minimum of the daily net outflow of all mobile phone accessory models over the past 12 months, and the upper bound is the maximum value among them. Each value is linearly mapped to the interval between 0 and 1. During inference, values ​​exceeding the upper and lower bounds are pruned according to those bounds. When the lower and upper bounds are equal, the normalization result is set to 0.5. This process unifies the numerical scale while preserving the magnitude difference in net outflow between different accessory models, ensuring consistency between the dimensionality of the network input and the training labels.

[0024] The temporal sequence extraction network employs a Long Short-Term Memory (LSTM) network, which consists of an input layer, two hidden layers, and an output mapping layer. The input layer receives seven time steps, each containing a normalized value. Each of the two hidden layers has 64 hidden units, with a random deactivation rate of 0.2 between layers to suppress overfitting during training. The sigmoid function is used for activation within the forget gate, input gate, and output gate, while the tanh function is used for updating cell and hidden states. Each sequence segment is input into the network sequentially at each time step, and the network outputs a hidden state at each time step. The hidden state from the last time step is taken as the hidden layer temporal vector for that segment, with a dimension of 1×64.

[0025] It is worth noting that the training data was collected from historical inbound and outbound records in the warehouse management database, extracting daily net outbound volume sequences for at least 30 different mobile phone accessory models over the past 12 months, spanning 90 consecutive days. For each 90-day sequence, the same abnormal fluctuation filtering process as described above was performed first, followed by the same sliding truncation and intra-segment min-max normalization processing. Only segments where the following 3 consecutive days still fall within the 90-day sequence were retained as training samples. The 7-day data for each segment were used as input features, and the actual net outbound volume for the following 3 consecutive days was used as the training label. All training samples were divided into training and validation sets in a 7:3 ratio. The loss function used was mean squared error, the optimizer was Adam, the initial learning rate was set to 0.001, and the batch size was set to 32. The maximum number of training epochs was 50, and the learning rate decayed to 0.8 times the previous value after every 10 epochs. Training was stopped early when the validation set loss did not decrease for 5 consecutive epochs. After training, all weight and bias parameters within the network were fixed.

[0026] Then, the hidden layer time series vector is input into a fully connected layer to perform nonlinear dimension mapping calculation. The fully connected layer contains a hidden mapping layer and an output layer connected sequentially. The hidden mapping layer has an input dimension of 64 and an output dimension of 64, and uses the tanh function as the activation function. The output layer has an input dimension of 64 and an output dimension of 3, and no activation function is set. The nonlinear dimension mapping is implemented by the tanh function of the hidden mapping layer. The fully connected layer maps the 64-dimensional hidden layer time series vector into 3 output values, with the output values ​​covering negative values, corresponding to the predicted values ​​of the daily net outflow for the next 3 days, in units / day. After performing forward calculation on each of the 84 sequence segments, the 3-dimensional output corresponding to the last segment is taken as the initial demand trend at the current moment, with a dimension of 1×3.

[0027] In step S12, the step of performing clustering calculations based on the initial demand trend, pre-extracted seasonal rotation features, and life cycle features to obtain the initial consumption rhythm includes: The initial demand trend, the seasonal rotation feature, and the life cycle feature are subjected to vector concatenation processing to construct a multi-dimensional fusion feature matrix; Obtain preset initial cluster centers, calculate the feature distance between the multidimensional fusion feature matrix and the initial cluster centers, and iteratively update the cluster centers to obtain convergent classification clusters; The initial consumption rhythm is obtained by mapping the business attributes corresponding to the convergent classification clusters.

[0028] It should be noted that, firstly, vector concatenation processing is performed on the initial demand trend, seasonal rotation characteristics, and life cycle characteristics. The initial demand trend comes from the time-series memory extraction results of the previous step, with a dimension of 1×3. The three values ​​correspond to the predicted values ​​of the daily net outflow for the next three days, in units of pieces / day. The seasonal rotation characteristics are extracted using seasonal intensity mapping based on historical sales data. The average daily net outflow of mobile phone accessories of the same category in the same warehouse for each natural month over the past 12 months is extracted. The ratio of the historical average monthly net outflow corresponding to the current month's sequence number m to the average monthly net outflow of all months over the past 12 months is calculated to obtain the seasonal intensity ratio. This seasonal intensity ratio is then subtracted by 1, multiplied by 2, and cropped to the range of -1 to 1 to obtain the seasonal rotation characteristic value. For example, for protective case accessories of a certain brand's flagship mobile phone, its seasonal rotation characteristics usually peak during the annual new product launch period, when the demand for accessories driven by the new phone launch is concentrated. In the traditional off-season, the seasonal intensity ratio of this accessory is at a low level, reflecting the typical non-monotonic seasonality of the mobile phone accessory market. For such accessories, seasonal rotation features can effectively capture the concentrated sales periods brought about by new product launches and promotional cycles, avoiding the use of linear monthly mapping as the sole basis for seasonality strength. The lifecycle feature is calculated based on the cumulative number of days from the accessory's first entry date to the current date. This cumulative number of days is divided by the historical average lifecycle days of similar accessories. This average lifecycle days is the arithmetic mean of the number of days from the first entry date to the last exit date for similar accessories in the historical dataset. When there are no lifecycle records for similar accessories in the historical dataset or the arithmetic mean is less than or equal to 0, the lifecycle feature takes a default value of 0.5 and is cropped to the 0-1 range, where 0 represents newly launched and 1 represents about to be discontinued. The three values ​​from the initial demand trend, the seasonal rotation feature value, and the lifecycle feature value are concatenated sequentially into a 5-dimensional vector. Then, min-max normalization is performed on this 5-dimensional vector, using the minimum and maximum values ​​obtained from each dimension in the historical mobile phone accessory dataset as the upper and lower bounds for normalization. The values ​​of each dimension of this 5-dimensional vector are linearly mapped to the 0-1 range, resulting in a multi-dimensional fusion feature matrix with dimensions of 1×5.

[0029] Next, the preset initial cluster centers are obtained. These initial cluster centers are extracted from the historical mobile phone accessories dataset. The dataset contains multi-dimensional fused feature vectors of all mobile phone accessories of the same model from the same warehouse over the past 12 months, under the same seasonal and lifecycle conditions, totaling approximately 2000 records. The number of clusters is set to three categories, corresponding to three business attributes: high consumption rhythm, stable consumption rhythm, and low consumption rhythm. Three records are randomly selected from the dataset as initial cluster centers. Feature distance is calculated using Euclidean distance. Iterative updates are performed on all approximately 2000 multi-dimensional fused feature vectors in the historical mobile phone accessories dataset: In each iteration, each record in the dataset is assigned to the nearest category cluster, and the arithmetic mean of all records in each dimension within each category cluster is recalculated as the updated cluster center. This process of distance calculation, classification, and center update is repeated until the movement distance of the cluster center between two consecutive iterations is less than 0.01 or the maximum number of iterations (10) is reached, resulting in a converged category cluster. After the multidimensional fusion feature matrix of the current accessories converges to the cluster centers, it is assigned to the corresponding convergent classification cluster in one go according to the nearest distance principle. This clustering process is unsupervised learning, requires no labeled data, and does not involve loss functions or optimizers.

[0030] Secondly, the initial consumption rhythm is obtained by mapping the business attributes corresponding to the converged category clusters. After convergence, the 30-day average daily net outflow corresponding to the historical consumption rhythm template is calculated for each of the three category clusters. Based on this average daily net outflow, the three category clusters are sequentially classified into three business attributes: high consumption rhythm, stable consumption rhythm, and low consumption rhythm. Each category cluster corresponds to a historical consumption rhythm template, which is a sequence of average daily net outflow over 30 consecutive days, with a dimension of 30×1 and a unit of pieces / day. This template is obtained by calculating the arithmetic mean of the net outflow of all historical parts within the category cluster over their respective 30-day periods, and then arranging them in daily order. After the multidimensional fusion feature matrix of the current part is assigned to the converged category cluster, the historical consumption rhythm template corresponding to that category cluster is directly extracted as the initial consumption rhythm. If the Euclidean distance between the current accessory and the center of the convergent cluster is greater than 0.5, the two nearest cluster templates are extracted. The reciprocals of the distances from the current accessory to the two cluster centers are used as weights, and the two weights are normalized proportionally so that the sum of the weights is 1. The two templates are then weighted and averaged daily to obtain the initial consumption rhythm. The distance threshold of 0.5 is set based on historical dataset statistics. Under the same seasonal and life cycle conditions, the maximum Euclidean distance between the fused feature vectors of accessories of the same type usually does not exceed 0.5. If the distance exceeds this, it indicates that the accessory has a weak match with the cluster and is suitable for interpolation processing using neighboring templates.

[0031] Subsequently, the initial consumption rhythm is output. The initial consumption rhythm is a 30×1 time series, where each element represents the estimated net outflow on a given day between day 1 and day 30, in units per day. This series is used for subsequent fluctuation calculations and stability assessments.

[0032] In step S13, the fluctuation amplitude of the initial consumption rhythm is calculated; if the fluctuation amplitude does not meet the preset stability determination condition, a smoothing correction process is performed to obtain the target consumption rhythm, including: Extract the local maxima and local minima of the initial consumption rhythm within the time segment, calculate the difference between the local maxima and the local minima, and obtain the fluctuation amplitude value; Obtain the historical mean squared error parameters and generate the stationarity determination criteria; If the fluctuation amplitude value does not meet the stability determination condition, a moving average calculation is performed on the fluctuation amplitude value to obtain a corrected amplitude value, and the target consumption rhythm is reconstructed based on the corrected amplitude value.

[0033] It should be noted that, firstly, the local maxima and minima of the initial consumption rhythm within the time segments are extracted. The initial consumption rhythm is a 30×1 time series, where each element represents the estimated net outflow for a future day, in units / day. The 30-day series is truncated using a 5-day window with a step size of 1 day, resulting in 26 consecutive time segments. Within each time segment, the magnitudes of 5 values ​​are compared. Points with values ​​greater than their two adjacent values ​​are designated as local maxima, and points with values ​​less than their two adjacent values ​​are designated as local minima. The first and last points within the window are each compared only with their innermost adjacent point. For each time segment, the absolute value of the difference between the largest local maximum and the smallest local minimum is taken as the fluctuation amplitude value for that segment. When there are no local maxima or local minima within the window, the absolute value of the difference between the largest and smallest of the 5 values ​​within the window is taken as the fluctuation amplitude value for that segment. This results in 26 fluctuation amplitude values, forming a fluctuation amplitude sequence of length 26, in units / day.

[0034] For example, when a new generation of a certain mobile phone model is released, the demand for accessories such as screen protectors and phone cases for the older model often drops sharply in the short term. This is reflected in the initial consumption rhythm as a significant decrease in net inventory output for some days, with increased fluctuations. By extracting local extreme differences and performing stability assessments, the system can identify abnormal consumption rhythms caused by model rotation without manual intervention, thereby triggering smoothing corrections. This avoids misjudging short-term market withdrawal shocks as normal demand fluctuations, thus ensuring the stability of replenishment decisions.

[0035] Next, the historical mean squared error (MSE) parameter is obtained, and a stationarity criterion is generated. The historical MSE parameter is derived from the initial consumption rhythm of mobile phone accessories of the same category in the same warehouse over the past 12 months. A 30-day initial consumption rhythm sequence of these accessories under the same seasonal and lifecycle conditions is collected. The standard deviation of the net outflow of different accessories is calculated for each day, and the 90th quantile of the 30-day standard deviation is taken as the historical MSE parameter. The 90th quantile is chosen because it covers 90% of the normal fluctuation levels in historical data, excluding interference from a few extreme fluctuation samples, while retaining the ability to identify anomalies. The stationarity criterion is set as follows: if the maximum value in the fluctuation amplitude sequence does not exceed 1.2 times the historical MSE parameter, the current initial consumption rhythm is considered to be in a stationary state. 1.2 times is a safety factor introduced based on historical statistics and can be adjusted between 0.8 and 1.5 times. The adjustment direction depends on the business's requirements for inventory stability; the higher the stability requirement, the closer the safety factor is to 0.8 times, and vice versa. This criterion requires no training and is directly determined by the historical data distribution.

[0036] Secondly, if the fluctuation amplitude values ​​do not meet the stability criteria, a moving average operation is performed on the fluctuation amplitude sequence. The moving average window length is set to 5 fluctuation amplitude values, with a step size of 1. For each position, the arithmetic mean of the two fluctuation amplitude values ​​before and after it, plus the value itself, is taken as the correction amplitude value for that position. The boundary positions within the window are averaged based on the actual available quantity. After moving average, 26 correction amplitude values ​​are obtained, forming a correction amplitude sequence. When reconstructing the target consumption rhythm, the 26 correction amplitude values ​​are used as window fluctuation amplitude constraints to perform window-by-window fluctuation amplitude correction on the initial 30-day consumption rhythm. For the j-th 5-day window, j takes values ​​from 1 to 26, and this window corresponds to days j to j+4 in the initial consumption rhythm; the 5 net outflow values ​​within this window are extracted, and the window mean is calculated. Subtract each value Obtain the residual sequence within the window ; Scale the residual sequence by a factor equal to the correction magnitude. Divide by the difference between the original maximum and minimum values ​​within the window; if the original maximum difference is 0, then the multiplier is 1; then the window mean is... Adding the scaled residuals, we obtain the reconstructed window sequence corresponding to that window. After performing the above operation on all 26 windows, for each day from day 3 to day 28, we calculate the arithmetic mean of the reconstructed values ​​obtained by multiple windows covering that day, and use this as the target net outflow for that day. For days 1, 2, 29, and 30, as the preceding and following boundary days, we directly retain the values ​​of the corresponding days in the initial consumption rhythm as the target values. Thus, we obtain the target consumption rhythm D'=[d1',d2',...,d30']. The target consumption rhythm is still a 30×1 time series, with units per day.

[0037] If the stability criteria are met, the initial consumption rhythm is directly used as the target consumption rhythm, and the moving average and reconstruction operations are not performed. Then, the target consumption rhythm is output. The target consumption rhythm includes the estimated daily net outflow for the next 30 days, in units per day.

[0038] For example, Figure 2 This is a schematic diagram illustrating the smoothing correction of fluctuation amplitude values. The diagram shows the fluctuation amplitude values ​​and correction amplitude values ​​for 26 consecutive time segments. The correction amplitude values ​​reduce abrupt changes between adjacent time segments while maintaining the overall trend, providing continuous amplitude constraints for the rhythm of reconstructing the target consumption.

[0039] In step S14, the step of generating a first replenishment cycle for a stockout state if the target consumption rate is greater than a preset stockout determination threshold includes: If the target consumption rate is greater than the stockout determination threshold, then extract the corresponding physical inventory balance and the quantity of replenishment in transit. The actual available inventory is obtained by adding the physical inventory balance to the quantity of replenishment in transit. Calculate the ratio of the actual available inventory to the target consumption rate to obtain the estimated depletion time; If the estimated depletion time is less than the pre-obtained comprehensive replenishment time, a first replenishment cycle is generated based on the difference between the comprehensive replenishment time and the estimated depletion time; if the estimated depletion time is not less than the comprehensive replenishment time, a normal replenishment cycle is generated.

[0040] It's important to note that, firstly, the target consumption rhythm is compared with the stockout threshold. The target consumption rhythm is derived from the stability processing in the previous step, representing a 30×1 time series, where each element represents the estimated net outbound volume for the corresponding future date, in units per day. The stockout threshold is determined by increasing the historical average daily consumption of this component by 40%. The historical average daily consumption is calculated as the arithmetic mean of the daily net outbound volumes for the 15 days before and after the same date within the past 12 months. The 40% increase is based on the fact that the daily consumption corresponding to the 95th percentile in the historical outbound volume distribution is typically about 40% higher than the average; exceeding this level indicates a significantly faster inventory depletion rate. This increase can be adjusted between 0.8 and 1.2 times, meaning the actual increase range is between 32% and 48%. A 0.8-fold increase is used when the warehouse prioritizes avoiding stockouts, and a 1.2-fold increase is used when the warehouse prioritizes reducing inventory backlog. When comparing, the arithmetic mean of the net outflow volume over the next 7 days within the target consumption rhythm is used as the comparison value. This comparison value is a scalar, with units per day. If the comparison value is greater than the larger of the stockout judgment threshold and the backlog judgment threshold, this branch is processed; otherwise, this branch is not executed.

[0041] Next, the physical inventory balance and the quantity of replenishment in transit are retrieved. The physical inventory balance is read from the inventory ledger in the warehouse management database by material code, recording the available quantity that has been received but not yet dispatched at the current moment, in units of pieces. The quantity of replenishment in transit is read from the purchase order data by material code, extracting the quantities corresponding to effective purchase orders that have not yet been received and summing them, in units of pieces. Both reads are based on the timestamp of the current batch, ensuring that both sets of data reflect the inventory status at the same moment. Both the inventory ledger and purchase order data are indexed by material code to avoid material code mismatches.

[0042] Secondly, the physical inventory balance is added to the quantity of replenishment in transit to obtain the actual available inventory. Actual available inventory is a scalar quantity, measured in units. It represents the total inventory currently at hand, including not only the actual stock in the warehouse but also the quantity about to be received in transit, thus more accurately reflecting the available scale of the component under the risk of stockout.

[0043] Subsequently, the ratio of actual available inventory to the target consumption rhythm is calculated to obtain the estimated depletion duration. Since the target consumption rhythm is a 30×1 sequence, the arithmetic mean of the net outflows for the next 7 days extracted in the previous step is used as the divisor, in units of pieces / day. When the average daily net outflow is less than or equal to 0, it is determined that there is no risk of inventory depletion, and no division operation is performed; a normal replenishment cycle is directly generated. When the average daily net outflow is greater than 0, the actual available inventory is divided by the average daily net outflow, and the resulting estimated depletion duration is a scalar, in days. This value represents the number of days that the actual available inventory can sustain at the current consumption level. Dimensionally, pieces divided by pieces / day yields days, ensuring dimensional matching.

[0044] Then, the overall replenishment time is obtained and compared with the estimated depletion time. The overall replenishment time consists of three parts: supplier production cycle, logistics and transportation time, and safety buffer time. The supplier production cycle is the average number of days from order placement to shipment in the three most recent purchase orders; the logistics and transportation time is the average number of days from shipment to warehousing in the three most recent purchase orders; and the safety buffer time is fixed at 1 day. The three values ​​are added together to obtain the overall replenishment time, in days. If the estimated depletion time is less than the overall replenishment time, it means that the existing inventory cannot support the arrival of a new batch of replenishment. In this case, the overall replenishment time is subtracted from the estimated depletion time, and the difference is used as the first replenishment cycle, in days. The first replenishment cycle is used to advance the replenishment action to make up for the time gap between inventory depletion and replenishment arrival. If the estimated depletion time is not less than the overall replenishment time, a normal replenishment cycle is generated. The normal replenishment cycle follows the standard replenishment lead time in the purchase parameter table. In this embodiment, the standard replenishment lead time is set to 7 days.

[0045] In one implementation, the step of generating a second replenishment cycle for a backlogged state if the target consumption rate is less than a preset backlog determination threshold includes: If the target consumption rate is less than the backlog judgment threshold, the difference between the total inventory balance and the safety stock floor is calculated to obtain the amount of stagnant inventory. The inventory holding load is obtained by multiplying the stagnant inventory quantity with the pre-acquired storage occupancy conversion factor. Based on the aforementioned inventory holding load and the preset return discount coefficient, a multiple regression analysis is performed to obtain the inventory lag difficulty index; The inventory consumption rate is obtained by performing a discount response simulation based on the inventory lag difficulty index, and a second replenishment cycle is generated based on the inventory consumption rate and the pre-obtained remaining shelf life.

[0046] It should be noted that, firstly, the target consumption rhythm is compared with the backlog judgment threshold. The target consumption rhythm is a 30×1 time series, where each element represents the estimated net outbound volume for the corresponding future date, in units of pieces / day. The backlog judgment threshold is determined based on a 30% reduction from the historical average daily consumption of the component during the same period. The historical average daily consumption during the same period is the arithmetic mean of the daily net outbound volumes for the 15 days before and after the same date within the past 12 months. The 30% reduction is based on the fact that the daily consumption corresponding to the 5th quantile in the historical shipment distribution is usually about 30% lower than the average. A level below this indicates a significant slowdown in inventory consumption and the formation of backlog. This reduction ratio can be adjusted between 0.8 and 1.2 times, meaning the actual reduction range is between 24% and 36%. A ratio of 0.8 is used when the warehouse prioritizes reducing backlog, and 1.2 times when prioritizing ensuring inventory supply. The comparison is based on the arithmetic mean of the net outbound volumes for the next 7 days within the target consumption rhythm. This comparison value is a scalar quantity, in units of pieces / day. If the comparison value is less than the smaller of the out-of-stock threshold and the backlog threshold, then proceed to this branch for processing.

[0047] For example, for original battery accessories of a mobile phone model that is about to be discontinued, its lifecycle characteristic is close to 1, and the target consumption rate is often consistently lower than the backlog judgment threshold. At this time, the system will calculate the amount of stagnant inventory and the inventory consumption rate, and combine this with the remaining shelf life to generate a second replenishment cycle, prompting the warehouse to postpone the next purchase and prioritize clearance sales to avoid long-term stagnant inventory due to model discontinuation. This scenario, directly related to the end of the lifecycle of mobile phone accessories and model rotation characteristics, demonstrates the practical applicability of this method in mobile phone accessory inventory management.

[0048] Next, the difference between the total remaining inventory and the safety stock floor is calculated to obtain the stagnant inventory. The total remaining inventory is read from the inventory ledger in the warehouse management database by material code, including all quantities currently in stock but not yet out of stock, in units of pieces. The safety stock floor is read from the material file by material code, set according to the supplier's delivery cycle and daily demand fluctuations for that component, in units of pieces. When the total remaining inventory is less than or equal to the safety stock floor, the stagnant inventory is set to 0. In this case, the calculation of the subsequent inventory holding load to the second replenishment cycle is not performed, and the second replenishment cycle directly uses the standard replenishment lead time of 7 days. When the total remaining inventory is greater than the safety stock floor, the difference between the two is the stagnant inventory, a scalar quantity in units of pieces, representing the quantity of inventory exceeding the safety stock floor and unable to be consumed through normal turnover. The inventory holding load is obtained by multiplying the stagnant inventory by a pre-obtained warehouse occupancy conversion factor. The warehouse occupancy conversion factor is converted from the storage fee and management fee in the warehousing contract into a daily occupancy cost per piece, in yuan / (piece·day). Multiplying the amount of stagnant inventory by the storage occupancy conversion factor yields the daily inventory holding load, expressed in yuan / day, reflecting the daily occupancy scale of backlogged inventory at the current holding level.

[0049] Subsequently, a multiple regression analysis was performed based on the inventory holding load and the preset return discount coefficient to obtain the inventory stagnation difficulty index. The return discount coefficient was determined based on the discount ratio statistics of the past 12 months of return disposal records for the same type of parts, taking the average of historical discount ratios, with a value ranging from 0.1 to 0.5, which can be adjusted within this range according to the supplier negotiation results. The multiple regression model adopted a linear regression structure, with the input features being the inventory holding load and the return discount coefficient, and the output being the inventory stagnation difficulty index. The training set was collected from the backlog disposal records completed in the same warehouse over the past 12 months. For each record, the inventory holding load, return discount coefficient, and corresponding inventory stagnation difficulty index were extracted. The inventory stagnation difficulty index was calculated by subtracting the disposal recovery discount ratio from 1. The disposal recovery discount ratio is the ratio of the final disposal recovery discount value of the batch of backlog parts to the initial inventory discount value, with a value ranging from 0 to 1. The larger the value, the lower the disposal recovery ratio and the higher the disposal difficulty. The training sample consisted of approximately 800 records, divided into a training set and a validation set in a 7:3 ratio. The loss function is mean squared error, and the optimizer is the adaptive moment estimator. The learning rate is set to 0.001, the batch size is 32, and the maximum number of training epochs is 50. Every 10 epochs, the learning rate decays to 0.8 times the previous value. Training stops when the validation set loss does not decrease for 5 consecutive epochs. Before training, the input features are subjected to min-max normalization, mapping inventory holding load and return discount coefficients to the interval between 0 and 1, respectively. The upper and lower bounds of the normalization are taken as the maximum and minimum values ​​in the training set, respectively. After training, the weights and bias parameters are fixed.

[0050] Secondly, the inventory consumption rate is calculated by simulating the discount response based on the inventory stagnation difficulty index. Multiplying the inventory stagnation difficulty index by the upper limit of the discount intensity coefficient (0.3) yields the suggested discount intensity coefficient. 0.3 is a commonly used upper limit for industry clearance discounts and can be adjusted between 0.2 and 0.4. The discount response coefficient is extracted from historical sales records. This coefficient is defined as the percentage increase in shipment volume when the promotion intensity increases by 10%. For example, a discount response coefficient of 1.5 means that a 10% increase in promotion intensity results in a 15% increase in shipment volume. Multiplying the discount intensity coefficient by the discount response coefficient gives the shipment volume increase percentage. Adding 1 to this percentage and multiplying it by the arithmetic average of the net outflow volume over the next 7 days gives the inventory consumption rate, in units per day.

[0051] Then, a second replenishment cycle is generated based on the inventory consumption rate and the pre-obtained remaining shelf life. The remaining shelf life is read from the material file by material code, in days, representing the time that the stockpiled parts can still be sold normally in the warehouse. When the inventory consumption rate is less than or equal to 0, no division operation is performed; it is directly determined that the digestion time is greater than the remaining shelf life, and the remaining shelf life is used as the second replenishment cycle, simultaneously triggering a return or damage reporting process. When the inventory consumption rate is greater than 0, the stagnant inventory is divided by the inventory consumption rate to obtain the digestion time, in days. The digestion time is compared with the remaining shelf life. If the digestion time is less than the remaining shelf life, the digestion time is used as the second replenishment cycle; if the digestion time is not less than the remaining shelf life, the remaining shelf life is used as the second replenishment cycle, simultaneously triggering a return or damage reporting process. The second replenishment cycle represents the number of days from the current moment until the next purchase replenishment action is performed while the stockpiled inventory is in a state of stockpiling, in days.

[0052] In one implementation, if the target consumption rhythm is within the normal range, the current state is maintained. The normal range is defined by the smaller and larger values ​​of the backlog judgment threshold and the stockout judgment threshold. The arithmetic average of the net outflow over the next 7 days is used as a comparison value. When this comparison value is not less than the smaller value and not greater than the larger value, it is considered normal. At this time, no first or second replenishment cycle is generated, the standard replenishment lead time of 7 days is used, the inventory status table remains marked as normal, and the material files, purchase plan, and safety stock baseline are not modified. In subsequent integrated decision evaluation, the replenishment cycle input is filled with the standard replenishment lead time of 7 days, and the target consumption rhythm continues to participate in the calculation as a 30×1 time series.

[0053] For example, Figure 3 This is a diagram illustrating the target consumption rhythm threshold comparison. The diagram shows the relationship between the comparison value, formed by the arithmetic average of the net outflow over the next 7 days, and the stockout and backlog judgment thresholds. The comparison value consistently exceeds the stockout judgment threshold; the current comparison value is 56.74 units / day. Based on this, the system enters a stockout state and generates the first replenishment cycle.

[0054] In step S15, the integrated decision evaluation of the first replenishment cycle and the second replenishment cycle to obtain the target replenishment decision includes: Calculate the absolute difference between the first replenishment cycle and the second replenishment cycle to obtain the cycle difference magnitude; The decision feature vector is constructed by combining the periodic difference magnitude with the pre-acquired inventory turnover parameters. The decision feature vector is input into a preset integrated decision model to perform node splitting evaluation calculation, and the leaf node weights are obtained. A weighted voting statistical calculation is performed on the weights of the leaf nodes to obtain the voting classification results, and a target replenishment decision is generated based on the voting classification results.

[0055] It should be noted that, firstly, the absolute difference between the first and second replenishment cycles is calculated to obtain the cycle difference magnitude. The first and second replenishment cycles are derived from the state branch processing results of the previous step. If the current state is out of stock, the first replenishment cycle is the actually generated out-of-stock replenishment cycle, and the second replenishment cycle has not yet been generated, using the standard replenishment lead time of 7 days. If the current state is overstocked, the second replenishment cycle is the actually generated overstock replenishment cycle, and the first replenishment cycle uses the standard replenishment lead time of 7 days. If the current state is normal, both the first and second replenishment cycles use the standard replenishment lead time of 7 days. The absolute value of the difference between the two cycle values ​​is then used to obtain the cycle difference magnitude, which is a scalar quantity in days. The first and second replenishment cycles, along with the standard replenishment lead time, are used in this step as decision inputs reflecting the execution rhythm of the next procurement replenishment. The cycle difference magnitude is only used to measure the deviation of the replenishment rhythm between out-of-stock and overstocked states and is not directly used as the replenishment time length.

[0056] Next, the periodicity variation magnitude is combined with pre-acquired inventory turnover parameters to construct a decision feature vector. The inventory turnover parameter combination includes four items: inventory turnover rate, inventory availability days, the ratio of in-transit inventory to safety stock, and the arithmetic mean of net outflows over the next 7 days. The inventory turnover rate is read from the material files in the warehouse management database, representing the total outflow of the material over the past 30 days divided by the average inventory balance, measured in cycles. Inventory availability days are obtained by dividing the current physical inventory balance by the arithmetic mean of net outflows over the next 7 days, measured in days. When the arithmetic mean of net outflows over the next 7 days is less than or equal to 0, the inventory availability days are set to a preset value of 30 days. The ratio of in-transit inventory to safety stock is obtained by dividing the quantity of replenishment in transit by the safety stock floor, and is dimensionless. The arithmetic mean of net outflows over the next 7 days is the arithmetic mean of the first 7 values ​​extracted from the target consumption rhythm, measured in units per day. The periodicity variation magnitude, along with these four items, forms a 5-dimensional decision feature vector. The arithmetic mean of inventory turnover rate, inventory availability days, in-transit inventory to safety stock ratio, and net outflows in the next 7 days are subjected to min-max normalization, mapped to the interval between 0 and 1, with the upper and lower bounds being the maximum and minimum values ​​of the corresponding features in the historical data set, respectively. The magnitude of periodic differences retains its original dimensions and is not included in the normalization.

[0057] Secondly, the 5-dimensional decision feature vector is input into the pre-defined ensemble decision model to perform node splitting evaluation calculations, obtaining the leaf node weights. The ensemble decision model adopts a random forest classification model, consisting of 50 decision trees, with a maximum depth of 8 for each tree and a minimum number of samples per node for splitting, using the Gini coefficient as the splitting criterion. The 50 trees, maximum depth of 8, and minimum number of split samples of 10 are determined by grid search on historical datasets. The number of candidate trees is 20, 50, and 80, the maximum candidate depth is 6, 8, and 10, and the minimum number of split samples is 5, 10, and 20. The combination with the highest classification accuracy on the validation set is selected as the fixed hyperparameter. The training set is collected from mobile phone accessory replenishment events in the same warehouse over the past 12 months. Each sample includes five features: periodic difference magnitude, inventory turnover rate, inventory availability days, ratio of in-transit inventory to safety stock, and arithmetic mean of net outflows in the next 7 days. The label is the actual replenishment action ultimately performed in that replenishment event. The tags are derived from replenishment actions confirmed after manual review of historical inventory transfer orders, categorized into three types: emergency replenishment, deferred replenishment, and normal replenishment. Approximately 8,000 training samples are used, divided into training and validation sets in a 7:3 ratio. After training, each decision tree's leaf nodes store the frequency proportion of the three replenishment actions among the training samples falling on that node; these are the leaf node weights, with a dimension of 3×1. When splitting a node, the optimal split point is selected by calculating the decrease in the Gini coefficient of each candidate split point, until the maximum depth is reached or the number of samples in the node falls below the minimum number of samples required for splitting.

[0058] Subsequently, a weighted voting statistical calculation is performed on the leaf node weights to obtain the voting classification result. The 5-dimensional decision feature vector is input into 50 decision trees, and each tree completes a split path from the root node to a leaf node, obtaining a 3-dimensional weight vector for the corresponding leaf node. Each tree is assigned a voting weight, calculated based on its classification accuracy on the validation set; higher accuracy results in a larger weight, and the sum of the voting weights for all 50 trees is 1. The 3-dimensional weight vector output by each tree is multiplied by its corresponding voting weight and then summed class-wise to obtain a weighted voting score for the three categories. The three scores are compared, and the category corresponding to the highest score is taken as the voting classification result.

[0059] Then, a target replenishment decision is generated based on the voting classification results. The voting classification result is one of three categories: urgent replenishment, delayed replenishment, or normal replenishment; the target replenishment decision directly adopts this category. The target replenishment decision is used to subsequently extract target cumulative values ​​and verify the actual consumption rhythm.

[0060] In one implementation, the target cumulative value is extracted based on the target replenishment decision and the actual consumption rhythm is verified. If the verification is successful, a replenishment trigger instruction is generated to execute inventory allocation.

[0061] It should be noted that the target cumulative value is first extracted based on the target replenishment decision. The target replenishment decision falls into one of three categories: emergency replenishment, deferred replenishment, or normal replenishment. The extraction period for the target cumulative value is determined by the target replenishment decision: emergency replenishment corresponds to the next 3 days, deferred replenishment to the next 7 days, and normal replenishment to the next 7 days. After the extraction period is determined, the net outbound volume for the corresponding number of days is extracted from the target consumption rhythm and accumulated daily to obtain the target cumulative value, in units of pieces. This target cumulative value represents the predicted total outbound volume for the current replenishment decision.

[0062] Next, verify the actual consumption rhythm. The actual consumption rhythm is calculated from the actual outbound records of the most recent 7 days, taking the arithmetic mean of the daily net outbound volume of the most recent 7 calendar days, in units / day. Multiply the actual consumption rhythm by the number of extraction days corresponding to the target replenishment decision to obtain the actual cumulative value, in units. Subtract the target cumulative value from the actual cumulative value, take the absolute value, and then divide it by the actual cumulative value to obtain the consumption deviation ratio. When the actual cumulative value is less than or equal to 0, the above division operation is not performed, and the verification is directly judged as failing, and processed according to the verification failure procedure. The consumption deviation ratio is a dimensionless value. The verification pass condition is set at a consumption deviation ratio not exceeding 15%. The 15% setting is based on the deviation distribution of the target cumulative value and the actual cumulative value of effective replenishment events in historical inventory transfer data. Under normal replenishment decisions, this deviation usually does not exceed 15%. Exceeding this ratio indicates that the target consumption rhythm deviates too much from the recent actual consumption level, and the reliability of the replenishment decision is insufficient. This threshold can be adjusted within the range of 12% to 18%, with 12% used for warehouses with large inventory fluctuations and 18% used for warehouses with stable inventory.

[0063] If the consumption deviation ratio does not exceed 15%, the verification passes. Upon successful verification, a replenishment trigger instruction is generated. The replenishment trigger instruction includes five fields: material code, replenishment quantity, replenishment cycle, execution action, and generation timestamp. The material code is read from the currently processed inventory file. The replenishment quantity is calculated based on the target replenishment decision and the actual available inventory. The actual available inventory is the sum of the physical inventory balance and the replenishment quantity in transit. For emergency replenishment, the recommended replenishment quantity is the sum of the target cumulative value and the safety stock floor minus the actual available inventory; if the result is negative, it is set to 0. When replenishment is postponed, no positive replenishment quantity is generated, and the execution action is written as "postponed replenishment." For normal replenishment, if the actual available inventory is lower than the safety stock floor, the positive recommended replenishment quantity is the safety stock floor minus the actual available inventory, i.e., the inventory gap; if the actual available inventory is not lower than the safety stock floor, no replenishment quantity is generated. The replenishment cycle is read from the first or second replenishment cycle corresponding to the target replenishment decision. Emergency replenishment corresponds to the first replenishment cycle, while delayed replenishment corresponds to the second replenishment cycle. Normal replenishment uses a standard replenishment lead time of 7 days. The action to be executed is directly written to the target replenishment decision category. A timestamp is generated and written to the current date, accurate to the day.

[0064] Subsequently, the replenishment trigger instruction is written to the inventory status table in the warehouse management database. The inventory status table uses the material code as the primary key and records the status changes after this replenishment trigger. If verification fails, no replenishment trigger instruction is generated. Instead, the target replenishment decision, target cumulative value, actual cumulative value, and consumption deviation ratio are written to the replenishment verification anomaly record table, and the inventory status table is kept in a pending review state, awaiting manual processing or reprocessing in the next round of dynamic inventory calculation. This forms a complete closed loop, preventing erroneous replenishment instructions from being directly issued. While the inventory status table is in the pending review state, issuing new replenishment trigger instructions for that material code is suspended. The pending review state is lifted after manual review is completed and the review results are written, after which automatic processing for the next round of dynamic inventory calculation resumes.

[0065] For example, Figure 4 This is a diagram illustrating the consumption deviation ratio verification. The diagram shows the consumption deviation ratios for 48 replenishment events and the 15% verification pass condition. 45 of the replenishment events passed verification, while the 3 events that exceeded the verification pass condition entered a pending review state and did not generate replenishment trigger instructions.

[0066] In summary, this invention discloses an automatic replenishment method for mobile phone accessories based on dynamic inventory, which realizes dynamic adaptive control of the mobile phone accessory inventory allocation process and meets the core requirement of supply chain inventory management to maintain the reliability of supply and demand matching when facing complex demand changes.

[0067] The second embodiment of the present invention provides an automatic replenishment system for mobile phone accessories based on dynamic inventory, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the automatic replenishment method for mobile phone accessories based on dynamic inventory as described above.

[0068] It should be noted that the mobile phone accessory automatic replenishment system based on dynamic inventory provided in this embodiment of the invention is used to execute all the process steps of the mobile phone accessory automatic replenishment method based on dynamic inventory in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0069] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for automatic replenishment of mobile phone accessories based on dynamic inventory, characterized in that, include: Obtain the entry and exit record sequence of mobile phone accessories and perform time-series memory retrieval processing to obtain the initial demand trend; Based on the initial demand trend, pre-extracted seasonal rotation features, and life cycle features, clustering calculations are performed to obtain the initial consumption rhythm; Calculate the fluctuation range of the initial consumption rhythm; if the fluctuation range does not meet the preset stability determination condition, perform smoothing correction processing to obtain the target consumption rhythm; if the stability determination condition is met, use the initial consumption rhythm as the target consumption rhythm. If the target consumption rate is greater than the preset out-of-stock determination threshold, a first replenishment cycle in the out-of-stock state is generated; if the target consumption rate is less than the preset backlog determination threshold, a second replenishment cycle in the backlog state is generated; if the target consumption rate is within the normal range, the current state is maintained. An integrated decision evaluation is performed on the first replenishment cycle and the second replenishment cycle to obtain a target replenishment decision; the target cumulative value is extracted based on the target replenishment decision and the actual consumption rhythm is verified. If the verification is successful, a replenishment trigger instruction is generated to execute inventory allocation.

2. The automatic replenishment method for mobile phone accessories based on dynamic inventory according to claim 1, characterized in that, The acquisition of the entry and exit record sequence of mobile phone accessories includes: Based on the pre-acquired inventory records and transaction vouchers, perform data matching processing to generate initial transaction records; If the business type of the initial transaction record is inbound or outbound, then the aggregation and compression process is performed according to the transaction time to obtain the daily change record; The daily change records are subjected to abnormal fluctuation filtering to obtain smoothed change quantities; Based on the smoothed change quantity, a time span summary data is constructed, and the inflow and outflow record sequences are extracted from the summary data.

3. The automatic replenishment method for mobile phone accessories based on dynamic inventory according to claim 1, characterized in that, The execution of the time-series memory retrieval process to obtain the initial demand trend includes: Perform sliding truncation processing on the ingress and egress record sequence to construct multiple consecutive sequence segments; The sequence fragment is input into a preset temporal extraction network to perform implicit gating update calculation, thereby obtaining the hidden layer temporal vector; A nonlinear dimension mapping calculation is performed on the hidden layer time series vector to obtain the initial demand trend.

4. The automatic replenishment method for mobile phone accessories based on dynamic inventory according to claim 1, characterized in that, The step of performing clustering calculations based on the initial demand trend, pre-extracted seasonal rotation features, and life cycle features to obtain the initial consumption rhythm includes: The initial demand trend, the seasonal rotation feature, and the life cycle feature are subjected to vector concatenation processing to construct a multi-dimensional fusion feature matrix; Obtain preset initial cluster centers, calculate the feature distance between the multidimensional fusion feature matrix and the initial cluster centers, and iteratively update the cluster centers to obtain convergent classification clusters; The initial consumption rhythm is obtained by mapping the business attributes corresponding to the convergent classification clusters.

5. The automatic replenishment method for mobile phone accessories based on dynamic inventory according to claim 1, characterized in that, The calculation of the fluctuation amplitude of the initial consumption rhythm; If the fluctuation amplitude value does not meet the preset stability determination condition, a smoothing correction process is performed to obtain the target consumption rhythm, including: Extract the local maxima and local minima of the initial consumption rhythm within the time segment, calculate the difference between the local maxima and the local minima, and obtain the fluctuation amplitude value; Obtain the historical mean squared error parameters and generate the stationarity determination criteria; If the fluctuation amplitude value does not meet the stability determination condition, a moving average calculation is performed on the fluctuation amplitude value to obtain a corrected amplitude value, and the target consumption rhythm is reconstructed based on the corrected amplitude value.

6. The automatic replenishment method for mobile phone accessories based on dynamic inventory according to claim 1, characterized in that, If the target consumption rate is greater than a preset stockout determination threshold, a first replenishment cycle for a stockout state is generated, including: If the target consumption rate is greater than the stockout determination threshold, then extract the corresponding physical inventory balance and the quantity of replenishment in transit. The actual available inventory is obtained by adding the physical inventory balance to the quantity of replenishment in transit. Calculate the ratio of the actual available inventory to the target consumption rate to obtain the estimated depletion time; If the estimated depletion time is less than the pre-obtained comprehensive replenishment time, a first replenishment cycle is generated based on the difference between the comprehensive replenishment time and the estimated depletion time; if the estimated depletion time is not less than the comprehensive replenishment time, a normal replenishment cycle is generated.

7. The automatic replenishment method for mobile phone accessories based on dynamic inventory according to claim 1, characterized in that, If the target consumption rate is less than a preset backlog determination threshold, a second replenishment cycle for the backlog status is generated, including: If the target consumption rate is less than the backlog judgment threshold, the difference between the total inventory balance and the safety stock floor is calculated to obtain the amount of stagnant inventory. The inventory holding load is obtained by multiplying the stagnant inventory quantity with the pre-acquired storage occupancy conversion factor. Based on the aforementioned inventory holding load and the preset return discount coefficient, a multiple regression analysis is performed to obtain the inventory lag difficulty index; The inventory consumption rate is obtained by performing a discount response simulation based on the inventory lag difficulty index, and a second replenishment cycle is generated based on the inventory consumption rate and the pre-obtained remaining shelf life.

8. The automatic replenishment method for mobile phone accessories based on dynamic inventory according to claim 1, characterized in that, The step of performing integrated decision evaluation on the first replenishment cycle and the second replenishment cycle to obtain the target replenishment decision includes: Calculate the absolute difference between the first replenishment cycle and the second replenishment cycle to obtain the cycle difference magnitude; The decision feature vector is constructed by combining the periodic difference magnitude with the pre-acquired inventory turnover parameters. The decision feature vector is input into a preset integrated decision model to perform node splitting evaluation calculation, and the leaf node weights are obtained. A weighted voting statistical calculation is performed on the weights of the leaf nodes to obtain the voting classification results, and a target replenishment decision is generated based on the voting classification results.

9. An automatic replenishment system for mobile phone accessories based on dynamic inventory, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the automatic replenishment method for mobile phone accessories based on dynamic inventory as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the automatic replenishment method for mobile phone accessories based on dynamic inventory as described in any one of claims 1 to 8.

Images