AI bi business decision method based on big data analysis

Abstract

This invention relates to the field of artificial intelligence business decision-making and warehouse scheduling control technology, specifically an AI BI business decision-making method based on big data analysis. This method is applied to a warehouse scheduling system, with inventory goods as the target business object. The method includes: collecting front-end user click timestamps, warehouse system inventory deduction records, and baseline decision vectors generated by the backbone AI model; extracting features and constructing a transient disturbance feature matrix within a sliding time window; calculating the matrix norm and comparing it with a historical benchmark threshold; outputting the baseline decision vector when the threshold is not exceeded, and generating and superimposing incremental compensation instructions when the threshold is exceeded; updating the linear adjustment weight matrix with negative feedback suppression based on task cancellation rate and resource occupancy dwell time within the next sliding time window. This invention aims to decouple low-frequency stable decision-making from high-frequency transient compensation, quickly respond to sudden business fluctuations on the warehouse side, and reduce the computational overhead of frequent backbone model wake-ups.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence business decision-making and warehouse scheduling control technology, specifically to an AI BI business decision-making method based on big data analysis. Background Technology

[0002] In retail warehousing scheduling scenarios, the status of front-end user access, order placement intent, warehouse inventory deduction, picking and replenishment in transit, etc., will change continuously in a short period of time. The warehousing execution end also needs to simultaneously complete the adjustment of pricing parameters and the allocation of picking priorities. Therefore, the efficiency of generating business decision results and the stability of scheduling response will directly affect the turnover efficiency of inventory goods and the business capacity. In traditional approaches, business decisions mostly rely on the backbone model to perform uniform calculations at a low frequency, or to re-trigger full analysis after fluctuations occur. However, in the event of sudden increases in access, rapid changes in local inventory, or abnormal traffic surges, problems such as delayed decision response, high computing power consumption, and the coupling degree between local disturbances and global decisions exceeding the preset tolerance range are likely to occur, making the warehouse system still relatively unadaptable to short-term business fluctuations. Summary of the Invention

[0003] The purpose of this invention is to provide an AI BI business decision-making method based on big data analysis, and to solve the following technical problems: By decoupling low-frequency stable decision-making from high-frequency transient compensation, we can achieve a rapid response to sudden business fluctuations on the warehouse side and effectively control the computing power overhead caused by frequent calls to the backbone model.

[0004] The objective of this invention can be achieved through the following technical solutions: This AI-driven business decision-making approach, based on big data analytics, is applied to a warehouse scheduling system that includes both warehouse execution and dispatch vehicles. The target business object is inventory goods, including: Step 1: Collect the front-end user click timestamps of the inventory goods as event trigger time data and the inventory deduction records of the warehousing system as node resource status flow data, and extract the baseline decision vector generated by the preset backbone artificial intelligence model; Step 2: Within a preset sliding time window, construct a transient disturbance feature matrix based on the event trigger time data and the node resource status transition data; Step 3: Calculate the matrix norm of the transient disturbance feature matrix and compare it with a preset historical benchmark threshold. Step 4: When the matrix norm is not greater than the historical benchmark threshold, the baseline decision vector is output as a control command; when the matrix norm is greater than the historical benchmark threshold, the transient disturbance feature matrix and the linear adjustment weight matrix pre-constructed using historical feature samples are multiplied to obtain an incremental compensation command, which is then superimposed on the baseline decision vector to generate an updated control command, which is output to the warehouse execution end. The baseline decision vector and the incremental compensation command are both pre-divided into two control dimensions: resource allocation weight and picking priority. The control command includes instructions for modifying resource allocation weight parameters and increasing the picking priority of the dispatch cart. Step 5: Within the next sliding time window after the updated control command is output, monitor the task cancellation rate and resource occupation dwell time. When the task cancellation rate is greater than a preset cancellation rate threshold and the resource occupation dwell time is greater than a preset duration threshold, use the preset attenuation coefficient in the (0,1) interval to perform negative feedback suppression update on the linear adjustment weight matrix. When the task cancellation rate is not greater than the preset cancellation rate threshold or the resource occupation dwell time is not greater than the preset duration threshold, keep the linear adjustment weight matrix unchanged.

[0005] Furthermore, step one specifically includes: The event trigger time data is read through a preset application programming interface gateway; The node resource status flow data is obtained through a database log monitoring mechanism; Write the event trigger time data and the node resource status flow data into a distributed message queue; Extract the baseline decision vector calculated yesterday by the backbone AI model from the preset memory cache.

[0006] Furthermore, step two specifically includes: Set a uniform window length and sliding step size. Within the sliding time window, calculate the task conversion rate of the inventory goods based on the event trigger time data and the order conversion data obtained from the warehousing system in advance, and calculate the first time derivative of the task conversion rate. Based on the node resource status flow data, extract the local resource consumption rate and the amount of replenishment in transit, and calculate the difference between the two. Extract the traffic transfer coefficient of the associated object from the preset association rule configuration table; The transient perturbation feature matrix is ​​generated by concatenating the first-order time derivative of the task conversion rate, the difference, and the traffic transfer coefficient of the associated object.

[0007] Furthermore, step three specifically includes: Perform a Frobenius norm calculation operation on the transient perturbation feature matrix to obtain the matrix norm; Extract the mean and standard deviation of the matrix norm for the same period in the historical time period, and add the mean to the standard deviation by a preset multiple as the historical benchmark threshold.

[0008] Furthermore, step four involves generating and outputting updated control commands, specifically including: When the matrix norm is greater than the historical benchmark threshold, a preset lightweight matrix multiplication operator is triggered. The lightweight matrix multiplication operator is used to perform matrix multiplication between the transient perturbation feature matrix and the linear adjustment weight matrix to generate the incremental compensation instruction; The incremental compensation command is added to the baseline decision vector to generate the updated control command; The updated control instructions are converted into Hypertext Transfer Protocol Request format and sent to the warehouse execution terminal.

[0009] Furthermore, step five, which involves updating the linear adjustment weight matrix using a preset attenuation coefficient in the (0,1) interval with negative feedback suppression, specifically includes: Based on the business dimensions related to the task cancellation rate predefined by the system, the elements corresponding to the business dimensions in the linear adjustment weight matrix are extracted through the preset mapping index, and the corresponding elements are multiplied by the decay coefficient to obtain the updated linear adjustment weight matrix, which is used for the extrapolation calculation of the next sliding time window.

[0010] Furthermore, it also includes steps to prevent unauthorized access: A preset abnormal extreme value threshold is provided, wherein the abnormal extreme value threshold is greater than the historical same period benchmark threshold. When the matrix norm is greater than the historical benchmark threshold and less than or equal to the outlier threshold, the step of generating updated control instructions and outputting them to the warehouse execution terminal in step four is executed; when the matrix norm is greater than the outlier threshold, a conservative degradation instruction is generated to replace the updated control instructions. The transient disturbance feature matrix at this time is marked as abnormal traffic feature, and the abnormal traffic feature is output to the preset security risk control gateway for interception processing.

[0011] Furthermore, it also includes circuit breaker steps for extreme scenarios: Count the number of inventory items whose matrix norm is greater than the historical benchmark threshold within the current sliding time window; Calculate the proportion of the stated quantity to the total number of global inventory items; When the ratio exceeds the preset system global overload threshold, the active circuit breaker mechanism is triggered; In response to the active circuit breaker mechanism, the matrix multiplication operation is stopped, and the backbone artificial intelligence model is awakened to perform a global re-initialization operation; When the ratio is lower than or equal to the system global overload threshold, the generation and output of the updated control commands are maintained.

[0012] The beneficial effects of this invention are: 1) This invention proposes an incremental compensation mechanism based on transient perturbation feature matrix and baseline decision vector; it decouples low-frequency global decision and high-frequency local perturbation, avoids full recalculation of the backbone model under short-term fluctuations, significantly reduces computing power overhead, and achieves S-level rapid response to sudden business fluctuations and transient anomalies on the warehouse side. 2) This invention constructs a multi-source data acquisition and benchmark extraction link on a unified timeline; through application programming interface gateway, database log monitoring and distributed message queue, real-time behavior stream and warehouse resource status are spliced ​​together with low latency, eliminating time drift and statistical deviation caused by cross-system acquisition, and ensuring the timeliness of underlying data; 3) This invention establishes a multi-dimensional transient disturbance feature matrix; introduces the conversion rate time derivative, the supply and demand difference of local and in-transit resources and the related traffic transfer coefficient, and structures the original business flow to accurately distinguish between daily natural fluctuations and sudden disturbances, thereby realizing an explicit and comprehensive expression of short-term abnormal demand and resource shortage status. 4) This invention innovates a dynamic threshold discrimination method for abnormal disturbance intensity; it uses the Frobenius norm to reduce the dimensionality of multidimensional features into a comprehensive judgment index, and combines the historical average value and the preset variance to generate a benchmark threshold, thereby eliminating false triggering caused by time period differences and improving the stability and accuracy of the system's decision-making on sudden abnormal triggering. 5) This invention introduces a lightweight matrix multiplication operator to perform incremental compensation calculations; after confirming that the threshold has been exceeded, pricing and priority adjustment instructions are directly output through the matrix mapping resident in memory, without relying on deep inference services, and the execution actions are quickly generated and issued with extremely low computational cost, ensuring the absolute real-time performance of local business responses. Attached Figure Description

[0013] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating an intelligent scheduling decision-making method based on big data analysis provided in an embodiment of this application. Detailed Implementation

[0014] 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.

[0015] Please see Figure 1 This AI BI business decision-making method, based on big data analytics, is applied to a warehouse scheduling system that includes a warehouse execution terminal and dispatch vehicles. The target business object is inventory goods, including: Step 1: Collect the timestamps of user clicks on inventory items as event trigger time data, and the inventory deduction records of the warehousing system as node resource status flow data, and extract the baseline decision vector generated by the preset backbone artificial intelligence model; Step 2: Within a preset sliding time window, construct a transient disturbance feature matrix based on event trigger time data and node resource status transition data; Step 3: Calculate the matrix norm of the transient disturbance feature matrix and compare it with the preset historical benchmark threshold. Step 4: When the matrix norm is not greater than the historical benchmark threshold, the baseline decision vector is output as the control command; when the matrix norm is greater than the historical benchmark threshold, the transient disturbance feature matrix is ​​multiplied with the linear adjustment weight matrix pre-constructed using historical feature samples to obtain the incremental compensation command, which is superimposed on the baseline decision vector to generate the updated control command and output to the warehouse execution end. The baseline decision vector and the incremental compensation command are pre-divided into two control dimensions: resource allocation weight and picking priority. The control command includes the modification of resource allocation weight parameters and the instruction to increase the picking priority of the dispatching cart. Step 5: Within the next sliding time window after the updated control command is output, monitor the task cancellation rate and resource occupancy dwell time. When the task cancellation rate is greater than the preset cancellation rate threshold and the resource occupancy dwell time is greater than the preset duration threshold, use the preset attenuation coefficient in the (0,1) interval to perform negative feedback suppression update on the linear adjustment weight matrix. When the task cancellation rate is not greater than the preset cancellation rate threshold or the resource occupancy dwell time is not greater than the preset duration threshold, keep the linear adjustment weight matrix unchanged.

[0016] This embodiment provides a real-time business scheduling and control mechanism for retail warehouse scheduling; specifically, the main scenario is set as follows: a chain retail enterprise sets up a central warehouse in East China, and the target business object is a single product A of folding umbrellas; During the period from 17:40 to 18:10 on weekdays, the weather in the region changed suddenly, and the front-end mall saw a rapid increase in visits to and order intentions for inventory products. At the same time, the warehouse side saw changes in inventory deduction, picking space occupancy, and replenishment in transit. Instead of re-invoking the backbone model for every fluctuation, the system first uses the low-frequency generated baseline decision as the anchor point, and then performs local compensation based on short-term disturbances. In step one, the system collects two types of real-time data: one is the timestamp of the front-end user clicking on the inventory product details page, shopping cart page, or checkout page, which serves as the event trigger time data; Another type is the deduction, replenishment, and space release records of inventory corresponding to the inventory in the warehouse management system, which serve as node resource status flow data; at the same time, the baseline decision vector of the inventory has been calculated at midnight of the day or after the store closed the previous day is extracted from the preset backbone artificial intelligence model. The baseline decision vector can be expressed in a lightweight two-dimensional or three-dimensional form. For example, the first dimension can represent the resource allocation weight coefficient, and the second dimension can represent the picking priority coefficient of the dispatching cart. The default backbone AI model is a deep learning model based on a long short-term memory network and a fully connected layer. Its input features include: the daily average access time series of the target business object within a historical preset period, warehouse inventory turnover rate features, and average resource request amount per task features. The model uses feature vectors extracted from historical real warehouse scheduling logs and the corresponding human optimal scheduling decision results as labels. After supervised training and convergence through the backpropagation algorithm, the data is solidified and deployed to generate a baseline decision vector. For ease of understanding, assume that the baseline decision vector is [0.90, 3], which means that the current resource allocation weight adjustment rate for the corresponding pricing parameter modification instruction is 0.90, and the dispatching car corresponding to the dispatching car picking priority upgrade instruction, such as the automatic guided car, has a picking priority of level 3. In step two, the system extracts transient disturbance features within a sliding time window; instead of directly feeding all the original logs back into the backbone model, it extracts short-term structured features that reflect sudden changes and constructs a transient disturbance feature matrix. Taking a window with a length of 15 minutes and a step size of 1 minute as an example, within the window that ends at 18:00, the system can form a 3×2 matrix, where the 3 rows correspond to demand changes, supply deviations, and correlation migrations, respectively, and the 2 columns correspond to the statistical values ​​of the first half and the second half of the window, respectively. To illustrate the data flow, we can assume that the matrix is... The first line reflects increased demand, the second line reflects decreased local resource safety margins, and the third line reflects the migration of similar goods to inventory goods. In step three, the system calculates the matrix norm of the matrix and compares it with the historical benchmark threshold; the norm can be understood as a single-valued expression of the comprehensive disturbance intensity within the window. Taking the above matrix as an example, if the matrix norm is calculated to be 16.2, and the benchmark threshold for the same working day at 18:00 in the past 30 days is 11.5, then it can be determined that the inventory is in a transient abnormal fluctuation state that exceeds the preset fluctuation benchmark threshold. In step four, the system selects the control path based on the comparison results; when the matrix norm is lower than or equal to the threshold, it indicates that the current fluctuation is still within the range that can be covered by daily decisions, and the baseline decision vector can be output directly at this time; for example, in a certain window at 17:20, the norm is only 6.1, so [0.90, 3] remains unchanged; When the matrix norm is higher than the threshold, instead of calling a full recalculation, the transient perturbation feature matrix is ​​multiplied by the linear adjustment weight matrix to obtain the incremental compensation instruction. For ease of understanding, the linear adjustment weight matrix can be viewed as a lightweight mapper that maps disturbances to changes in pricing and priority. The initial value of the preset linear adjustment weight matrix is ​​obtained by using the least squares method to perform multiple linear regression fitting calculation based on real samples of transient disturbance characteristics and corresponding valid manual compensation instruction samples collected by the warehouse scheduling system during the historical testing phase. The value is preloaded into the memory configuration of the computing node during system initialization. Assuming the calculation result yields an incremental compensation instruction of [0.03, 2], the updated control instruction becomes [0.93, 5], indicating that the resource allocation weight adjustment rate increases from 0.90 to 0.93, and the picking priority increases from level 3 to level 5. The system outputs the result to the warehouse execution end, which is used as a resource allocation weight parameter modification instruction to synchronously modify the resource scheduling center parameters, and as a dispatch cart picking priority promotion instruction to update the picking task sorting in the scheduling system. In step five, the system performs closed-loop verification in the next sliding window after outputting the updated control command; here we focus on two feedback indicators: task cancellation rate and resource occupancy and resource occupancy retention time; The former corresponds to the proportion of canceled orders and released inventory space, while the latter corresponds to the duration of occupancy of locked storage locations, pallets, or cart resources. If both indicators exceed the threshold, it indicates that the compensation parameter is greater than the upper limit of resource carrying capacity, causing resource stagnation or deviation in the estimation of the demand system state. In this case, an interval is used. The attenuation coefficient is used to suppress the updating of the linear adjustment weight matrix; For example, the cancellation rate threshold is set to 8%, and the duration threshold is set to 12 minutes. If the next window detects a cancellation rate of 11% and a resource holding time of 16 minutes, the corresponding weights involved in the inventory compensation calculation will be multiplied by 0.8. If only one threshold is exceeded, or if neither threshold is exceeded, the weight matrix will remain unchanged to avoid excessive contraction. As an exception handling mechanism, if the event trigger time data arrives normally but the warehouse deduction log is temporarily missing, the system can temporarily use the node resource status flow data of the most recent complete window for temporary data replacement calculation, but this is only allowed to last for one window period. If multiple consecutive windows are missing, incremental compensation is stopped, and only baseline decisions are output to prevent incomplete supply information from amplifying erroneous scheduling. For example, if the amount of data in a window is too small, such as when the number of click events is less than the preset lower limit, the system can mark the window as a low-confidence window and will not start the compensation process. In the above rainstorm scenario, the number of clicks on umbrella item A in the store increased from 300 in the previous window to 780 in the 18:00 window, the actual available inventory in the warehouse dropped rapidly from 260pcs to 120pcs, and some of the visits to similar raincoat item B were transferred to umbrella item A. After the system calculates that the disturbance intensity exceeds the threshold, it does not re-execute the global AI inference. Instead, it directly adds incremental compensation to the cached baseline result. First, it adjusts the resource allocation weight parameter from 0.90 to 0.93 through the pricing parameter modification instruction. Then, it raises the picking priority of the dispatch cart corresponding to the shelf by two preset priority levels through the dispatch cart picking priority promotion instruction. If a large number of order cancellations occur in subsequent windows and the car remains in place for an extended period, the compensation weight will be automatically reduced to prevent continuous over-adjustment. The purpose of this step is to decouple low-frequency stable decision-making from high-frequency transient compensation, thereby enabling a rapid response to sudden business fluctuations on the warehouse side and controlling the computing power overhead caused by frequent calls to the backbone model.

[0017] In a preferred embodiment of the present invention, step one specifically includes: reading event trigger time data through a preset application programming interface gateway; obtaining node resource status flow data through a database log monitoring mechanism; writing the event trigger time data and node resource status flow data into a distributed message queue; and extracting the baseline decision vector calculated yesterday by the backbone artificial intelligence model from a preset memory cache.

[0018] This embodiment provides a data access and baseline extraction mechanism applicable to the aforementioned main scenario. Specifically, in the previous embodiment, although the sources of event trigger time data, node resource status flow data and baseline decision vector have been given, in high-concurrency scenarios, if the front-end direct connection to the database or the timed batch export method is still used, it is easy to encounter problems such as inconsistent collection delay, timing misalignment and cache miss, leading to high-frequency direct access to the underlying database. Therefore, this embodiment further refines the data collection process, enabling front-end logs, warehouse logs, and baseline results to be aligned on a unified timeline; For event trigger time data, the system reads it through a preset application programming interface gateway. This gateway does not handle complex reasoning; it only records access events containing product identifier, user terminal identifier, access action type, and millisecond-level timestamps when actions such as accessing the product details page, adding items to the shopping cart, or initiating checkout occur. For ease of explanation, assuming that three click events related to umbrella item A are received at 18:00:03, 18:00:05, and 18:00:07 respectively, the gateway can generate three messages with a uniform format, such as item A, click, 18:00:03; the system obtains node resource status flow data through the database log listening mechanism. Warehouse management systems typically generate insert, update, or delete actions in inventory tables, task tables, and placeholder tables. These actions can be reflected in the database logs in real time. The monitoring component does not directly scan the entire table, but only parses the change fragments related to the target field. For example, if the inventory record changes from 150 to 148, then the deduction can be extracted for item A, and the inventory is reduced by 2 pieces at 18:00:04; if the task slot changes from idle to locked, then it can be extracted for item A, and the task is automatically locked at 18:00:06. The system writes two types of data, from the gateway and the database log listener, into a distributed message queue. The message queue can be partitioned by product identifier or warehouse identifier to ensure that the same product is consumed in chronological order within the same partition. In the simplified example, partition P1 corresponds to product A. The click event at 18:00:03, the deduction event at 18:00:04, and the task lock event at 18:00:06 are sequentially entered into P1 and read by the stream processing engine in order. This design can reduce the window statistics deviation caused by seeing inventory changes before seeing click events. The baseline decision vector is extracted from the preset memory cache. Considering that this vector is calculated infrequently by the backbone model, accessing the offline model warehouse or querying a large table for each window calculation would introduce additional latency. Therefore, after the daily offline calculation is completed, the system writes the product identifier, baseline vector, and generation date to the memory cache. For example, if the result for product A yesterday is [0.90, 3], the cache validity period can cover the next day's business cycle; when a stream processing task needs to generate control instructions, it only needs to retrieve the vector directly based on the SKU identifier; As an exception fault tolerance mechanism, if the application programming interface gateway is temporarily unavailable, but the front-end business continues to provide product browsing services through the degraded link, the system can temporarily receive events that are buffered locally by the client and then retransmitted. However, events that are delayed for more than one window length are marked as late data and are not written back to the closed window to avoid disrupting the existing control instructions. If the database log listening is interrupted, the system can enable low-frequency polling of the inventory table to fill the gap, but this is only used for coarse-grained estimation during the recovery phase, and the log stream will be used as the primary reference after recovery. If the baseline vector of a certain product is missing in the memory cache, the inventory product will not perform incremental compensation within the current window, but will instead output a preset safe default vector, such as [1.00, 1], to avoid using null value operations; During the sudden rainstorm at the East China Central Warehouse, the front-end e-commerce platform generated a large number of access logs for umbrella product A between 18:00 and 18:01, and the warehouse database also showed deduction and placeholder logs simultaneously. The gateway sends front-end events into the user behavior routing channel of the message queue, and the database log listener sends inventory and task changes into the resource status routing channel. The two types of data are then aligned downstream using product identifiers and timestamps. Meanwhile, the streaming task directly retrieves the baseline vector [0.90, 3] of product A given by yesterday's backbone model from the memory cache, thus providing a starting point for subsequent window calculations; The purpose of this mechanism is to unify the data acquisition entry point and the baseline extraction entry point, so as to achieve low-latency splicing of real-time behavior flow, warehouse resource flow and low-frequency decision results, and reduce the time drift caused by cross-system data retrieval; In a preferred embodiment of the present invention, step two specifically includes: setting a uniform window length and sliding step size; within the sliding time window, calculating the task conversion rate of inventory goods based on event trigger time data and order conversion data obtained in advance from the warehousing system, and calculating the first time derivative of the task conversion rate; extracting the local resource consumption rate and the amount of replenishment in transit based on the node resource status flow data, and solving the difference between the two. Extract the traffic transfer coefficients of associated objects from the preset association rule configuration table; concatenate the first-order time derivative of the task conversion rate, the difference, and the traffic transfer coefficients of associated objects to generate a transient disturbance feature matrix.

[0019] This embodiment provides a mechanism for constructing a transient disturbance feature matrix. Specifically, although the previous embodiment was able to stably connect to the front-end event stream and the warehouse log stream, if the absolute values ​​such as click volume and inventory volume are directly spliced ​​together, it is still easy to blur the boundaries between daily fluctuations and sudden disturbance features, making it impossible to achieve effective feature isolation and resulting in inaccurate compensation triggering. Therefore, this embodiment introduces three types of features—rate of change, supply difference, and correlation migration—to make the matrix more focused on expressing what changes have occurred in a short period of time. Set a uniform window length and sliding step size; for example, set the window length to 15 minutes and the sliding step size to 1 minute; in this way, at 18:00, the system will collect data from 17:46 to 18:00; at 18:01, it will collect data from 17:47 to 18:01; this setting can balance response speed and statistical stability. Within this window, the system calculates the task conversion rate based on the event trigger time data and the pre-acquired order conversion data; the task conversion rate here can be understood as the ratio of the number of orders that trigger scheduling or outbound tasks to the number of valid visits or valid order placement intentions. In the simplified example, if product A receives 200 valid clicks and generates 40 valid outbound tasks during the first half of the window from 17:46 to 17:53, the conversion rate for the first half is 0.20; if there are 300 valid clicks and generates 105 valid outbound tasks during the second half of the window from 17:54 to 18:00, the conversion rate for the second half is 0.35. Therefore, the difference in conversion rate within this window is To strictly conform to the mathematical definition of the first-order time derivative, the system further divides the difference by the time span between the center points of the two sub-windows. In stream computing design, if the sub-window length is taken as one standardized observation period, that is, if the time span is 1, then the difference in the conversion rate is numerically equivalent to the first time derivative, that is, 0.15 divided by 1 equals 0.15. This structured approach ensures a closed loop of mathematical dimensions, eliminates the need for complex calculus derivations, and can intuitively reflect the rate of demand growth. The system extracts the local resource consumption rate and the amount of replenishment in transit based on the node resource status flow data, and calculates the difference. Assuming that within the same window, the local inventory of item A consumes an average of 12 pieces per minute, while replenishment items that have been dispatched but not yet received can be replenished at an average rate of 5 pieces per minute, then the difference between the two is: pcs / min; If the result is positive, it means that the demand is being consumed faster than the replenishment; if the result is negative, it means that the replenishment capacity is temporarily higher than the consumption. To maintain a consistent matrix representation, the sign direction can also be preserved. Next, the system extracts the traffic transfer coefficient of the associated objects from the preset association rule configuration table; this configuration table can be generated by business rules such as historical co-purchase, similar substitution, and page recommendation links. Specifically, the configuration table is generated by the system backend server periodically reading the order details database of all users within a preset historical time period, and using a priori association analysis algorithm to calculate the frequent itemsets and confidence scores between different product identifiers; the system extracts product combinations with confidence scores greater than a preset confidence threshold, and converts the confidence score value into a traffic transfer coefficient for related objects, and writes it into the distributed cache configuration table; For example, raincoat item B and umbrella item A are similar weather emergency products. The configuration table records a migration weight of 0.25 for item B to item A, which means that when item B is out of stock, the call threshold is increased, or the recommendation is weakened, 25% of the relevant traffic will be redirected to item A. In the current window, if the number of visits to product B page decreases while the number of visits to product A increases, the coefficient is extracted into the matrix as the third type of perturbation feature; after obtaining the above three types of features, the system concatenates them to generate a transient perturbation feature matrix. A simple and business continuity-compliant approach is to divide the window into two observation segments of equal length. By saving the feature state of the previous statistical period, each segment can independently form three types of features. Then, the feature column vectors corresponding to the two segments are horizontally concatenated in chronological order to construct a 3-row, 2-column transient perturbation feature matrix. Taking the calculation of the first time derivative as an example, the system can extract the first time derivative of the first half of the current window as 0.05, which is the growth rate of the first half relative to an earlier period, while the first time derivative of the second half of the current window is 0.15, which is the difference between the conversion rate of 0.35 in the second half and 0.20 in the first half in the example above. Combining the corresponding supply difference and migration coefficient, if the three complete features in the first half form a column vector of [0.05, 4, 0.10] and the three features in the second half form a column vector of [0.15, 7, 0.25], then by horizontally concatenating the columns, the transient disturbance feature matrix can be strictly and clearly written as [[0.05, 0.15], [4, 7], [0.10, 0.25]]. If the actual system requires finer granularity, it can also be expanded into a 3×15 matrix with one column per minute, as long as the downstream norm calculation and weight mapping dimension remain consistent. As an exception handling mechanism, if order conversion data is missing in a certain window but event trigger data is complete, the system can mark the window as a conversion pending state, temporarily use the conversion rate of the previous window as an approximation, and reduce the confidence level of the window. If there is no associated item for the target product in the association rule configuration table, the migration coefficient is set to zero directly instead of using a null value; if the in-transit replenishment data is not updated due to transportation system delays, only the most recent valid replenishment value is used within one window, and matrix generation stops and reverts to the baseline output after the time limit is exceeded. During the 18:00 window when the rainstorm intensified, the conversion rate of product A was 0.20 in the first half and increased to 0.35 in the second half, resulting in an increase of 0.15; the local resource consumption rate was 12 pcs per minute, and the replenishment in transit was 5 pcs per minute, with a difference of 7; the traffic transfer coefficient of product B to product A was 0.25. The system combines these three indicators to generate a disturbance matrix that reflects the current local outbreak situation. Compared to just looking at the total number of clicks, this matrix is ​​better able to indicate the simultaneous occurrence of three phenomena: rapidly increasing demand, supply surplus below the preset safety threshold, and the convergence of alternative traffic. The purpose of this step is to reconstruct the original business flow into a disturbance carrier suitable for rapid identification, thereby enabling an explicit expression of short-term abnormal demands and resource shortages.

[0020] In a preferred embodiment of the present invention, step three specifically includes: performing a Frobenius norm calculation operation on the transient disturbance feature matrix to obtain the matrix norm; extracting the mean and standard deviation of the matrix norm for the same period in the historical time period, and adding a preset multiple of the standard deviation to the mean as the historical benchmark threshold for the same period.

[0021] This embodiment provides a threshold calculation mechanism for judging disturbance intensity; specifically, in the previous embodiment, the transient disturbance feature matrix can already characterize short-term changes, but without a unified compression index, it is difficult to quickly determine whether the current fluctuation is sufficient to trigger compensation. If the three types of features are compared separately, it is easy for one type of indicator to be accidentally too large and trigger false alarms. Therefore, this embodiment uses matrix norm to merge multidimensional perturbations into a directly comparable intensity value, and further introduces historical thresholds to avoid seasonal deviations between different time periods. The system performs Frobenius norm calculation on the transient disturbance characteristic matrix; this calculation can be understood as: squaring each element in the matrix, summing the results, and then taking the square root to obtain an overall value. Taking the aforementioned matrix [[0.05, 0.15], [4, 7], [0.10, 0.25]] as an example, we can first calculate the sum of squares of the elements, that is... Taking the square root gives approximately 8.07; The larger this value, the more obvious the superposition of multiple perturbations within the window; this norm is used here because it has low requirements for matrix shape and is suitable for fast computation in streaming environments. Norms alone are not enough, because the natural fluctuation levels of weekday morning and evening peak hours and early morning low hours are different; therefore, the system extracts the mean and variance of the matrix norm for the same period in historical time periods, and uses the mean plus a preset multiple of the variance as the historical benchmark threshold for the same period. It should be clarified that in traditional mathematical statistics, the dimension of the mean is a power of one, while the dimension of the variance is a power of two. Directly adding the two results in a problem of inconsistent physical dimensions. However, in the context of streaming risk control, in order to avoid the extra computing power overhead caused by high-frequency square root operations on historical variance, this system innovatively assigns a unit to the preset multiple: the preset multiple is given a scaling factor attribute with a unit correction reciprocal. For example, setting the preset multiple parameter to 2 essentially involves mapping the second power of the variance to reduce the dimension, so that the product can directly form a safe redundancy that can be added to the mean in numerical calculation. In the simplified explanation, if the mean norm of item A is 5.2 and the variance is 1.3 within the same window from 17:46 to 18:00 over the past 30 working days, based on the above dimensionality reduction mapping and a preset multiplier of 2, then the benchmark threshold is directly calculated as follows: Thus, when the current window norm of 8.07 is greater than 7.8, it can be determined that it exceeds the normal fluctuation range of the same period in history. The synchronous sampling can be selected based on clock time, weekday attribute, promotional period type, or warehouse area attribute; for example, different synchronous sample pools can be established for regular weekdays, weekend business peaks, and the eve of holidays; if a more granular classification is used, the problem of threshold being diluted by global averaging can be reduced; As an anomaly tolerance mechanism, if there are insufficient historical samples for a certain product, such as a new product that has been on the market for less than 7 days, the system can first use the average mean and average variance of products in the same category as a temporary threshold benchmark; after the samples have accumulated to the preset number of days, it can then switch to the threshold of the inventory product itself. If the historical variance is extremely small, causing the threshold to be too close to the mean, the system can further set a minimum safety interval, such as adding a fixed offset based on the mean, to prevent repeated triggering due to slight fluctuations. If there are empty columns or abnormally missing columns in the window matrix, the current window will not participate in the threshold comparison, but will be directly recorded as an incomplete window; In the days leading up to the rainstorm, umbrella item A usually experiences a natural temperature rise around 18:00, so its mean norm is not close to zero, but rather about 5.2. After calculating 8.07 at 18:00 that day, the system did not directly determine that the value was too large and therefore abnormal. Instead, it judged it within the historical scale for the same period. Since 8.07 had exceeded 7.8, the system determined that the current period was in the sudden enhancement range and allowed it to enter the subsequent compensation path. If the same 8.07 appears in the historical context of the Double Eleven promotion, and the threshold for that period is above 12, compensation will not be triggered. The purpose of this step is to uniformly map multidimensional disturbances into single-value indicators that can be compared with thresholds, and to eliminate time period differences through historical correction, thereby achieving a more stable trigger judgment.

[0022] In a preferred embodiment of the present invention, step four of generating and outputting the updated control command specifically includes: when the matrix norm is greater than the historical benchmark threshold, triggering a preset lightweight matrix multiplication operator; using the lightweight matrix multiplication operator to perform matrix multiplication of the transient disturbance feature matrix and the linear adjustment weight matrix to generate an incremental compensation command. The incremental compensation command is added to the baseline decision vector to generate the updated control command; the updated control command is then converted into a Hypertext Transfer Protocol request format and sent to the warehouse execution terminal.

[0023] This embodiment provides a lightweight compensation calculation and execution output mechanism; specifically, in the previous embodiment, the system has determined whether a certain window exceeds the historical threshold, but if the backbone model is still called to recalculate after exceeding the threshold, the problem of high latency and high computing power consumption will return. Therefore, this embodiment introduces a lightweight matrix multiplication operator to quickly map the overthreshold window into executable incremental compensation instructions; When the matrix norm is greater than the historical benchmark threshold, the system triggers a preset lightweight matrix multiplication operator. This operator can be deployed in the stream processing node to perform resident memory calculations on small-dimensional matrices without relying on external deep inference services. In order to strictly follow the feature dimension mapping rules and ensure the transparency and traceability of the calculation logic between features and control quantities, the system directly performs matrix multiplication on the aforementioned 3-row 2-column transient disturbance feature matrix and the preset 2-row 3-column linear adjustment weight matrix. For ease of explanation, taking the aforementioned transient feature matrix as [[0.05, 0.15], [4, 7], [0.10, 0.25]] as an example, the linear adjustment weight matrix is ​​[[0.10, 0.002, 0.05], [1, 0.15, 0.5]], where the first row is used to map the price change and the second row is used to map the picking priority change; According to the rules of matrix multiplication, the multiplication results in a 2x2 initial compensation matrix. In the specific calculation, the second column of this compensation matrix represents the latest perturbation result of the second half of the window, and its first dimension is "". The second dimension is ; Since the control instructions need to cover the business responses of both pricing and priority dimensions, in order to generate the final execution and include multi-dimensional incremental compensation instructions, and to avoid confusion with single-element one-dimensional scalars at the terminology level, the system directly extracts the second column vector of the result matrix as the incremental compensation instruction; after business rule constraints and rounding, it can be output as [0.04, 1] or [0.04, 2]. The system superimposes the incremental compensation instruction onto the baseline decision vector; if the baseline decision vector is [0.90, 3], then the updated control instruction can be [0.94, 4]; the superposition here can be a simple addition or a boundary clipping applied to each dimension; for example, the pricing coefficient is not allowed to exceed 1.00, and the priority is not allowed to be higher than level 7; If the sum exceeds the upper limit, it will be automatically truncated to the upper limit value; after receiving the updated control command, the system will convert it into a Hypertext Transfer Protocol request format and send it to the warehouse execution terminal. In the simplified example, two requests corresponding to the two execution dimensions can be generated: one is sent to the resource call parameter center interface as a pricing parameter modification instruction, which is used to change the resource allocation weight adjustment rate of single product A from 0.90 to 0.94; Another instruction is sent to the scheduling system interface as a priority upgrade instruction for the dispatching cart, which is used to change the priority of the dispatching cart for the location of item A to level 4; In addition to the target parameters, the request can also include the window number, product identifier, instruction generation time, and validity period, which facilitates idempotent processing and backtracking auditing at the execution end. As an exception handling mechanism, if an outlier occurs in the matrix multiplication result, such as a dimension becoming a value far exceeding the business limit due to input contamination, then amplitude clipping and validity verification are performed before output. If the resource call parameter center interface returns failure while the scheduling interface succeeds, the system can mark this instruction as partially executed and retry the failed dimension in the next window, without rolling back the successful dimension; If the warehouse execution terminal times out, the result generated in the current window can be retained, but it will not be repeatedly issued more than a preset number of times to avoid generating redundant control commands with high frequency of concurrency. During the 18:00 window of the peak of the rainstorm, after item A is determined to exceed the threshold, the lightweight matrix multiplication operator runs immediately to generate a compensation value relative to the baseline; based on this, the system adjusts the resource allocation weight adjustment rate from 0.90 to 0.94 and increases the picking priority of the dispatching cart from 3 to 4 or 5. The entire process does not rely on re-executing a full AI inference. Instead, it directly maps local disturbances into executable control actions and then sends them to the resource call parameter center and warehouse management system scheduling interface via the Hypertext Transfer Protocol. The purpose of this mechanism is to generate incremental control actions quickly with low computational cost in the case of exceeding the threshold, thereby achieving local scheduling and parameter adjustment with a response time of seconds.

[0024] In a preferred embodiment of the present invention, step five, which involves updating the linear adjustment weight matrix using a preset attenuation coefficient in the (0,1) interval with negative feedback suppression, specifically includes: Based on the business dimensions related to task cancellation rate predefined by the system, the elements corresponding to the business dimension in the linear adjustment weight matrix are extracted through the preset mapping index. The corresponding elements are then multiplied by the decay coefficient to obtain the updated linear adjustment weight matrix, which is used for the extrapolation calculation of the next sliding time window.

[0025] This embodiment provides a negative feedback suppression mechanism for overcompensation. Specifically, in the previous embodiment, lightweight matrix multiplication can quickly provide compensation results, but if certain business dimensions are too sensitive to disturbances, they may be repeatedly amplified within a continuous window, leading to system overshoot or control oscillation phenomena such as increased cancellation rate and increased resource retention. Therefore, this embodiment further attenuates local elements of the weight matrix based on the feedback results, rather than resetting the entire matrix; The system monitors the task cancellation rate and resource occupancy dwell time in the next sliding window after the output control command; if both exceed the threshold at the same time, it indicates that the compensation in the previous window does not match the actual handling capacity. For example, in the 18:01 window, although the click conversion rate of product A increased after compensation, there was a 12% cancellation rate, and the related auto-guided cart occupied an average of 16 minutes, exceeding the preset thresholds of 8% and 12 minutes. At this time, negative feedback suppression was activated. To avoid indiscriminately reducing overall compensation capacity, the system locates relevant elements in the weight matrix based on the business dimension corresponding to the task cancellation rate through a preset mapping index; this mapping index can be understood as a comparison relationship of which decision dimension mainly causes the cancellation phenomenon. For example, if cancellations mainly occur for orders sensitive to call thresholds, the index points to pricing-related weights; if cancellations mainly occur for timeout orders caused by warehouse congestion, the index points to scheduling priority-related weights. Assuming that the first row of the weight matrix above corresponds to the pricing dimension and the second row corresponds to the priority dimension, and since the problem in this case mainly stems from over-scheduling, only the elements in the second row related to the supply difference and migration coefficient are selected for updating; Then, the system multiplies the selected element by a decay coefficient between zero and one; for example, if the relevant element in the second row of the original matrix is ​​[1, 0.15, 0.5], and the decay coefficient is 0.8, then it can be updated to [0.8, 0.12, 0.4]. Elements not matched by the mapped index remain unchanged; in the next window, the system continues to use the updated matrix for deduction, thereby automatically reducing the compensation magnitude under similar conditions; This local decay method is better than a global reset; if the entire weight matrix is ​​directly restored to its initial value, although the risk can be suppressed, the previously adapted dimensional information will be lost; if no adjustment is made, the same overcompensation may be repeated continuously in a continuous window. As an exception handling mechanism, if the task cancellation rate exceeds the threshold but the resource retention time does not exceed the threshold, the system can determine that the authenticity of the demand is lower than the preset conversion benchmark rather than the scheduling resources reaching the upper limit of the slot. In this case, only the attenuation of relevant elements in the pricing dimension is considered. If the resource holding time exceeds the threshold but the cancellation rate does not exceed the threshold, it means that the order is still valid, but the execution pace is slow. In this case, you can reduce the priority increase. If the same element is decayed multiple times in a row, the system can set a minimum lower limit, such as not lower than 20% of the initial value, to prevent the compensation capability from failing for a long time. In the rainstorm scenario, the system prioritized umbrella item A at a higher level within the 18:00 window. As a result, at the 18:01 window, it was found that the associated shelf cart was congested, and some orders were canceled due to waiting timeout. The system identified the problem as originating from the scheduling dimension rather than the price dimension through the mapping index. Therefore, it only multiplied the priority-related weight by 0.8, while keeping the pricing-related weight unchanged. At the 18:02 window, even if the disturbance is still greater than the baseline threshold, the new incremental compensation will use a smaller linear adjustment weight value to avoid the same shelf repeatedly reaching the occupancy limit. The purpose of this mechanism is to feed back the posterior business results into the feedforward compensation parameters, thereby achieving adaptive suppression of local over-adjustment.

[0026] In a preferred embodiment of the present invention, the method further includes the step of preventing abnormal access: setting an abnormal extreme value threshold, wherein the abnormal extreme value threshold is greater than the historical benchmark threshold; when the matrix norm is greater than the historical benchmark threshold and less than or equal to the abnormal extreme value threshold, the step of generating an updated control instruction and outputting it to the warehouse execution terminal in step four is executed. When the matrix norm is greater than the abnormal extreme value threshold, a conservative degradation instruction is generated to replace the updated control instruction; the transient disturbance feature matrix at this time is marked as an abnormal flow feature, and the abnormal flow feature is output to the preset security risk control gateway for interception processing.

[0027] This embodiment provides an abnormal access prevention mechanism based on perturbation extreme values. Specifically, in the previous embodiment, as long as the matrix norm exceeds the historical threshold, it will enter the compensation path. However, in real business, extremely high-frequency access does not always mean a real demand explosion. It may also come from high-frequency concurrent requests from malicious crawlers, fictitious transaction instructions, or abnormal order locking. If the system misjudges such abnormal traffic as real demand, it may incorrectly raise prices and increase scheduling priority, thereby disrupting the warehousing order. Therefore, this embodiment adds a second layer of discrimination between normal exceedance and extreme abnormality. The system presets an abnormal extreme value threshold, which is higher than the historical benchmark threshold for the same period. The method for calculating the abnormal extreme value threshold is as follows: The system extracts log data from historical security audit logs that are clearly marked as suffering from malicious network attacks or unauthorized high-frequency fake transaction requests by high-frequency crawlers, replays them to construct the corresponding transient disturbance feature matrix, calculates the statistical mean of the norm of the known abnormal matrix, and sets the mean as the abnormal extreme value threshold. In the simplified example, if the historical baseline threshold is 7.8, then the extreme value threshold can be set to 15.0; in this way, the matrix norm of the current window will fall into three intervals: the normal interval not exceeding 7.8, the business surge interval between 7.8 and 15.0, and the extreme value interval exceeding 15.0. When the matrix norm is greater than the historical benchmark threshold and less than or equal to the outlier threshold, the system will still perform compensation calculations according to the established procedure. This indicates that the system believes that although the current fluctuations are large, they may still be a real business peak that can be handled. Conversely, when the matrix norm is greater than the abnormal extreme value threshold, the system no longer outputs the usual update control instructions, but generates conservative downgrade instructions. Conservative downgrade can include freezing price increases, stopping further increases in picking priority, maintaining only the baseline decision, or switching to the manual review channel. For example, if the resource allocation weight adjustment rate should be adjusted from 0.90 to 0.95 and the priority from 3 to 5, but the norm reaches 18.6, the system will instead keep the rate unchanged at 0.90, only keep the priority at 3, or only allow a slight increase to 4. At the same time, the system marks the corresponding transient disturbance feature matrix as abnormal traffic features and outputs it to the security risk control gateway; the security risk control gateway can further combine it with other security signals such as Internet Protocol address access frequency, terminal fingerprint, account association, and order behavior consistency for interception processing; In this implementation, the BI decision system does not need to complete all security identification on its own. It only needs to hand over the matrix form that has been confirmed on the business side as having a norm greater than the outlier threshold to the risk control side for review. The introduction of an abnormal extreme value threshold here is an enhancement to the aforementioned scheme; relying solely on a single threshold trigger will confuse genuine surges with malicious shocks; by adding a second layer of extreme value judgment, the system can distinguish between the two types of high volatility sources without significantly increasing computational complexity. As an anomaly tolerance mechanism, if the norm of a certain window matrix is ​​exactly equal to the anomaly extreme value threshold, it will be processed according to the first preset degradation control logic, and the update control instruction can still be executed, but an additional check can be added, such as requiring that the proportion of independent users exceeds the lower limit at the same time. If the security risk control gateway is temporarily unavailable, the system will still perform conservative downgrade, but the abnormal traffic characteristics will be written to disk locally and re-uploaded after the gateway is restored; if a product falls into the abnormal extreme value range in multiple consecutive windows, the inventory product can be added to the temporary observation list and a conservative strategy will be uniformly adopted for a period of time. If visits to umbrella product A do increase during heavy rain, but mainly come from scattered users and the ordering rhythm is consistent with the inventory deduction rhythm, then the matrix norm may fall between 10 and 12, and the system will perform compensation normally. If thousands of highly concentrated price queries and order locking requests suddenly occur, and the actual payment conversion rate is lower than the preset benchmark value, the first-order change and resource occupation will be abnormally amplified at the same time, and the matrix norm may jump to more than 18. At this point, the system no longer treats this as a regular high-conversion-rate concurrent business request, but instead outputs a conservative degradation instruction and sends the window matrix as an abnormal traffic feature to the security risk control gateway for interception. The purpose of this mechanism is to embed an abnormal access identification threshold into the business compensation chain, thereby achieving passive isolation of malicious high-frequency traffic and avoiding incorrect scheduling and pricing.

[0028] In a preferred embodiment of the present invention, an extreme scenario circuit breaker step is further included: counting the number of inventory goods whose matrix norm is greater than the historical benchmark threshold within the current sliding time window; and calculating the proportion of the number to the total number of global inventory goods. When the ratio exceeds the preset global system overload threshold, the active circuit breaker mechanism is triggered. In response to the active circuit breaker mechanism, matrix multiplication operations are stopped, and the backbone artificial intelligence model is awakened to perform global re-initialization operations. When the ratio is lower than or equal to the global system overload threshold, the generation and output of updated control commands are maintained.

[0029] This embodiment provides a circuit breaker mechanism for global abnormal shocks. Specifically, the aforementioned scheme is suitable for handling sudden fluctuations in local SKUs, local warehouse areas, or local time periods. However, if a large number of goods exceed the threshold simultaneously in the global scope, such as a surge in request concurrency after the global communication interruption is restored, a supply chain interruption caused by a major natural disaster, or a platform-level activity configuration error, the assumption based on local linear compensation may no longer hold. If matrix multiplication compensation is still performed on each product individually, the error will be amplified when the system becomes unstable. Therefore, this embodiment adds an active circuit breaker step to identify overload or failure of global system nodes and switch to the global re-initialization path. At the end of each sliding window, the system counts the number of inventory items in the current window whose matrix norm is greater than the historical benchmark threshold. For example, if the East China central warehouse currently has 10,000 monitored products, and 6,200 of these products have disturbance norms exceeding their respective thresholds within the 18:00 window, then the number of products exceeding the threshold is 6,200; the system then calculates the proportion of this number to the total number of products in global inventory, i.e., 6,200 / 10,000 = 62%. The system compares this ratio with the preset global overload threshold. If the global overload threshold is set to 60%, then 62% has already exceeded the threshold, indicating that the current situation is not a localized sudden event, but rather a near-global simultaneous instability. At this point, the active circuit breaker mechanism is triggered. After triggering, the system stops executing matrix multiplication operations for individual products, stops generating update control instructions based on local compensation, and instead calls the backbone artificial intelligence model to perform global re-initialization calculations. This re-initialization may include reloading the latest global features, refreshing cross-warehouse inventory status, rebuilding regional demand profiles, and notifying human operators to intervene when necessary. If the proportion is lower than or equal to the system's global overload threshold, it means that most products are still in the normal or locally controllable range, and the system continues to maintain the generation and output of the aforementioned update control instructions; for example, if only 800 out of 10,000 products exceed the threshold, accounting for 8%, then there is no need to trigger the circuit breaker, and the local compensation strategy will still be used. This design embodies a hierarchical shift from local adaptation to global protection; the aforementioned implementations assume that the overall system background is relatively stable, with only some products experiencing significant disturbances; while this embodiment acknowledges that this premise may be violated in extreme cases and actively stops local compensation to avoid error accumulation; As an anomaly tolerance mechanism, if the statistical caliber of the total number of products fluctuates, for example, if some products are not listed on the shelves or are temporarily removed from the shelves on a certain day, the denominator should be based on the number of products currently under monitoring and with a valid baseline vector. If the system exceeds the global overload threshold in two consecutive windows, in addition to calling the main model, the new compensation request can be locked until the global re-initialization is completed; if the ratio is exactly equal to the global overload threshold, it can be handled with a conservative strategy, that is, it is considered to be still within the controllable boundary, and the original output is maintained, but a warning prompt is issued at the same time. In heavy rain scenarios, only a few categories such as rain gear and waterproof shoe covers are usually affected. At this time, the proportion of products exceeding the threshold is usually less than 10%, and the system can continue to use local compensation. If, due to subsequent regional communication failures, marketing platform configuration errors, or large-scale delays in the upstream inventory system, thousands of SKUs simultaneously exhibit abnormal norms within the same window, the percentage exceeding the threshold may quickly exceed 60%. Based on this, the system determines that the applicable conditions of the local compensation model have failed, immediately stops adding incremental instructions to each SKU one by one, and instead calls the main model to perform a global re-initialization. If necessary, it switches control to the manual takeover process. The purpose of this mechanism is to set clear applicable boundaries and system-level safety and fault tolerance guarantees for local compensation algorithms, thereby achieving overall stability protection under extreme operating conditions.

[0030] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An AI BI business decision-making method based on big data analytics, applied to a warehouse scheduling system including a warehouse execution terminal and a dispatch vehicle, characterized in that, The target business object is inventory goods, including: Step 1: Collect the front-end user click timestamps of the inventory goods as event trigger time data and the inventory deduction records of the warehousing system as node resource status flow data, and extract the baseline decision vector generated by the preset backbone artificial intelligence model; Step 2: Within a preset sliding time window, construct a transient disturbance feature matrix based on the event trigger time data and the node resource status transition data; Step 3: Calculate the matrix norm of the transient disturbance feature matrix and compare it with a preset historical benchmark threshold. Step 4: When the matrix norm is not greater than the historical benchmark threshold, the baseline decision vector is output as a control command; when the matrix norm is greater than the historical benchmark threshold, the transient disturbance feature matrix and the linear adjustment weight matrix pre-constructed using historical feature samples are multiplied to obtain an incremental compensation command, which is then superimposed on the baseline decision vector to generate an updated control command, which is output to the warehouse execution end. The baseline decision vector and the incremental compensation command are both pre-divided into two control dimensions: resource allocation weight and picking priority. The control command includes instructions for modifying resource allocation weight parameters and increasing the picking priority of the dispatch cart. Step 5: Within the next sliding time window after the updated control command is output, monitor the task cancellation rate and resource occupation dwell time. When the task cancellation rate is greater than a preset cancellation rate threshold and the resource occupation dwell time is greater than a preset duration threshold, use the preset attenuation coefficient in the (0,1) interval to perform negative feedback suppression update on the linear adjustment weight matrix. When the task cancellation rate is not greater than the preset cancellation rate threshold or the resource occupation dwell time is not greater than the preset duration threshold, keep the linear adjustment weight matrix unchanged.

2. The AI ​​BI business decision-making method based on big data analysis according to claim 1, characterized in that, Step one specifically includes: The event trigger time data is read through a preset application programming interface gateway; The node resource status flow data is obtained through a database log monitoring mechanism; Write the event trigger time data and the node resource status flow data into a distributed message queue; Extract the baseline decision vector calculated yesterday by the backbone AI model from the preset memory cache.

3. The AI ​​BI business decision-making method based on big data analysis according to claim 2, characterized in that, Step two specifically includes: Set a uniform window length and sliding step size. Within the sliding time window, calculate the task conversion rate of the inventory goods based on the event trigger time data and the order conversion data obtained from the warehousing system in advance, and calculate the first time derivative of the task conversion rate. Based on the node resource status flow data, extract the local resource consumption rate and the amount of replenishment in transit, and calculate the difference between the two. Extract the traffic transfer coefficient of the associated object from the preset association rule configuration table; The transient perturbation feature matrix is ​​generated by concatenating the first-order time derivative of the task conversion rate, the difference, and the traffic transfer coefficient of the associated object.

4. The AI ​​BI business decision-making method based on big data analysis according to claim 3, characterized in that, Step three specifically includes: Perform a Frobenius norm calculation operation on the transient perturbation feature matrix to obtain the matrix norm; Extract the mean and standard deviation of the matrix norm for the same period in the historical time period, and add the mean to the standard deviation by a preset multiple as the historical benchmark threshold.

5. The AI ​​BI business decision-making method based on big data analysis according to claim 4, characterized in that, Step four involves generating and outputting updated control commands, specifically including: When the matrix norm is greater than the historical benchmark threshold, a preset lightweight matrix multiplication operator is triggered. The lightweight matrix multiplication operator is used to perform matrix multiplication between the transient perturbation feature matrix and the linear adjustment weight matrix to generate the incremental compensation instruction; The incremental compensation command is added to the baseline decision vector to generate the updated control command; The updated control instructions are converted into Hypertext Transfer Protocol Request format and sent to the warehouse execution terminal.

6. The AI ​​BI business decision-making method based on big data analysis according to claim 5, characterized in that, Step five, which involves updating the linear adjustment weight matrix using a pre-set attenuation coefficient in the (0,1) interval with negative feedback suppression, specifically includes: Based on the business dimensions related to the task cancellation rate predefined by the system, the elements corresponding to the business dimensions in the linear adjustment weight matrix are extracted through the preset mapping index, and the corresponding elements are multiplied by the decay coefficient to obtain the updated linear adjustment weight matrix, which is used for the extrapolation calculation of the next sliding time window.

7. The AI ​​BI business decision-making method based on big data analysis according to claim 6, characterized in that, It also includes steps to prevent unauthorized access: A preset abnormal extreme value threshold is provided, wherein the abnormal extreme value threshold is greater than the historical same period benchmark threshold. When the matrix norm is greater than the historical benchmark threshold and less than or equal to the outlier threshold, the step of generating updated control instructions and outputting them to the warehouse execution terminal in step four is executed; when the matrix norm is greater than the outlier threshold, a conservative degradation instruction is generated to replace the updated control instructions. The transient disturbance feature matrix at this time is marked as abnormal traffic feature, and the abnormal traffic feature is output to the preset security risk control gateway for interception processing.

8. The AI ​​BI business decision-making method based on big data analysis according to claim 7, characterized in that, It also includes circuit breaker steps in extreme scenarios: Count the number of inventory items whose matrix norm is greater than the historical benchmark threshold within the current sliding time window; Calculate the proportion of the stated quantity to the total number of global inventory items; When the ratio exceeds the preset system global overload threshold, the active circuit breaker mechanism is triggered; In response to the active circuit breaker mechanism, the matrix multiplication operation is stopped, and the backbone artificial intelligence model is awakened to perform a global re-initialization operation; When the ratio is lower than or equal to the system global overload threshold, the generation and output of the updated control commands are maintained.

Images