An AI-based ERP data multi-dimensional analysis and evaluation method

Abstract

This application provides an AI-based method for multi-dimensional analysis and evaluation of ERP data, including: extracting a unique dynamic root anchor point from the entire purchase order chain, obtaining the hash association between the corresponding random seed and timestamp, and generating a shared time series benchmark for simulating branch generation; transmitting the time series alignment benchmark calibration result back to the fund tracking link, supplementing the real-time verification signal under the scenario of uneven node latency, and generating a closed-loop calibration strategy adapted to the real-time state of the fund tracking link; adjusting the preset AI purchase prediction model for time series and link adaptation through the closed-loop calibration strategy, quantitatively defining the risk boundary range of inventory deviation, and obtaining the optimized inventory prediction result adapted to the real-time state of the link; performing multi-dimensional data fusion of the optimized inventory prediction result and the fund path feedback signal to obtain the comprehensive risk score of the anomaly detection module, quantitatively determining whether the risk of fund misappropriation is within the preset controllable range, and obtaining the final risk assessment report.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to an AI-based method for multi-dimensional analysis and evaluation of ERP data. Background Technology

[0002] In the areas of procurement forecasting and financial anomaly detection within Enterprise Resource Planning (ERP) systems, the stability of risk control and business collaboration is paramount. This area directly impacts the security of inventory management, cash flow, and the reliability of business decisions. With the expansion of business scale and the increase in transaction complexity, ensuring the coordinated operation of procurement simulation and cash flow monitoring in a dynamic environment has become a pressing challenge for enterprise management. Currently, many solutions often separate procurement forecasting and financial anomaly detection, neglecting their close connection in actual business operations. This approach is inadequate when facing frequently changing procurement demands and complex cash flow environments, especially in multi-party collaborative cash flow processes. The system cannot effectively adapt to time differences at different stages, leading to distortions in information transmission and feedback, thus affecting the reliability of overall decision-making. There is a lack of a unified benchmark to coordinate the temporal and data correspondence between the procurement forecasting simulation process and the real-time monitoring of cash flow. This deficiency makes simulation results prone to deviating from actual business trajectories, and time delays in cash flow further exacerbate this deviation. Furthermore, existing AI-powered procurement forecasting models cannot adapt to the real-time status of the cash flow, and corrected simulation data cannot be reused in the benchmark calibration stage. Risk assessment also fails to integrate inventory forecasting with cash flow signals. For example, in a large procurement project, the system may simulate multiple procurement plans based on historical data, but because it fails to link these plans to the specific order's timing, the simulation results may be completely inconsistent with the actual order execution. Simultaneously, the cash payment process involves multiple external interfaces with varying time delays, preventing the system from obtaining complete cash flow information in a timely manner. This ultimately leads to a severe disconnect between forecast results and actual conditions, resulting in chaos in inventory planning and cash allocation. Therefore, establishing a unified and coordinated time benchmark for procurement forecasting simulation and cash flow monitoring in a dynamic and ever-changing business environment, ensuring the timing alignment and collaborative operation of procurement simulation and cash monitoring, has become a critical issue that urgently needs to be addressed. Summary of the Invention

[0003] This invention provides an AI-based method for multi-dimensional analysis and evaluation of ERP data, mainly including: Extract the unique dynamic root anchor point of the entire purchase order chain, obtain the hash association between the random seed and the timestamp, and construct a shared time series benchmark generated by the simulated branch; The actual delay of the entire link node of the fund flow is measured according to the shared time series benchmark. The tracking timeout threshold of each node is adjusted according to the actual delay. Feedback signals are collected based on the tracking timeout threshold. The integrity of the fund path feedback signal is determined according to the matching degree between the feedback signal and the execution time series of the entire link node of the purchase order. When the integrity of the funding path feedback signal is lower than a preset integrity threshold, for the corresponding purchase order, the abnormal distortion simulation branch corresponding to the feedback signal that is lower than the preset integrity threshold is removed, the bound random seed is reset, and the corrected purchase prediction simulation data is output. Extract historical procurement multi-path simulation results, combine them with the corrected procurement prediction simulation data, eliminate contamination and interference branches, and output the time-series alignment benchmark calibration results. Based on the time-series alignment benchmark calibration results, extract the real-time verification signal and construct a closed-loop calibration strategy based on the real-time verification signal. The preset AI procurement forecasting model is adjusted in terms of timing and link according to the closed-loop calibration strategy, the risk boundary range of inventory deviation is defined, and the optimized inventory forecasting result is obtained based on the risk boundary range. The optimized inventory forecast results are combined with the feedback signal to obtain a comprehensive risk score. The comprehensive risk score is compared with the risk controllable range threshold, and risk assessment data is output.

[0004] Furthermore, the extraction of the unique dynamic root anchor point across the entire purchase order chain, obtaining the hash association between the random seed and the timestamp, and constructing a shared time-series benchmark for simulated branch generation includes: Traverse the node identifiers of the entire purchase order chain, determine the unique initial node corresponding to the order creation time, and concatenate its unique code with the order generation timestamp value to obtain the original string of the dynamic root anchor point; The original string is subjected to a one-way SHA-256 hash operation to obtain a hash value. The high-order bits of the string with a preset number of bits are extracted as the seed factor. The seed factor is XORed bit by bit with the unique link level number corresponding to the initial node to obtain the target random seed uniquely bound to the purchase order. Establish a hash association binding table between the target random seed and the order generation timestamp value and read the corresponding records. Use the timestamp value as the starting coordinate of the time series and the target random seed as the generation control parameter of the procurement forecast simulation branch to construct a shared time series benchmark for the generation of the simulation branch.

[0005] Furthermore, the step of measuring the actual delay of all nodes in the fund flow chain based on the shared timing benchmark, adjusting the tracking timeout threshold of each node based on the actual delay, collecting feedback signals based on the tracking timeout threshold, and determining the integrity of the fund path feedback signal based on the matching degree between the feedback signal and the execution timing of all nodes in the purchase order chain, includes: Based on the time-series starting coordinates of the shared time-series benchmark, the actual delay values ​​of each node in the entire capital flow chain are calculated, and a node delay sequence is formed according to the node order. For each actual delay value, if it exceeds the preset baseline timeout threshold of the corresponding node, the tracking timeout threshold will be increased to the sum of the actual delay value and the preset tolerance increment; if it does not exceed the threshold, the original threshold will remain unchanged. Collect fund path feedback signals with fund node response timestamps within the corresponding tracking window period, compare the response timestamps with the corresponding order node trigger times, and convert the absolute value of the time deviation into a time series matching degree value in the range of 0-1. The system calculates the percentage of nodes whose time-series matching degree reaches the preset matching threshold, and outputs the integrity judgment result of whether the node meets or does not meet the preset integrity threshold.

[0006] Furthermore, when the integrity of the funding path feedback signal is lower than a preset integrity threshold, for the corresponding purchase order, the abnormal distortion simulation branches corresponding to feedback signals lower than the preset integrity threshold are pruned, the bound random seed is reset, and the corrected purchase forecast simulation data is output, including: Extract the integrity judgment result of the fund path feedback signal. If the judgment result is not up to standard, extract the corresponding fund link identifier, reverse the triggering purchase order, and locate all purchase forecast simulation branches bound to the order. Compare the timestamps of simulated branch generation with the timestamps of responses to nodes with insufficient funding. Branches whose absolute time difference is lower than a preset time interval threshold are marked as abnormally distorted simulated branches and pruned to obtain a set of retained simulated branches. Read the target random seed originally bound to the purchase order, perform SHA-256 one-way hash operation based on the concatenation result of the current timestamp and the order number to generate the corresponding reset random seed, reset the random seed bound to the order, replace the original bound target random seed with the reset random seed, and integrate it with the retained simulation branch set to form the corrected purchase prediction simulation data; If the determination result is satisfactory, the current procurement forecast simulation data bound to the order will be directly used as the corrected procurement forecast simulation data.

[0007] Furthermore, the step of extracting historical procurement multi-path simulation results, combining them with the corrected procurement prediction simulation data, eliminating contamination and interference branches, and outputting time-series alignment benchmark calibration results includes: Based on the business category and purchase amount range corresponding to the current purchase order, select historical simulation records from the pre-stored historical purchase multi-path simulation results that belong to the same business category as the current purchase order and whose purchase amount is within a preset range. Extract the timestamps of each simulated branch and the response timestamps of the corresponding funding nodes from the historical simulation records to construct a historical time-series comparison baseline under normal execution conditions of similar businesses; Read the generation timestamp of the retained simulation branch in the corrected procurement forecast simulation data, and calculate the absolute value of the difference between it and the corresponding position timestamp in the historical time series comparison baseline; When the absolute value of the difference exceeds the preset deviation threshold, the corresponding retained branch is marked as a pollution interference branch and excluded to eliminate the pollution interference on the abnormal detection benchmark and obtain the purified procurement prediction simulation data. Align the time-series coordinates of each branch of the purified procurement forecast simulation data with the response time-series of the corresponding nodes in the entire capital flow chain, and output the time-series alignment benchmark calibration result containing the time-series mapping relationship and correction offset record through zero-correction calculation of time-series offset.

[0008] Furthermore, the step of extracting a real-time verification signal based on the time-series alignment reference calibration result and constructing a closed-loop calibration strategy based on the real-time verification signal includes: Extract the time-series mapping relationship and correction offset record from the time-series alignment benchmark calibration result, number them sequentially according to the fund tracking link nodes, and write the offset record into the corresponding node attribute field to obtain a node sequence carrying the correction offset; traverse the actual delay value collected in real time by each node in the node sequence, and use the written correction offset as the delay benchmark reference value to calculate the absolute value of the deviation between the two; if the absolute value of the deviation exceeds the preset unevenness threshold, generate a corresponding real-time verification signal based on the deviation value and the node number to supplement the verification basis for the node delay unevenness scenario; combine the closed-loop calibration rule set containing the node offset update value and the real-time verification signal trigger judgment condition, deploy the signal and rules to each node, and generate a closed-loop calibration strategy adapted to the real-time status of the fund tracking link.

[0009] Furthermore, the step of adjusting the preset AI procurement forecasting model in terms of time series and links according to the closed-loop calibration strategy, defining the risk boundary range of inventory deviation, and obtaining the optimized inventory forecasting result based on the risk boundary range includes: Based on the offset update value and trigger judgment condition of the closed-loop calibration strategy, the set of nodes with uneven delay is extracted, and the deviation values ​​are extracted in the order of nodes to form a link state deviation vector. The vector is used as an additional input feature and input into the long short-term memory network used by the preset AI procurement prediction model. The vector is concatenated with the main input feature at the input layer. The model timing and link adaptation are adjusted through hidden layer operations to obtain the adapted network model. Input the historical inventory consumption data and procurement cycle information of the corresponding purchase order into the network model to obtain the inventory demand forecast value; The extreme values ​​of the deviation values ​​in the vector are extracted to form the risk boundary interval. The predicted value and the interval are encapsulated to obtain the optimized inventory prediction result. The time-series matching degree values ​​of each node in the entire capital flow chain are collected simultaneously.

[0010] Furthermore, the integrated risk score is obtained by fusing the optimized inventory forecast results with the feedback signal. The integrated risk score is then compared with a risk controllable range threshold to output risk assessment data, including: The inventory demand forecast value in the optimized inventory forecast result is read as the forecast center value, and the difference between the upper and lower limits of the risk boundary interval is read as the forecast fluctuation range. Read the time-series matching degree values ​​of each node in the synchronously collected fund path feedback signal; The predicted center value, the predicted floating range, and each of the time-series matching degree values ​​are sequentially concatenated to form a multi-dimensional fusion feature vector; The comprehensive risk score is obtained by multiplying each dimension parameter in the multidimensional fusion feature vector with the corresponding preset weight coefficient and then summing the results. The comprehensive risk score is compared with the risk controllable range threshold. If the comprehensive risk score is lower than or equal to the risk controllable range threshold, the risk of misappropriation of funds is determined to be within the preset controllable range. If the comprehensive risk score is higher than the risk controllable range threshold, the risk of misappropriation of funds is determined to be beyond the preset controllable range. The risk assessment data is output based on the determination result and the comprehensive risk score.

[0011] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses an AI-based multi-dimensional analysis and evaluation method for ERP data. It proposes a complete solution to business scenario problems such as time-series misalignment, uneven node delays, and the risk of fund misappropriation in the entire procurement order and fund flow chain. This invention extracts dynamic root anchor points in the procurement order chain to generate a shared time-series benchmark, measures and adjusts the delay and tracking timeout thresholds of each node, detects the integrity of the fund path by combining feedback signal matching degree, prunes abnormal branches and resets the random seed, thereby calibrating the time-series alignment benchmark. Finally, a closed-loop calibration strategy is formed to optimize the AI ​​prediction model, quantifying the inventory deviation risk boundary and the fund misappropriation risk score. The core innovation of this invention lies in constructing a dynamic adaptation mechanism for procurement forecasting and fund monitoring through multi-dimensional data fusion and real-time verification. This ensures real-time response and risk controllability of the chain status, significantly improving the decision-making accuracy and risk management capabilities of the ERP system in complex business scenarios. Attached Figure Description

[0012] Figure 1This is a flowchart of an AI-based multi-dimensional analysis and evaluation method for ERP data according to the present invention.

[0013] Figure 2 This is a schematic diagram of an AI-based multi-dimensional analysis and evaluation method for ERP data according to the present invention.

[0014] Figure 3 This is another schematic diagram of an AI-based multi-dimensional analysis and evaluation method for ERP data according to the present invention. Detailed Implementation

[0015] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0016] like Figures 1-3 This embodiment of an AI-based multi-dimensional analysis and evaluation method for ERP data may specifically include: S101. By extracting a unique dynamic root anchor point from the entire purchase order chain, the hash association between the corresponding random seed and the timestamp is obtained, and a shared time series benchmark for simulated branch generation is constructed.

[0017] The process iterates through the node identifiers in the entire purchase order chain to locate the initial node corresponding to the order creation time. The unique code of the initial node and the order generation timestamp value are extracted. The unique code and timestamp value are concatenated to form the original string of the dynamic root anchor. A one-way mapping operation is performed on the original string using the SHA-256 hash function to obtain a fixed-length dynamic root anchor hash value. Based on the dynamic root anchor hash value, a numerical fragment of a preset number of bits is extracted from its high-order bits as a seed factor. The seed factor is then XORed bit-by-bit with the link level number of the purchase order to obtain the target random seed bound to the purchase order. The link level number is defined as the depth level of the order in the entire chain, obtained by traversing the link nodes and counting from the initial node. The input is a list of all link nodes, the algorithm is a depth-first traversal counting, and the output is an integer. For example, if the link depth is 3 levels, the number is 3. A hash association binding table is established between the target random seed and the timestamp value. Based on the hash association binding table, the binding record of the target random seed and the timestamp value is read. The timestamp value is used as the time series starting point coordinate, and the target random seed is used as the generation control parameter of the simulated branch. A shared time series benchmark containing the association structure of the time series starting point coordinate and the generation control parameter is constructed. The shared time series benchmark provides a unified time reference and branch generation control basis for the simulated branch generation process.

[0018] The specific inputs to the procurement forecast simulation branch include a shared time series baseline, current procurement order basic data, and historical procurement baseline data from the ERP system. The generation of the simulation branch is based on the unified scheduling of the time series starting point of the shared time series baseline. Using the target random seed as the random disturbance control factor, multiple parallel procurement forecast simulation paths are generated by traversing the hierarchy of fund flow nodes. Each path corresponds to an independent simulation branch. All simulation branches under the same procurement order maintain time series alignment and are independent of each other. The specific output of the simulation branch is structured data, which includes four core contents: a unique branch identifier, a generation timestamp, time series coordinates of each node, and a simulated fund flow path sequence.

[0019] Throughout the entire procurement order chain, each node carries unique node identification information, including a node code and the node creation time. For the initial node corresponding to the order creation time, a unique code is extracted from the initial node as an anchor identifier, and the order generation timestamp value is obtained as a time-series positioning basis.

[0020] Specifically, in the formation of the original string of the dynamic root anchor, the 16-bit alphanumeric unique code of the initial node is placed at the beginning of the string, and the order generation timestamp value is placed at the end of the string. The two are connected by an underscore "_" as a preset separator. The order generation timestamp value is a 13-bit numerical sequence accurate to the millisecond level.

[0021] For example, if the unique code of the initial node is a 16-digit alphanumeric sequence and the order generation timestamp is a numerical sequence accurate to the millisecond level, then the two are concatenated in a fixed order to form the original string of the dynamic root anchor. This original string of the dynamic root anchor is processed by the SHA-256 hash function. This hash function accepts an input string of arbitrary length and outputs a hash digest of a fixed length of 256 bits through multiple rounds of compression operations. This hash digest is the hash value of the dynamic root anchor. The SHA-256 hash function has a one-way characteristic: the same input produces the same output, and different inputs produce significantly different output results.

[0022] In one embodiment, the process of generating a target random seed based on the dynamic root anchor hash value is as follows: The dynamic root anchor hash value is represented as a 256-bit binary string. Starting from its highest bit position (the leftmost bit), the first 32 bits are extracted as the seed factor S. The seed factor S and the link level number of the purchase order are converted to 32-bit binary format, with the highest bits padded with zeros to ensure complete consistency. Then, a bitwise XOR operation is performed, where identical bits are zeroed and dissimilar bits are oneed. The result is the target random seed bound to the purchase order.

[0023] It should be noted that the hash-associative binding table uses a key-value pair structure to store the binding relationship between the target random seed and the timestamp value, where the target random seed serves as the index key and the timestamp value as the corresponding value. Each record corresponds to the time-series anchoring information of a purchase order. Based on the hash-associative binding table, the binding record of the target random seed and the timestamp value is read. The timestamp value is used to determine the time-series starting coordinates, and the target random seed is used to determine the generation control parameters of the simulation branch. Together, they constitute the association structure of a shared time-series benchmark. This shared time-series benchmark provides a unified time reference and branch generation control basis for the simulation branch generation process.

[0024] S102. Measure the actual delay of each node in the entire fund transfer chain according to the shared timing benchmark, adjust the tracking timeout threshold of each node according to the actual delay, collect the feedback signal of each node in the entire fund transfer chain based on the tracking timeout threshold, compare the matching degree of the feedback signal with the node execution timing of the entire purchase order chain, and determine the integrity of the fund path feedback signal.

[0025] Based on the time-series starting point coordinates in the shared time-series benchmark, each fund node in the entire fund transfer chain is traversed. The actual arrival time of the fund transfer instruction received by each fund node is recorded. The difference between the actual arrival time and the time-series starting point coordinates is calculated to obtain the actual delay value of each fund node. According to the order of the fund nodes in the entire fund transfer chain, the actual delay values ​​are organized into a node delay sequence. For each actual delay value in the node delay sequence, the preset benchmark timeout threshold of that fund node is read. If the actual delay value exceeds the preset benchmark timeout threshold, the tracking timeout threshold of that fund node is increased to the sum of the actual delay value and the preset tolerance increment. If the actual delay value does not exceed the preset benchmark timeout threshold, the tracking timeout threshold of that fund node remains unchanged, thus obtaining the adjusted tracking timeout threshold for each fund node. Based on the tracking timeout threshold, fund flow feedback signals containing the response timestamps of fund nodes are collected within the tracking window period of each fund node. Simultaneously, the execution time sequence records of the corresponding order nodes in the entire purchase order chain are obtained. The response timestamps are compared one by one with the node trigger times in the execution time sequence records, and the absolute value of the time deviation between the two is calculated to obtain the time sequence matching degree of each node. Based on the time sequence matching degree, the ratio of the number of fund nodes with a time sequence matching degree reaching a preset matching threshold to the total number of nodes in the entire fund flow chain is calculated. If the ratio reaches a preset integrity threshold, the integrity of the fund path feedback signal is determined to be satisfactory; if the ratio is lower than the preset integrity threshold, the integrity of the fund path feedback signal is determined to be unsatisfactory, thus obtaining the integrity determination result of the fund path feedback signal.

[0026] In the process of delay measurement across the entire capital flow chain, the shared time series benchmark provides a unified time reference starting point. The time series starting point coordinates in the shared time series benchmark are derived from the timestamp value of the purchase order creation time. This timestamp value is then bound to the target random seed after hash association mapping, thereby establishing the time series correlation basis between the purchase order and the capital flow.

[0027] Specifically, the actual delay value is measured as follows: when the fund transfer instruction is triggered from the purchase order, it passes through each fund node in the entire fund transfer chain in sequence. Each fund node records the current time as the actual arrival time when it receives the fund transfer instruction.

[0028] For example, in a fund transfer chain including a payment application node, a financial approval node, a bank interface node, and a receipt confirmation node, the payment application node records its arrival time after receiving the payment trigger signal of the purchase order, the financial approval node records its processing completion time after completing the approval action, and so on. The difference between the actual arrival time of each fund node and the starting coordinate of the time sequence is calculated, i.e., the actual arrival time is subtracted from the starting coordinate of the time sequence. Legal holidays and non-working hours are excluded during the difference calculation. Specifically, the algorithm takes a start time T1 and an end time T2 as input, calls the national legal holiday API to query the work calendar table, iterates through each day from T1 to T2, and only counts the effective working hours H from 8:00 to 18:00 if it is a working day, excluding weekends and holidays. Finally, H is accumulated as the effective working hours, i.e., the actual delay value. For example, if T1 is 2023-10-01-09:00 and T2 is 2023-10-03-17:00, after excluding the National Day holiday, H=18. This delay value reflects the cumulative effective working time that a fund transfer instruction takes from order creation to reaching this node.

[0029] It should be noted that the node delay sequence is organized according to the physical transmission order of the funding nodes in the entire funding flow chain. Each element in the sequence corresponds to the actual delay value of a funding node, and the order of the sequence is consistent with the actual path of the funding flow.

[0030] In one possible implementation, the dynamic adjustment of the tracking timeout threshold employs the following mechanism: Each funding node is pre-configured with a baseline timeout threshold, which is determined based on the node's average processing time in historical fund flow data. When the actual delay value of a funding node in the node delay sequence exceeds its baseline timeout threshold, it indicates that the node's current fund flow speed is lower than the historical average. In this case, the tracking timeout threshold for that funding node is adjusted to the sum of the actual delay value and the preset tolerance increment.

[0031] For example, the preset baseline timeout threshold for the bank interface node is 1800 seconds, and the preset baseline timeout threshold for the payment application / financial approval node is 300 seconds. These preset baseline timeout thresholds are adjusted by 20% based on the historical average processing time of the corresponding node. If the actual delay value of the bank interface node reaches 2100 seconds, the tracking timeout threshold is adjusted to 2100 seconds plus a preset tolerance increment. The preset tolerance increment defaults to 300 seconds, with a range of 30 seconds to 600 seconds, providing a buffer time for the node's timeout threshold. If the actual delay value does not exceed the baseline timeout threshold, the tracking timeout threshold for that funding node remains unchanged, indicating that the node's fund transfer is within the normal speed range. The tracking window period refers to the time interval from when the funding node begins processing the fund transfer instruction to when the tracking timeout threshold expires. During this tracking window period, the fund transfer feedback signal returned by the funding node is continuously monitored and collected. This feedback signal includes a response timestamp, which records the specific time when the funding node completes processing and returns a response.

[0032] In one embodiment, the calculation process for the timing matching degree is as follows: The execution timing record of the order node corresponding to the current funding node in the entire procurement order chain is obtained. The execution timing record includes the trigger time of the order node, which represents the time point when the procurement order is activated and executed at that stage. The response timestamp in the fund flow feedback signal is compared with the node trigger time in the execution timing record, and the absolute value of the time deviation between the two is calculated. The smaller the absolute value of the time deviation, the closer the timing of the fund flow and the procurement order at that node, and the higher the degree of synchronization. The absolute value of the time deviation is compared with a preset upper limit value of deviation. The preset upper limit value of deviation is set to 600 seconds by default, with a range of 300 seconds to 1800 seconds. A preset linear mapping rule is used to calculate the timing matching degree value in the interval of 0 to 1 using the formula: TMD=max(0,1-(ATD / PDUL)), where TMD represents the timing matching degree, max represents the maximum value function, ATD represents the absolute value of the time deviation (in seconds), and PDUL represents the preset upper limit value of deviation (in seconds). For example, if the input ATD is 200 seconds and PDUL is 600 seconds, then TMD = max(0, 1 - (200 / 600)) = max(0, 0.6667) = 0.6667. The matching degree is 1 when the deviation is zero, 0 when the deviation reaches the preset upper limit, and fixed at 0 when the deviation exceeds the preset upper limit.

[0033] Understandably, after calculating the time-series matching degree for all funding nodes, the entire funding flow chain is traversed, and the number of funding nodes whose time-series matching degree reaches a preset matching threshold is counted. The specific process for integrity determination involves counting the number of funding nodes whose time-series matching degree reaches the preset matching threshold. The preset matching threshold defaults to 0.75, with a range of 0.6-0.9. The number of nodes meeting the threshold is divided by the total number of nodes in the entire funding flow chain to obtain the matching compliance ratio. This matching compliance ratio is compared with the preset integrity threshold, which defaults to 0.85, with a range of 0.7-0.95. If the matching compliance ratio reaches or exceeds the preset integrity threshold, the integrity of the funding path feedback signal is deemed to be compliant, indicating that the feedback signals of most nodes in the entire funding flow chain are synchronized with the procurement order execution timeline. If the matching compliance ratio is lower than the preset integrity threshold, the integrity of the funding path feedback signal is deemed to be non-compliant, indicating that the feedback signals of many nodes deviate from the procurement order execution timeline. The integrity determination result serves as a quantitative assessment output of the quality of the fund path feedback signal, and this result identifies the temporal coordination status between the current fund flow chain and the purchase order chain.

[0034] S103. If the integrity is lower than a preset integrity threshold, for the purchase order corresponding to the fund flow link to which the fund path feedback signal belongs, the abnormal distortion simulation branch corresponding to the fund path feedback signal that is lower than the preset integrity threshold is removed, the random seed bound to the order is reset, and the target random seed originally bound to the purchase order is replaced with the reset random seed to obtain the corrected purchase prediction simulation output data; if the integrity reaches the preset integrity threshold, the purchase prediction simulation data currently bound to the purchase order is directly used as the corrected purchase prediction simulation output data.

[0035] If the integrity determination result of the funding path feedback signal is unsatisfactory, the corresponding funding path identifier is extracted from the entire funding flow chain. This identifier is then used to reverse-link the funding flow to the purchase order that triggered the funding flow, obtaining the order number of the purchase order and locating all procurement forecast simulation branches bound to that purchase order. For each procurement forecast simulation branch, the generation timestamp of each simulation branch and the response timestamp of the funding node whose integrity is unsatisfactory are read. If the time difference between the generation timestamp of a simulation branch and the response timestamp of a funding node is lower than a preset time interval threshold, this simulation branch is marked as an abnormally distorted simulation branch. Abnormally distorted simulation branches are then removed from the procurement forecast simulation branches, resulting in a set of retained simulation branches. Based on the retained simulation branch set, the target random seed originally bound to the purchase order is read, and the timestamp accurate to the millisecond level for the current correction operation is obtained. The order number is placed at the beginning of the string, and the current timestamp is placed at the end. The strings are concatenated in a fixed order using underscores "_" as separators to form a new seed original string. The seed original string is subjected to a SHA-256 one-way hash operation to obtain a reset random seed. The random seed bound to the order is reset, and the reset random seed is written into the random seed binding record of the purchase order, replacing the original target random seed. Based on the reset random seed and the retained simulation branch set, the corrected purchase prediction simulation output data is output.

[0036] When the integrity judgment result of the funding path feedback signal is shown as unqualified, it indicates that there are many nodes in the entire funding flow chain whose feedback signals deviate from the execution sequence of the purchase order. At this time, the correction process of the procurement forecast simulation branch is triggered.

[0037] Specifically, a pre-established mapping relationship exists between the fund transfer link identifier and the purchase order. Each fund transfer link is assigned a unique fund transfer link identifier upon creation, and this identifier is bound one-to-one with the purchase order that triggers the fund transfer. By reading the fund transfer link identifier carried by a fund transfer link that fails to meet the integrity standard, and then querying the mapping relationship table in reverse, the corresponding purchase order and its order number can be located.

[0038] In one possible implementation, the determination of abnormal distortion simulation branches adopts a time difference comparison method. For each procurement forecast simulation branch bound to the procurement order, the generation timestamp recorded when the simulation branch was created is read, and the response timestamps returned by funding nodes that have not met the integrity standards are obtained. The absolute value of the time difference between the simulation branch generation timestamp and the funding node response timestamp is calculated, and this absolute value of the time difference is compared with a preset time interval threshold.

[0039] For example, if the generation timestamp of a simulated branch is the fifth minute after the purchase order is created, while the response timestamp of the corresponding funding node shows that the node experienced a delay anomaly in the sixth minute, and the absolute value of the time difference between the two is lower than a preset time interval threshold under the premise of only calculating the effective working time, and the generation timestamp of the simulated branch is earlier than the abnormal response timestamp of the funding node, then it is determined that the simulated branch is affected by the abnormality of the funding node and is marked as an abnormally distorted simulated branch; if it is only triggered once due to system network fluctuations and the final timing matching degree of the node meets the standard, it is exempted from being marked as an abnormally distorted simulated branch; the preset time interval threshold is set to 60s by default, and the value range is 30s-300s, which is adjusted according to the business execution cycle of the purchase order.

[0040] It should be noted that the branch pruning operation removes all simulated branches marked as abnormally distorted from the procurement forecast simulation branch set, and the remaining unmarked simulated branches constitute the retained simulation branch set. Furthermore, the random seed reset process maintains consistency with the aforementioned dynamic root anchor point formation method. The target random seed originally bound to the procurement order is read, the timestamp of the current correction operation is obtained, and this timestamp is concatenated with the order number of the procurement order in a fixed order to form a new original seed string. A SHA-256 one-way hash operation is performed on the original seed string to obtain the reset random seed. The random seed bound to this order is reset, and the reset random seed is written into the random seed binding record of the procurement order, replacing the original target random seed. Based on the reset random seed and the retained simulation branch set, the corrected procurement forecast simulation output data is output. The corrected procurement forecast simulation output data excludes distorted branches affected by abnormal funding nodes and carries new random seed binding information.

[0041] S104. Based on the pre-stored historical procurement multi-path simulation results, combined with the corrected procurement forecast simulation output data, the pollution interference of the anomaly detection benchmark is eliminated, and the time-series alignment benchmark calibration results of procurement forecast simulation and full-link monitoring of fund flow are obtained.

[0042] From the pre-stored historical procurement multi-path simulation results, the data is filtered according to the business category and procurement amount range of the procurement order. Historical simulation records belonging to the same business category and with similar procurement amount range as the current procurement order are extracted. The generation timestamps of each simulation branch and the response timestamps of the corresponding funding nodes in the historical simulation records are read to construct a historical time-series comparison baseline. Based on the historical time-series comparison baseline, the generation timestamps of the retained simulation branches in the corrected procurement forecast simulation output data are read. The absolute value of the difference between the generation timestamp of the retained simulation branch and the corresponding timestamp in the historical time-series comparison baseline is calculated. If the absolute value of the difference exceeds a preset deviation threshold, the retained simulation branch is marked as a contaminated interference branch, and the contaminated interference branch is excluded from the corrected procurement forecast simulation output data to obtain purified procurement forecast simulation data. The time-series coordinates of each branch in the purified procurement forecast simulation data are aligned one by one with the response timestamps of the corresponding funding nodes in the full-link monitoring of fund flow. The time-series offset between the two is calculated and zeroed out, and the time-series alignment benchmark calibration result of procurement forecast simulation and full-link monitoring of fund flow is output.

[0043] Historical procurement multi-path simulation results refer to the multiple simulated branch data and their corresponding cash flow response data recorded for different purchase orders during the execution of past procurement transactions. These historical simulation results are stored and categorized according to the business type of the purchase order, including different types such as raw material procurement, equipment procurement, and service procurement.

[0044] Specifically, the historical time-series baseline is constructed as follows: Based on the business category identifier of the current purchase order, a set of historical records belonging to the same business category is retrieved from the pre-stored historical simulation results. Within this historical record set, a second filtering is performed according to the purchase amount range, extracting historical simulation records whose amounts are close to those of the current purchase order.

[0045] For example, if the current purchase order is for raw materials and the amount is in a medium range, then according to the preset business category classification criteria, including three major categories: raw material purchase, equipment purchase, and service purchase, simulated records from the past that also belong to raw material purchases and whose purchase amounts are within a preset similar range fluctuating around the current order amount by 20%. From the selected historical simulated records, the generation timestamp sequence of each simulated branch and the response timestamp sequence of the corresponding funding node are read, and the two are combined in node order to form a historical time-series comparison baseline under the normal execution state of the same type of business.

[0046] It should be noted that the historical time series baseline reflects the time series distribution pattern of similar procurement operations under normal execution conditions, providing a reference for judging the deviation of the current data.

[0047] In one possible implementation, the identification of contamination branches employs a time difference comparison method. The generation timestamps of retained simulation branches are read from the corrected procurement forecast simulation output data. These timestamps are then compared one by one with the corresponding timestamps in the historical time-series baseline, and the absolute value of the difference is calculated. If the absolute value of the difference exceeds a preset deviation threshold (default value 120s, range 60s-600s), it indicates that the time-series characteristics of the retained simulation branch deviate from the historical normal range, and it is marked as a contamination branch. This contamination branch may originate from abnormal delays in the preceding fund transfer process or time-series drift during data acquisition. Further, all marked contamination branches are excluded from the corrected procurement forecast simulation output data to obtain purified procurement forecast simulation data. This purified data retains only simulation branches whose time-series characteristics are within the normal range. Based on the purified procurement forecast simulation data, the time-series coordinates of each branch are aligned one by one with the response time of the corresponding fund nodes in the full-link monitoring of fund transfer. The offset between each pair of time-series coordinates is calculated, and a linear translation method is used for zeroing correction. The correction method is that the corrected time-series coordinate = the original time-series coordinate - the corresponding offset, so that the time-series coordinates of the simulated branch are completely aligned with the reference starting point of the response time of the funding node. The time-series alignment reference calibration result of the procurement forecast simulation and the full-link monitoring of the fund flow is output. The calibration result includes the aligned time-series mapping relationship and the correction offset record.

[0048] S105. The timing alignment benchmark calibration result is sent back to the fund tracking link to supplement the real-time verification signal under the scenario of uneven node delay, supplement the verification basis under the scenario of uneven node delay, and generate a closed-loop calibration strategy adapted to the real-time status of the fund tracking link.

[0049] Based on the timing mapping relationship and correction offset records in the timing alignment benchmark calibration results, the correction offset records are written one by one into the node attribute field of the corresponding fund node according to the sequential numbering of each fund node in the fund tracking link. This completes the back-transmission of the timing alignment benchmark calibration results to the fund tracking link, resulting in a fund tracking link node sequence carrying the correction offset. For the fund tracking link node sequence, the current actual delay value of each fund node is traversed, and the correction offset is read as the delay benchmark reference value. The absolute value of the deviation between the actual delay value and the delay benchmark reference value is calculated. If the absolute value of the deviation exceeds a preset unevenness threshold, a real-time verification signal containing the node number and the current deviation value is generated for that fund node to supplement the verification basis in the scenario of uneven node delay. Based on the real-time verification signal and the correction offset, a closed-loop calibration rule set is constructed. The closed-loop calibration rule set includes the offset update value of each fund node and the trigger judgment condition of the real-time verification signal. The closed-loop calibration rule set is deployed to each fund node in the fund tracking link to form a closed-loop calibration strategy adapted to the real-time state of the fund tracking link.

[0050] The time-series alignment benchmark calibration results embody the time-series mapping relationship between procurement forecasting simulation and end-to-end monitoring of fund flows. The correction offset records in the calibration results reflect the adjustment range required for each fund node during the time-series alignment process. Feeding these calibration results back to the fund tracking link is a fundamental step in establishing a closed-loop calibration mechanism.

[0051] Specifically, the process of writing the correction offset back is executed sequentially according to the order number of each fund node in the fund tracking chain. For each fund node, the offset value corresponding to that node number is read from the correction offset record, and the offset value is written into the node attribute field of that fund node. The node attribute field is a reserved data storage location for the fund node, used to record calibration information related to that node.

[0052] In one possible implementation, the determination of uneven node latency is achieved through a deviation comparison method. The sequence of fund tracking link nodes carrying correction offsets is traversed. For each fund node, its current actual latency value is obtained. This actual latency value originates from the difference between the response timestamp and the starting coordinate of the time sequence in the fund flow feedback signal. The actual latency value is compared with the correction offset written to that node, which serves as a latency benchmark. The absolute value of the deviation between the two is calculated. If the absolute value of the deviation exceeds a preset unevenness threshold (the default value is 90s, with a range of 30s to 300s), it indicates that the current latency state of the fund node deviates from the calibrated benchmark level, indicating uneven latency.

[0053] It should be noted that for funding nodes exhibiting uneven latency, a real-time verification signal is generated as the verification basis. This real-time verification signal includes the node number of the funding node and its current deviation value, which is the absolute value of the deviation between the actual latency value and the latency benchmark reference value. Further, based on the real-time verification signal and the correction offset, a closed-loop calibration rule set is constructed. In this set, for each funding node, the offset update value and the triggering condition for the real-time verification signal are recorded. The triggering condition is the logic for determining whether the absolute value of the deviation exceeds a preset unevenness threshold. The closed-loop calibration rule set is deployed to each funding node in the funding tracking chain. Each node executes its own calibration judgment and signal triggering according to the rule set, forming a closed-loop calibration strategy adapted to the real-time state of the funding tracking chain.

[0054] S106. The preset AI procurement forecasting model is adjusted in terms of timing and link through a closed-loop calibration strategy to define the risk boundary range of inventory deviation, and the optimized inventory forecasting result is obtained based on the risk boundary range.

[0055] Based on the offset update values ​​and trigger conditions of each funding node in the closed-loop calibration rule set, a set of funding nodes in the current funding tracking link that are in a state of uneven delay is extracted. The current deviation value of each node in the funding node set is read, and the current deviation values ​​are organized according to the node order to form a link state deviation vector. The link state deviation vector is used as an additional input feature and input to a preset long short-term memory network. The long short-term memory network takes historical procurement data sequences as the main input and inventory demand as the output. The link state deviation vector and the current input of the long short-term memory network are concatenated and then fed into the network hidden layer. The hidden layer output is recalculated based on the concatenated input to obtain a long short-term memory network adapted to the real-time state of the link. Through the long short-term memory network adapted to the real-time state of the link, the historical inventory consumption data and procurement cycle information of the current procurement order are input to obtain the inventory demand forecast value. At the same time, the maximum value of the deviation value of each node in the link state deviation vector is extracted as the upper limit of the risk boundary and the minimum value is extracted as the lower limit of the risk boundary. The upper limit of the risk boundary and the lower limit of the risk boundary constitute the risk boundary interval of inventory deviation. The specific calculation follows the following formula. In the formula, This represents the inventory deviation risk boundary range. This is the central value for inventory demand forecast. This is the link state deviation vector. To adapt and correct the preset link state, This represents the maximum absolute deviation value of each element in the deviation vector. The inventory demand forecast and the risk boundary interval are combined and encapsulated, with the inventory demand forecast serving as the center value of the inventory forecast and the risk boundary interval serving as the fluctuation range of the inventory forecast. The optimized inventory forecast result, adapted to the real-time status of the link, is then output.

[0056] The closed-loop calibration rule set records the offset update values ​​and trigger conditions for each funding node. Based on the trigger conditions, funding nodes currently in a state of uneven latency can be identified. The current deviation value is extracted from these nodes, reflecting the degree of difference between the actual latency value and the latency benchmark reference value.

[0057] Specifically, the link state deviation vector is organized according to the physical order of each fund node in the fund tracking link.

[0058] For example, if the fund tracking link includes four nodes: payment application node, financial approval node, bank interface node, and receipt confirmation node, then the link status deviation vector is a four-dimensional vector. Each component in the vector corresponds to the current deviation value of a node, and the order of the components is consistent with the order of the nodes in the link.

[0059] In one embodiment, the pre-defined AI procurement prediction model uses a Long Short-Term Memory (LSTM) network with a pre-set basic structure and training rules. The network is a 2-layer LSTM + 1-layer fully connected output layer structure with 64 hidden layer neurons. The input gate, forget gate, and output gate all use the Sigmoid activation function, and the cell state uses the Tanh activation function. The input sequence length is fixed at 12 historical procurement cycles, and the main input features are the inventory consumption, procurement amount, and supplier delivery cycle time series data of the historical procurement cycles. The model uses the Adam optimizer, the loss function is the mean squared error, the training batch size is 32, the number of iterations is 100, and the convergence condition is that the validation set loss decreases to below 1e-4. The pre-training is completed using the enterprise's historical procurement data from the past 12 months.

[0060] In one possible implementation, a Long Short-Term Memory (LSTM) network is a recurrent neural network structure with temporal memory capabilities. The LSM network includes three gating units: an input gate, a forget gate, and an output gate, as well as a cell state storage unit. The input gate controls the proportion of the current input information written into the cell state, the forget gate controls the proportion of the cell state from the previous time step retained in the current time step, and the output gate controls the proportion of the cell state output to the hidden layer. In a procurement forecasting scenario, the LSM network uses historical procurement data sequences as its main input. These sequences contain time-series data such as inventory consumption, procurement amount, and supplier delivery cycles from multiple past procurement cycles. The network output is a predicted inventory demand value for the next procurement cycle.

[0061] It should be noted that the concatenation of the link state deviation vector with the input of the Long Short-Term Memory network is performed at each time step. When the network processes the input at the current time step, the original input vector and the link state deviation vector are concatenated end-to-end along the feature dimension to form the expanded input vector.

[0062] In one embodiment, the network adaptation adjustment process after vector concatenation is as follows: The link state deviation vector and the main input feature of the Long Short-Term Memory (LSTM) network at the current moment are concatenated at the input layer. The concatenated feature is processed by the network's hidden layer, and after gating operations through the input gate, forget gate, and output gate, the cell state is updated. The hidden layer output at the current moment is calculated, completing the model's temporal and link adaptation adjustment to obtain the adapted LSTM network model. Further, the historical inventory consumption data and procurement cycle information of the current purchase order are input into the LSTM network that adapts the real-time state of the link. After forward propagation calculation, the network outputs an inventory demand prediction value, which represents the estimated inventory demand for the next procurement cycle.

[0063] In one possible implementation, the risk boundary interval is determined based on the extreme values ​​of each component in the link state deviation vector. All components in the link state deviation vector are traversed, and the component with the largest value is extracted as the upper risk boundary, and the component with the smallest value is extracted as the lower risk boundary. The upper and lower risk boundaries together constitute the risk boundary interval for inventory deviation. The upper risk boundary reflects the node state with the most severe delay deviation in the fund tracking link, and the lower risk boundary reflects the node state with the slightest delay deviation. The upper and lower risk boundaries together constitute the risk boundary interval for inventory deviation, which characterizes the possible fluctuation range of inventory forecasting under the current link state.

[0064] It is understandable that the combination and encapsulation of the inventory demand forecast and the risk boundary range form a complete forecast output structure. The inventory demand forecast serves as the central value of the inventory forecast, and the risk boundary range serves as the fluctuation range of the inventory forecast; together, they constitute the optimized inventory forecast result. While encapsulating the optimized inventory forecast result, the time-series matching degree values ​​corresponding to each node in the entire capital flow chain are simultaneously collected and stored in association with the optimized inventory forecast result, serving as the core input data for subsequent risk scoring. The time-series matching degree value is calculated through the following process, with the input being the actual time-series data T of each node. a And expected time series data T e Using the formula M=1-(|T a -T e | / T e The output range is 0 to 1, where 0 represents a complete mismatch and 1 represents a perfect match. For example, the T value for node A is... a =5 days, T e =4 days, M=0.75; this value is subsequently used in the risk scoring to calculate the total score with a weight of 0.3. The optimized inventory forecast result includes the correspondence between the inventory demand forecast value and the risk boundary interval. This result provides both a point estimate of inventory demand and an interval estimate range considering the real-time status of the link, giving inventory planning decisions a quantitative basis for risk.

[0065] S107. The optimized inventory forecast result and the feedback signal are combined to obtain a comprehensive risk score. The comprehensive risk score is compared with the risk controllable range threshold, and the risk assessment data is output.

[0066] Based on the inventory demand forecast and risk boundary range in the optimized inventory forecast results, the inventory demand forecast is extracted as the forecast center value, and the upper and lower limits of the risk boundary range are extracted as the forecast fluctuation range. Simultaneously, the time-series matching degree values ​​of each funding node are read from the funding path feedback signal. The forecast center value, forecast fluctuation range, and time-series matching degree values ​​of each funding node are concatenated into a vector according to a preset feature dimension order to obtain a multi-dimensional fused feature vector. The forecast fluctuation range in the multi-dimensional fused feature vector is multiplied by a preset fluctuation range weight coefficient, and the time-series matching degree values ​​of each funding node are multiplied by preset matching degree weight coefficients and then summed. The sum of these products yields a comprehensive risk score. The comprehensive risk score is compared with a risk controllable range threshold, which is set to 70 by default and ranges from 50 to 100. A higher value indicates a more lenient risk control level. If the comprehensive risk score is lower than or equal to the risk controllable range threshold, the risk of misappropriation of funds is determined to be within the preset controllable range. If the comprehensive risk score is higher than the risk controllable range threshold, the risk of misappropriation of funds is determined to be beyond the preset controllable range. The risk assessment data is output based on the determination result and the comprehensive risk score.

[0067] Optimizing inventory forecasting results and cash flow feedback signals respectively carries business information from the dimensions of inventory forecasting and cash flow. Integrating these two data across multiple dimensions is a crucial step in comprehensively assessing procurement risks. This integration process transforms business data from different sources into a unified feature representation.

[0068] Specifically, the construction process of the multi-dimensional fusion feature vector is as follows: The inventory demand forecast value is extracted from the optimized inventory forecast results as the forecast center value, which reflects the estimated inventory demand for the next procurement cycle. Simultaneously, the upper and lower limits of the risk boundary interval are extracted, and the difference between the upper and lower limits is calculated as the forecast fluctuation range, which reflects the degree of uncertainty in inventory forecasting. The time-series matching degree values ​​of each funding node are read from the funding path feedback signals synchronously collected in step S106. These time-series matching degree values ​​originate from the real-time collection results of the corresponding funding nodes. The forecast center value, forecast fluctuation range, and time-series matching degree values ​​of each funding node are sequentially arranged and concatenated to form the multi-dimensional fusion feature vector.

[0069] It should be noted that the arrangement order of features in the multi-dimensional fusion feature vector follows the preset feature dimension order, which corresponds one-to-one with the weight coefficients in the subsequent weighted summation calculation.

[0070] In one possible implementation, the process of calculating the comprehensive risk score using the weighted summation method involves first normalizing all input features to the 0-1 range, where the prediction fluctuation range uses maximum-minimum normalization, and the time series matching degree value ranges from 0 to 1 and has been normalized; then reading the preset weight coefficients, which are fixed values ​​pre-configured based on business experience, with the sum of the weight coefficients being 1, and the default setting being a prediction fluctuation range weight coefficient of 0.4 and a single node time series matching degree weight coefficient of 0.6 / total number of funding nodes. The comprehensive risk score is calculated using the following formula: CRS = (NFR × WFR) + Σ[(1 - NTMD) × WTMD], where CRS is the comprehensive risk score, NFR is the normalized predicted fluctuation range, WFR is the fluctuation range weight coefficient, NTMD is the normalized single-node time series matching degree value, and WTMD is the single-node matching degree weight coefficient. The predicted fluctuation range is multiplied by the fluctuation range weight coefficient to obtain the fluctuation range weighted value. The (1 - time series matching degree value) of each funding node is multiplied by the corresponding single-node matching degree weight coefficient and then summed to obtain the matching degree weighted sum. The sum of the two is the final comprehensive risk score. The higher the score, the higher the overall risk level. For example, if the input prediction fluctuation range is 50, the normalized value is 0.5, the time series matching degree is 0.8 for node 1 and 0.9 for node 2, the total number of funding nodes is 2, and the weights WFR=0.4 and WTMD=0.3 respectively, then CRS=(0.5×0.4)+((1-0.8)×0.3+(1-0.9)×0.3)=0.2+0.06+0.03=0.29.

[0071] Understandably, a higher comprehensive risk score indicates greater uncertainty in inventory forecasting for the current purchase order or a more severe deviation in the timing of cash flow, resulting in a higher overall risk level. Further, the comprehensive risk score is compared to a risk controllable range threshold. If the comprehensive risk score is lower than or equal to the risk controllable range threshold, the risk of misappropriation of funds is considered controllable; if the comprehensive risk score is higher than the risk controllable range threshold, the risk of misappropriation of funds exceeds the controllable range. The judgment result and the comprehensive risk score are combined and packaged to form a final risk assessment report, which includes the basis for risk level determination and quantitative scoring.

[0072] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and additions without departing from the principle of the present invention, and these improvements and additions should also be considered within the scope of protection of the present invention.

Claims

1. A multi-dimensional analysis and evaluation method for ERP data based on AI, characterized in that, The method includes: Extract the unique dynamic root anchor point of the entire purchase order chain, obtain the hash association between the random seed and the timestamp, and construct a shared time series benchmark generated by the simulated branch; The actual delay of the entire link node of the fund flow is measured according to the shared time series benchmark. The tracking timeout threshold of each node is adjusted according to the actual delay. Feedback signals are collected based on the tracking timeout threshold. The integrity of the fund path feedback signal is determined according to the matching degree between the feedback signal and the execution time series of the entire link node of the purchase order. When the integrity of the funding path feedback signal is lower than a preset integrity threshold, for the corresponding purchase order, the abnormal distortion simulation branch corresponding to the feedback signal that is lower than the preset integrity threshold is removed, the bound random seed is reset, and the corrected purchase prediction simulation data is output. Extract historical procurement multi-path simulation results, combine them with the corrected procurement prediction simulation data, eliminate contamination and interference branches, and output the time-series alignment benchmark calibration results. Based on the time-series alignment benchmark calibration results, extract the real-time verification signal and construct a closed-loop calibration strategy based on the real-time verification signal. The preset AI procurement forecasting model is adjusted in terms of timing and link according to the closed-loop calibration strategy, the risk boundary range of inventory deviation is defined, and the optimized inventory forecasting result is obtained based on the risk boundary range. The optimized inventory forecast results are combined with the feedback signal to obtain a comprehensive risk score. The comprehensive risk score is compared with the risk controllable range threshold, and risk assessment data is output.

2. The method according to claim 1, characterized in that, The process of extracting the unique dynamic root anchor point of the entire purchase order chain, obtaining the hash association between the random seed and the timestamp, and constructing a shared time-series benchmark for simulating branch generation includes: Traverse the node identifiers of the entire purchase order chain, determine the unique initial node corresponding to the order creation time, and concatenate its unique code with the order generation timestamp value to obtain the original string of the dynamic root anchor point; The original string is subjected to a one-way SHA-256 hash operation to obtain a hash value. The high-order bits of the string with a preset number of bits are extracted as the seed factor. The seed factor is XORed bit by bit with the unique link level number corresponding to the initial node to obtain the target random seed uniquely bound to the purchase order. Establish a hash association binding table between the target random seed and the order generation timestamp value and read the corresponding records. Use the timestamp value as the starting coordinate of the time series and the target random seed as the generation control parameter of the procurement forecast simulation branch to construct a shared time series benchmark for the generation of the simulation branch.

3. The method according to claim 1, characterized in that, The process of measuring the actual delay of all nodes in the fund transfer chain based on the shared timing benchmark, adjusting the tracking timeout threshold of each node based on the actual delay, collecting feedback signals based on the tracking timeout threshold, and determining the completeness of the fund path feedback signal based on the matching degree between the feedback signal and the execution timing of all nodes in the purchase order chain, includes: Based on the time-series starting coordinates of the shared time-series benchmark, the actual delay values ​​of each node in the entire capital flow chain are calculated, and a node delay sequence is formed according to the node order. For each actual delay value, if it exceeds the preset baseline timeout threshold of the corresponding node, the tracking timeout threshold will be increased to the sum of the actual delay value and the preset tolerance increment; if it does not exceed the threshold, the original threshold will remain unchanged. Collect fund path feedback signals with fund node response timestamps within the corresponding tracking window period, compare the response timestamps with the corresponding order node trigger times, and convert the absolute value of the time deviation into a time series matching degree value in the range of 0-1. The system calculates the percentage of nodes whose time-series matching degree reaches the preset matching threshold, and outputs the integrity judgment result of whether the node meets or does not meet the preset integrity threshold.

4. The method according to claim 1, characterized in that, When the integrity of the feedback signal from the funding path is lower than a preset integrity threshold, for the corresponding purchase order, the abnormal distortion simulation branch corresponding to the feedback signal lower than the preset integrity threshold is pruned, the bound random seed is reset, and the corrected purchase prediction simulation data is output, including: Extract the integrity judgment result of the fund path feedback signal. If the judgment result is not up to standard, extract the corresponding fund link identifier, reverse the triggering purchase order, and locate all purchase forecast simulation branches bound to the order. Compare the timestamps of simulated branch generation with the timestamps of responses to nodes with insufficient funding. Branches whose absolute time difference is lower than a preset time interval threshold are marked as abnormally distorted simulated branches and pruned to obtain a set of retained simulated branches. Read the target random seed originally bound to the purchase order, perform SHA-256 one-way hash operation based on the concatenation result of the current timestamp and the order number to generate the corresponding reset random seed, reset the random seed bound to the order, replace the original bound target random seed with the reset random seed, and integrate it with the retained simulation branch set to form the corrected purchase prediction simulation data; If the determination result is satisfactory, the current procurement forecast simulation data bound to the order will be directly used as the corrected procurement forecast simulation data.

5. The method according to claim 1, characterized in that, The step of extracting historical procurement multi-path simulation results, combining them with the corrected procurement prediction simulation data, eliminating contamination and interference branches, and outputting time-series alignment benchmark calibration results includes: Based on the business category and purchase amount range corresponding to the current purchase order, select historical simulation records from the pre-stored historical purchase multi-path simulation results that belong to the same business category as the current purchase order and whose purchase amount is within a preset range. Extract the timestamps of each simulated branch and the response timestamps of the corresponding funding nodes from the historical simulation records to construct a historical time-series comparison baseline under normal execution conditions of similar businesses; Read the generation timestamp of the retained simulation branch in the corrected procurement forecast simulation data, and calculate the absolute value of the difference between it and the corresponding position timestamp in the historical time series comparison baseline; When the absolute value of the difference exceeds the preset deviation threshold, the corresponding retained branch is marked as a pollution interference branch and excluded to eliminate the pollution interference on the abnormal detection benchmark and obtain the purified procurement prediction simulation data. Align the time-series coordinates of each branch of the purified procurement forecast simulation data with the response time-series of the corresponding nodes in the entire capital flow chain, and output the time-series alignment benchmark calibration result containing the time-series mapping relationship and correction offset record through zero-correction calculation of time-series offset.

6. The method according to claim 5, characterized in that, The step of extracting a real-time verification signal based on the time-series alignment benchmark calibration result and constructing a closed-loop calibration strategy based on the real-time verification signal includes: Extract the time-series mapping relationship and correction offset record from the time-series alignment benchmark calibration result, number them sequentially according to the fund tracking link nodes, and write the offset record into the corresponding node attribute field to obtain a node sequence carrying the correction offset; traverse the actual delay value collected in real time by each node in the node sequence, and use the written correction offset as the delay benchmark reference value to calculate the absolute value of the deviation between the two; if the absolute value of the deviation exceeds the preset unevenness threshold, generate a corresponding real-time verification signal based on the deviation value and the node number to supplement the verification basis for the node delay unevenness scenario; combine the closed-loop calibration rule set containing the node offset update value and the real-time verification signal trigger judgment condition, deploy the signal and rules to each node, and generate a closed-loop calibration strategy adapted to the real-time status of the fund tracking link.

7. The method according to claim 6, characterized in that, The step of adjusting the preset AI procurement forecasting model according to the closed-loop calibration strategy in terms of time series and link, defining the risk boundary range of inventory deviation, and obtaining the optimized inventory forecasting result based on the risk boundary range includes: Based on the offset update value and trigger judgment condition of the closed-loop calibration strategy, the set of nodes with uneven delay is extracted, and the deviation values ​​are extracted in the order of nodes to form a link state deviation vector. The vector is used as an additional input feature and input into the long short-term memory network used by the preset AI procurement prediction model. The vector is concatenated with the main input feature at the input layer. The model timing and link adaptation are adjusted through hidden layer operations to obtain the adapted network model. Input the historical inventory consumption data and procurement cycle information of the corresponding purchase order into the network model to obtain the inventory demand forecast value; The extreme values ​​of the deviation values ​​in the vector are extracted to form the risk boundary interval. The predicted value and the interval are encapsulated to obtain the optimized inventory prediction result. The time-series matching degree values ​​of each node in the entire capital flow chain are collected simultaneously.

8. The method according to claim 7, characterized in that, The optimized inventory forecast results are fused with the feedback signal to obtain a comprehensive risk score. This comprehensive risk score is then compared with a risk controllable range threshold, and risk assessment data is output, including: The inventory demand forecast value in the optimized inventory forecast result is read as the forecast center value, and the difference between the upper and lower limits of the risk boundary interval is read as the forecast fluctuation range. Read the time-series matching degree values ​​of each node in the synchronously collected fund path feedback signal; The predicted center value, the predicted floating range, and each of the time-series matching degree values ​​are sequentially concatenated to form a multi-dimensional fusion feature vector; The comprehensive risk score is obtained by multiplying each dimension parameter in the multidimensional fusion feature vector with the corresponding preset weight coefficient and then summing the results. The comprehensive risk score is compared with the risk controllable range threshold. If the comprehensive risk score is lower than or equal to the risk controllable range threshold, the risk of misappropriation of funds is determined to be within the preset controllable range. If the comprehensive risk score is higher than the risk controllable range threshold, the risk of misappropriation of funds is determined to be beyond the preset controllable range. The risk assessment data is output based on the determination result and the comprehensive risk score.

Images