Big data-based manufacturing research and development whole-process financial and tax risk early warning management system

Abstract

The application discloses a manufacturing industry R&D whole-process financial and tax risk early warning management system based on big data, and relates to the technical field of big data processing and financial and tax risk control, comprising a data acquisition module, a risk index model and an early warning output module.The data acquisition module is used for acquiring enterprise financial data, business data and tax data, the risk index model is used for threshold comparison of the data, and the early warning output module is used for triggering an early warning prompt when an over-standard condition is detected.The manufacturing industry R&D whole-process financial and tax risk early warning management system based on big data solves the problem that each expense cannot be accurately attributed to a single R&D project in complex scenes such as common production lines and multiple project cross of personnel, realizes whole-link closed-loop tracing from financial numerical value backtracking to equipment operation curve, material flow atlas and personnel space-time trajectory, and provides a reliable data basis for the risk index model.

Description

Technical Field

[0001] This invention relates to the field of big data processing and financial and tax risk management technology, specifically to a big data-based financial and tax risk early warning management system for the entire R&D process in the manufacturing industry. Background Technology

[0002] In the field of R&D and tax management in the manufacturing industry, digital transformation is deepening, and data-driven risk warning technology has begun to be applied to R&D expense collection and compliance control. Patent document CN112734580A discloses a data-driven risk warning system. This system acquires corporate financial, business, and tax data through a data acquisition module, and uses a preset risk indicator model to compare data on expense ratios, account classifications, and other dimensions against thresholds. When exceeding or deviating from these thresholds, an early warning is triggered, achieving automated monitoring and alerting of financial and tax risks. This solution, to a certain extent, replaces manual accounting and improves the efficiency of risk identification.

[0003] However, when applied to the entire R&D process management in manufacturing, the aforementioned existing technologies reveal a fundamental bottleneck: they only achieve physical-level collection and surface-level comparison of multi-source data, failing to overcome the challenge of substantive integration at the data logic layer. In real-world manufacturing R&D scenarios, R&D activities are highly dynamic and overlapping. The same high-precision machining center might be used for trial production of new components in the morning and switched back to production orders in the afternoon; the same R&D engineer might simultaneously serve two parallel R&D projects on the same workday. The data association mechanism relied upon by this patent is mainly based on fixed account codes or expense types for single-dimensional matching and statistics, lacking the ability to perform timestamp alignment and business semantic parsing of time records, material consumption lists, equipment operation logs, and financial accounting vouchers. When faced with complex situations such as shared production lines, personnel working on multiple projects, and material requisition documents not clearly indicating the R&D project affiliation, the system cannot accurately determine the true R&D project affiliation of each piece of business data, nor can it establish a full-chain reverse traceability relationship from financial vouchers to original business documents. This results in a serious lack of credibility in the constructed project data map. The allocation of various expenses across different projects can only rely on financial personnel's experience to manually segment the data at the end of the period, rather than on an automated and substantive correlation based on underlying data logic. This inherent deficiency in data fusion directly leads to the input data upon which the upper-level risk indicator model relies being distorted or ambiguous. Consequently, the conclusions of risk warnings lose their objective basis, and false alarms and omissions become intractable systemic problems.

[0004] Therefore, the technical problem to be solved by this invention is how to achieve automated and substantial integration of multi-source heterogeneous business, financial and tax data in the entire R&D process of manufacturing, and generate a standardized, traceable, full-link data twin with a single R&D project as the smallest granularity. Summary of the Invention

[0005] The purpose of this invention is to provide a big data-based financial and tax risk early warning management system for the entire R&D process in the manufacturing industry, in order to solve the problems mentioned in the background.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a big data-based financial and tax risk early warning management system for the entire manufacturing R&D process, comprising a data acquisition module, a risk indicator model, and an early warning output module. The data acquisition module is used to acquire enterprise financial data, business data, and tax data. The risk indicator model is used to perform threshold comparisons on the data. The early warning output module is used to trigger an early warning notification when a threshold is detected.

[0007] The system also includes:

[0008] The R&D business scenario is reconstructed into an intelligent agent, with an industrial digital twin engine and causal inference algorithm built in. It is configured to actively pull multi-source heterogeneous raw data, including time-series data of equipment IoT, material flow records, and personnel workstation sensor data. It dynamically replays the physical activity flow within the R&D cycle in virtual space and generates the first version of the business flow mapping by inferring the time-series causal chain between material consumption, equipment occupancy and personnel activities.

[0009] The financial and tax regulations deconstruction and stress testing intelligent agent has a built-in knowledge computing engine that transforms the financial and tax policy clauses into a computable constraint model. It is configured to receive the first version of the business flow mapping graph and perform simulated audit path search on the nodes and edges in the mapping graph. It exhaustively lists the logical vulnerabilities and evidence chain breaks that lead to non-compliance in the allocation and generates a contradiction and conflict report carrying risk coefficients and attack paths.

[0010] An adaptive adversarial evolution controller is connected to the R&D business scenario reconstruction agent and the financial and tax regulations deconstruction and stress testing agent, respectively. It is configured to project the conflict report as a challenge input to the R&D business scenario reconstruction agent, drive it to introduce implicit evidence to adjust the parameters of the causal inference model to reconstruct the business logic, and submit the reconstructed business flow map to the financial and tax regulations deconstruction and stress testing agent again. This process is repeated for multiple rounds of adversarial iteration under non-cooperative game, until the financial and tax regulations deconstruction and stress testing agent can no longer generate an effective attack path. Finally, it outputs a high-confidence business and financial and tax digital twin that has converged after the adversarial game.

[0011] The risk indicator model is specifically configured to perform risk calculations based on the high-confidence business, finance, and tax digital twin.

[0012] Preferably, when the R&D business scenario reconstruction agent infers the temporal causal chain, it specifically executes the transfer entropy algorithm to quantify the information transmission intensity between the fluctuation of equipment operating parameters and the subsequent material consumption pulse signal, so as to identify the nonlinear driving relationship between seemingly unrelated events, thereby constructing a business flow mapping diagram containing the event context.

[0013] Preferably, when the tax and financial regulations deconstruction and stress testing agent performs simulated audit path search, it specifically adopts the Monte Carlo tree search algorithm to conduct game theory in the state space composed of all possible cost allocation schemes. For each cost allocation node in the current business flow map, it systematically explores the reverse evidence chain that can prove its improper allocation.

[0014] Preferably, when the intelligent agent reconstructing the R&D business scenario makes dynamic adjustments after receiving the conflict report, the implicit evidence it introduces includes retrieving historical equipment process parameter curves of the same production equipment during the same time period, and using a dynamic time warping algorithm to measure similarity, so as to verify whether the equipment operation mode during that period is closer to R&D trial production or mass production processing, thereby responding to the questioning of the stress test with quantitative results.

[0015] Preferably, the adaptive adversarial evolution controller has a built-in convergence judgment unit. This unit continuously monitors the conflict reports generated by the financial and tax regulations deconstruction and the stress test agent. When the risk coefficient of the reports generated in two consecutive iterations is lower than a preset threshold, or no new attack paths are generated, the adversarial relationship is considered to have reached Nash equilibrium. At this time, the output financial and tax digital twin is the final result.

[0016] Preferably, each data item in the final output of the business, finance and tax digital twin of the system is accompanied by a complete traceability evidence chain generated during the adversarial game to prove its compliance. The evidence chain clearly points to the stress test and business logic reconstruction records of a specific round.

[0017] Preferably, when collecting material flow records, the data acquisition module specifically obtains the unique code, movement trajectory and time node of the material through radio frequency identification technology, and uses this data as the basic spatiotemporal coordinates for the construction of physical activity flow by the intelligent agent reconstructing the R&D business scenario.

[0018] Preferably, the early warning output module is communicatively connected to the risk closed-loop handling unit. The risk closed-loop handling unit is used to automatically generate a risk early warning form containing a description of the risk point, rectification suggestions and rectification deadline after receiving the early warning information, and to track the rectification submission vouchers and rectification progress.

[0019] Preferably, the system further includes a policy dynamic adaptation unit, which has a built-in natural language processing engine for capturing and parsing rule elements in the latest fiscal and tax policy texts in real time, and automatically updating the parsed structured rules to the knowledge computing engine of the fiscal and tax regulations deconstruction and stress testing agent.

[0020] Preferably, the business, finance and tax digital twin uses the unique code of the R&D project as the primary key for data association, and organizes personnel labor costs, direct input costs, depreciation costs and intangible asset amortization into a standardized data system that can be traced throughout the entire chain with an immutable data structure, so that external auditors can directly access it.

[0021] This invention provides a big data-based financial and tax risk early warning management system for the entire R&D process in the manufacturing industry. It has the following beneficial effects:

[0022] By constructing an adversarial game mechanism between the R&D business scenario reconstruction agent and the financial and tax regulations deconstruction and stress testing agent, the previously fragmented equipment operation records, material flow trajectories, and personnel activity data in the entire manufacturing R&D process can be recreated in a virtual simulation space along a unified timeline. Furthermore, causal inference algorithms such as transfer entropy can automatically identify the true driving relationships between different business events. This mechanism no longer relies on manual annotations on material requisition documents or fixed subject mapping rules, solving the problem of accurately attributing each expense to a single R&D project in complex scenarios such as shared production lines and personnel working on multiple projects. The resulting business, financial, and tax digital twin, after multiple iterations, includes a complete traceability evidence chain originating from the adversarial process for each expense collection. This achieves a closed-loop traceability from financial values ​​back to equipment operation curves, material flow maps, and personnel spatiotemporal trajectories, providing a reliable data foundation for risk indicator models.

[0023] The adaptive adversarial evolution controller continuously monitors changes in risk coefficients and the emergence of new attack paths in conflict reports, automatically determining the convergence state of the adversarial process. This allows the system to terminate computation under appropriate conditions without requiring manual preset iteration counts, balancing the reliability of data fusion results with reasonable consumption of computational resources. The Monte Carlo tree search algorithm embedded within the tax and financial regulations deconstruction and stress testing agent systematically explores potential compliance vulnerabilities in a high-dimensional state space composed of all possible cost-sharing schemes. These vulnerabilities are then submitted to the R&D business scenario reconstruction agent in the form of explicit attack paths for targeted responses, thereby gradually eliminating auditable points during the iteration process. The policy dynamic adaptation unit continuously monitors and parses the latest policy texts through a natural language processing engine, automatically updating the structured rules to the knowledge computing engine. This enables the system to automatically adjust the judgment boundaries of stress testing when facing policy adjustments, avoiding compliance misjudgments due to outdated rules. The risk closed-loop handling unit automatically transforms early warning information into early warning forms containing rectification suggestions and deadlines, and continuously tracks the handling progress, thus breaking down management gaps from risk identification to rectification review and forming a complete control closed loop. Attached Figure Description

[0024] Figure 1 This is a data flow diagram between the modules of the big data-based manufacturing R&D full-process financial and tax risk early warning management system of the present invention;

[0025] Figure 2 This invention reconstructs the internal entropy transfer and causal inference process of the intelligent agent in the R&D business scenario of the big data-based manufacturing R&D full-process financial and tax risk early warning management system.

[0026] Figure 3 This is a state diagram of the convergence judgment mechanism of the adaptive adversarial evolution controller in the big data-based manufacturing R&D full-process financial and tax risk early warning management system of the present invention. Detailed Implementation

[0027] 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.

[0028] Please see Figures 1 to 3This invention provides a technical solution: a big data-based financial and tax risk early warning management system for the entire manufacturing R&D process. In the system, which is based on basic modules such as data collection, risk indicators and early warning output, this technical solution adds three core intelligent agents that work together to build a computing environment that can substantially integrate multi-source heterogeneous industrial and tax data.

[0029] After the system starts, the intelligent agent for reconstructing the R&D business scenario begins to work. This agent is equipped with an industrial digital twin engine and a causal inference algorithm. Its processing flow is not passively receiving manually formatted structured forms, but actively establishing connections with IoT sensor interfaces in the production execution system, warehouse management system, and human resources system. The data it retrieves includes, but is not limited to, second-by-second spindle load rate and speed time-series data output by CNC equipment, tag movement records captured by RFID readers during material flow, and in-situ status change sequences recorded by sensors at R&D personnel's workstations. After acquiring this fragmented data recorded at different time granularities and in different formats, the digital twin engine constructs a virtual simulation space in memory that completely corresponds to the spatiotemporal coordinate system of the physical workshop. This space advances forward according to a unified timeline, recreating the scenarios of overlapping or adjacent timestamps of equipment actions, material displacements, and personnel behaviors, vividly reproducing the trial production task switching process of a certain machining center during a specific morning period or the entire trajectory of a batch of experimental samples flowing on the production line. Based on this, the causal inference algorithm intervenes, making attribution judgments not only based on the item remarks field on the material requisition form. The algorithm employs transfer entropy analysis and nonlinear causal network discovery. When a high-speed, low-feed processing mode occurs, the algorithm traces material consumption pulse signals within the preceding and following time windows, such as the outbound records of a specific grade of engineering plastic used for experimental verification. This establishes a clearly defined driving and driven relationship between seemingly unrelated equipment behavior and material flow. In this way, the algorithm weaves discrete event points into a first-version business flow map with a complete causal framework. Each node in the map represents a specific R&D activity, and each edge represents a causal relationship of resource consumption or activity transfer.

[0030] After the first version of the business flow map was generated, it was immediately passed to the tax and financial regulations deconstruction and stress testing agent. This agent was pre-loaded with a knowledge computing engine. This engine is not a simple rule base, but rather it parses the textual clauses in relevant policy documents such as R&D expense deduction and high-tech enterprise certification regarding the scope of expense collection, allocation basis, asset depreciation standards, and prohibitions into a series of machine-executable computable constraint models. When nodes and edges in the business flow map enter this agent, the engine immediately initiates a multi-path deep search process simulating a tax audit. This process uses the Monte Carlo tree search algorithm to deduce in a high-dimensional state space composed of all possible expense allocation logics. The system puts itself in the audit perspective and performs an attack verification on the equipment depreciation allocation of a specific project in the map. It does not only look at the allocation ratio agreed in the project initiation document, but also retrieves the comprehensive operation logs of the equipment for processing other mass production orders in the adjacent time period, generating attack paths to prove that the equipment's operation mode during that period does not have the uniqueness of R&D trial production. Similarly, in the aggressive verification of R&D personnel's work hours attribution, the system automatically searches for digital trajectory records of the same person performing routine production tasks in the third-floor workshop during the same period, thereby questioning the exclusivity of their R&D work hour reporting. This process exhaustively examines all logical loopholes that could lead to a breach in the supporting evidence for a certain allocation being considered as compliance, ultimately forming a conflict report. Each item in the report is accompanied by a detailed deductive reasoning process and a quantified risk coefficient.

[0031] The aforementioned conflict report is pushed to the adaptive adversarial evolution controller. This controller acts as a link between the first two agents; it does not directly modify the data but instead projects the report as an adversarial challenge input to the R&D business scenario reconstruction agent. The R&D business scenario reconstruction agent must respond to this report at the underlying logical level.

[0032] In response to concerns about the lack of uniqueness in equipment operation modes, the intelligent agent automatically expands the scope of data retrieval and introduces implicit evidence. For example, it extracts frequent tool change records of the equipment's tool magazine within a certain time period, or directly compares the differences in instruction code sequences between historical trial production tasks and mass production tasks of the same CNC system. It uses a dynamic time warping algorithm to measure the similarity between the current machining curve and the historical R&D curve, thereby quantifying the data to prove the R&D attributes of its behavior mode.

[0033] In response to attacks that target employees' exclusive work hours, the intelligent agent will trace the movement trajectory of the location tag carried by the employee on the map, and combine it with the time of the employee's operation on the task card on the R&D project management board to reconstruct the exact time node when the employee switched from R&D tasks to production tasks. Based on this, the agent will accurately adjust the weight of the work hour ratio parameter in the causal model.

[0034] After completing this round of implicit evidence mining and model parameter correction, the new business flow map is submitted again to the tax and regulatory deconstruction and stress testing agent. This process, under the supervision of the controller, repeats itself at machine-to-machine processing speed. The controller has a built-in convergence judgment unit that continuously monitors changes in the risk coefficient of conflict reports and the number of discovered attack paths. When the tax and regulatory deconstruction and stress testing agent is no longer able to identify any valid logical vulnerabilities not covered by the evidence chain of the current business flow map in a certain round of testing, and all cost attribution nodes have been cross-verified by multi-source evidence, including equipment operation curves, material flow maps, and personnel spatiotemporal trajectories, the controller determines that the confrontation has reached an equilibrium state and then terminates the iteration.

[0035] After the iteration terminates, the final data structure extracted from virtual memory by the system is a highly confident business, financial, and tax digital twin. This twin uses R&D projects as the primary key for data organization. Its internal data, such as personnel costs, direct inputs, and depreciation expenses, are no longer isolated financial figures, but rather each item is connected to a set of immutable traceable evidence chains preserved in the adversarial game. This digital twin is then provided to the risk indicator model, which performs subsequent subject threshold comparisons and comprehensive risk calculations, thereby ensuring that the conclusions of the early warning output are based on irrefutable data.

[0036] In constructing the business flow mapping graph, the intelligent agent for reconstructing business scenarios does not use simple triggers based on preset rules. Instead, it uncovers deeper driving mechanisms by calculating the information transmission relationships between variables. Specifically, it employs the transfer entropy algorithm.

[0037] The agent first transforms the collected multi-source heterogeneous raw data into a time-series format suitable for information-theoretic computation. For equipment operating parameters, the agent discretizes continuous analog quantities such as spindle load rate, speed, and feed rate output by the CNC system according to a unified time window, forming a sequence of equipment state changes. For material consumption data, the agent arranges each material movement event captured by the RFID reader in chronological order and encodes attributes such as material type and direction of movement into discrete symbols, constructing a sequence of material flow events. For the workstation sensor data of R&D personnel, the agent converts the records of changes in their position status into a state sequence indicating whether a person is in the R&D area. These sequences constitute the set of input variables for transfer entropy calculation.

[0038] After data serialization, the agent begins calculating the transfer entropy. The transfer entropy algorithm, based on the Shannon entropy concept in information theory, measures the degree to which the past state of one time series reduces the uncertainty of the future state of another time series. This reduction represents the amount of information transferred from the source variable to the target variable. When processing the potential correlation between equipment behavior and material consumption, the agent selects a specific sequence of equipment operating parameters as the source variable and a specific sequence of material consumption events as the target variable. The algorithm first calculates the conditional entropy of the target variable itself, given its historical state. This conditional entropy represents the residual uncertainty in predicting future changes based solely on the historical patterns of material consumption. Subsequently, the algorithm introduces the historical information of the equipment operating parameter sequence as an additional known condition, recalculating the conditional entropy of the target variable when both its own history and the equipment history are known. The difference between the two conditional entropies represents the amount of information transferred from equipment behavior to material consumption, quantified in bits.

[0039] When the calculated transfer entropy is higher than a pre-set confidence threshold, it indicates that changes in equipment operating parameters indeed precede and drive changes in subsequent material consumption event patterns, and there is a clearly defined information transfer channel between the two. For example, when analyzing high-speed, low-feed processing modes and engineering plastic outbound records, if the calculated transfer entropy value from the former to the latter is significantly higher than the random noise level, it is determined that this specific processing mode is the driving factor leading to the use of experimental engineering plastics. Conversely, if the transfer entropy value is extremely low, it indicates that although the two are temporally adjacent, they are not in a driving-driven relationship. By iteratively performing this calculation on all possible equipment state variables and material consumption variables in the entire dataset, the agent can automatically identify which specific equipment operating modes and which subsequent material consumptions form stable and reliable information transfer relationships in a seemingly chaotic event sequence. These driving and dependency relationships quantified and confirmed by the transfer entropy algorithm are then used to construct causal edges in the business flow map, thus giving the business logic expressed by the map a rigorous mathematical foundation.

[0040] The specific algorithm used by the agent in deconstructing and stress testing financial and tax regulations to perform simulated audit path search was limited to the Monte Carlo tree search algorithm. This choice enables the agent to perform forward-looking game theory in complex situations regarding the verification process of cost attribution in the business flow map.

[0041] After receiving the first version of the business flow map, the tax and financial regulations deconstruction and stress testing agent first transforms each cost allocation node that needs to be verified in the map, such as the allocation of equipment depreciation to a certain R&D project, into an initial state node for Monte Carlo tree search. Based on this, the algorithm constructs a high-dimensional state space that combines all possible cost allocation schemes with audit judgment results. Each dimension of this state space corresponds to different combinations of evidence and policy interpretation branches for cost allocation.

[0042] The agent begins executing a Monte Carlo tree search in four alternating cyclical phases: selection, expansion, simulation, and backpropagation. In the selection phase, the algorithm starts from the root node of the current state tree, calculates the priority of each node based on its existing visit count and value score according to the confidence interval upper bound formula, and recursively selects a leaf node with further exploration value. In the expansion phase, the algorithm adds a new child node after the leaf node. This child node represents an unexplored audit branch, such as retrieving specific supporting evidence for the equipment depreciation, like the equipment's workpiece type switching record during the time period. In the simulation phase, the algorithm starts from the newly expanded node and performs a rapid deduction process, randomly generating a subsequent audit judgment sequence according to a preset probability distribution, proceeding downwards until a conclusion is reached. This conclusion either deems the cost allocation compliant or identifies a clear broken chain of evidence, thus determining non-compliance. In the backpropagation phase, the algorithm updates the visit count and value assessment values ​​of all nodes along the entire path from the expanded node back to the root node, based on the compliance or non-compliance conclusion obtained in this simulation.

[0043] Through massive iterations of the four-stage process described above, the Monte Carlo tree search gradually converges. The agent systematically extracts reverse evidence chains with high confidence that prove the current fee allocation is improper by tracking frequently visited branches in the tree where value assessments show a negative orientation. Each complete attack path clearly lists the key points of contention and the logical deductive relationships between the various pieces of evidence supporting those points, forming persuasive rebuttal evidence in the conflict report.

[0044] This study defines how an intelligent agent reconstructing R&D business scenarios should introduce and process implicit evidence for dynamic adjustments when responding to conflict reports. When the agent receives challenges from the financial and tax regulations deconstruction and stress testing agents, especially attacks alleging that the equipment operation mode during a certain period is not unique to R&D and trial production, its internal data mining mechanism is immediately activated to search data dimensions beyond direct business records.

[0045] The intelligent agent first identifies the complete historical operating records of the same production equipment within the questioned time period. These records are not simply on / off states or output counts, but rather time series of multi-channel process parameters sampled and stored at a fixed frequency by the equipment's CNC system. Common channels include spindle load rate, servo motor current values ​​for each axis, feed rate, coolant flow rate, and tool change signals. The agent's data access layer extracts a snapshot of all parameters for the equipment within a certain time window before and after the specified period from the historical database of the enterprise's manufacturing execution system or equipment IoT platform.

[0046] After acquiring the raw curve data, the agent executes a dynamic time warping algorithm. This algorithm is specifically designed to measure the similarity of two time series in their overall shape and can tolerate local scaling and shifts in the time axis. The agent uses the equipment process parameter curve extracted from the currently questioned time period as the comparison sequence. Simultaneously, it retrieves two types of reference sequences from the company's historical database based on pre-labeled typical operating condition tags for comparison. The first type of reference sequence is taken from the process parameter records of the equipment during its history of performing clearly defined R&D and trial production tasks. These tasks are typically assigned by the R&D department during project initiation, with the system locking the equipment's exclusive time period and leaving an R&D task number. The second type of reference sequence is taken from the equipment's operating records under stable mass production conditions. These records correspond to time periods confirmed by the production system as large-scale processing periods with continuous production of qualified products.

[0047] Before formally executing the dynamic time warping algorithm, the agent performs dimensionless and zero-mean processing on the three curves to eliminate systematic biases caused by differences in sensor calibration or different batch data acquisition formats. The algorithm then constructs a two-dimensional cost matrix, where rows correspond to the time steps of the sequence to be compared, and columns correspond to the time steps of the reference sequence. The value of each cell in the matrix represents the Euclidean distance between the multidimensional process parameter vectors at two time points. On this cost matrix, the algorithm uses dynamic programming to find a warped path from the lower left corner to the upper right corner of the matrix. This path satisfies boundary conditions, continuity, and monotonicity constraints, and minimizes the cumulative cost of each cell along the path. This minimum cumulative cost serves as the dynamic time warped distance between the sequence to be compared and the reference sequence; the smaller this distance, the closer the shapes of the two curves are.

[0048] The agent calculates a first distance between the curve to be compared and the R&D / prototype reference curve, and a second distance between the curve and the mass production reference curve. After obtaining these two quantitative results, the agent directly compares the first and second distance values. If the first distance value is lower than the second distance value, meaning the current equipment operating curve is more similar in shape to the historical R&D / prototype curve, this result serves as quantitative evidence and is directly incorporated into the reconstructed business flow map to maintain the rationality of the original allocation. If the second distance value is lower than the first distance value, it indicates that the actual behavior pattern of the equipment during this period is more inclined towards mass production. Based on this, the agent lowers the allocation weight of equipment depreciation for the R&D project during this period in the original map and transfers the corresponding proportion to the production cost center, achieving dynamic allocation correction based on curve shape similarity measurement.

[0049] The convergence judgment mechanism within the adaptive adversarial evolution controller is specifically defined. As the central coordinating unit connecting the R&D business scenario reconstruction agent and the financial and tax regulations deconstruction and stress testing agent, the controller's core task is to monitor the changes in conflict intensity during the adversarial game process and terminate the iteration at an appropriate time to lock in the final result.

[0050] The adaptive adversarial evolution controller has a built-in convergence judgment unit, which is triggered to execute after each complete adversarial iteration round. A complete adversarial iteration round refers to the tax and financial regulations deconstruction and stress testing agent completing a stress test on the current business flow map and generating a conflict report. Subsequently, the R&D business scenario reconstruction agent responds to the various questions in the report and outputs an updated business flow map. This completes one round.

[0051] At the end of each round, the convergence judgment unit retrieves the conflict report generated in that round from the tax and financial regulations deconstruction and stress testing agent. This report contains several specific attack paths, each carrying an independent risk coefficient. This risk coefficient is determined by the tax and financial regulations deconstruction and stress testing agent when generating the attack path, based on a comprehensive assessment of the strength of the evidence upon which the path relies and the clarity of the policy violation. The convergence judgment unit first extracts the highest risk coefficient value from the conflict report of this round as the overall conflict intensity index for this iteration. Simultaneously, this unit compares the identifiers of the attack paths listed in this round's report with the attack path identifiers in the previous round's report item by item.

[0052] The convergence judgment unit maintains two termination conditions that can be triggered alternately. The first condition is the risk coefficient decay condition, which is triggered when the overall conflict intensity index of two consecutive complete iterations is lower than a preset internal threshold. This preset threshold is set with reference to the upper limit of the acceptable residual tax compliance risk for enterprises. When the harm intensity of the attack path has been reduced to below this limit, it indicates that the tax regulation deconstruction and stress testing agent can no longer raise compliance vulnerabilities with substantial threats. The second condition is the attack path convergence condition, which is triggered when no new attack path identifiers that did not appear in the previous round of reports appear in the conflict report of this round. That is, when all the questions raised by the tax regulation deconstruction and stress testing agent are old questions that have been raised and attempted to be answered in previous rounds, and no previously unrevealed breaks in the chain of evidence are found.

[0053] When the convergence judgment unit detects that either of the above two conditions is met, it immediately notifies the adaptive adversarial evolution controller to stop the iterative loop. The controller then stops sending new adversarial tasks to the two agents and reconstructs the latest business flow mapping graph output by the agents in the current R&D business scenario. This mapping graph is then locked into the final business, finance, and tax digital twin that has been verified through adversarial game theory and has reached a Nash equilibrium state, for subsequent use by the risk indicator model.

[0054] The internal structure of the final output digital twin of business, finance and taxation was further defined, and the composition and source of the traceability evidence chain carried by each data link in the digital twin were clarified.

[0055] When the adaptive adversarial evolutionary controller determines that the iteration has terminated, and the system extracts the final version of the business flow map that has reached the Nash equilibrium state from virtual memory and solidifies it into a business, finance, and tax digital twin, it does not simply output a static cost allocation table. The data organization structure of this digital twin uses the unique code of the R&D project as the top-level index, expanding downwards to each specific cost allocation detail, such as the hourly rate of a certain R&D personnel allocated to a certain project in a certain week, or the depreciation and amortization amount of a certain piece of equipment allocated to a certain project on a certain day. For each of these detailed cost allocation records, the system embeds a related chain of traceability evidence within its data body.

[0056] The information in this traceability evidence chain originates from all the attack and defense records accumulated during the aforementioned multiple rounds of adversarial game. In each iteration, when the financial and tax regulations deconstruction and stress testing agent proposes a specific attack path for a particular expense attribution, this attack path is assigned a unique evidence package identifier and recorded. Subsequently, the R&D business scenario reconstruction agent extracts supporting data to respond to challenges from implicit data sources such as IoT devices, warehouse RFID records, or personnel location trajectories for this attack path. This supporting data, along with its extraction timestamp and data source interface information, is added to the evidence package. In subsequent rounds, if the expense attribution is attacked again, new attack paths and new response evidence are continuously added to the evidence package until the iteration terminates.

[0057] Upon termination of the iteration, for each expense attribution record ultimately retained in the digital twin, its associated traceability evidence chain fully records all the challenges and successful responses experienced throughout the entire adversarial game. The evidence package clearly identifies the stress test round number, pointing to the key questions raised by the financial and tax regulations deconstruction and stress test agent in each round, and correspondingly links to the specific storage location of the response evidence retrieved and adopted by the R&D business scenario reconstruction agent in the corresponding round. When subsequent internal audits or external tax inspections need to verify the compliance basis of a particular R&D expense collection, by accessing the traceability evidence chain attached to that expense record, all supporting materials can be viewed step-by-step, from the similarity comparison results of equipment operation curves and the causal inference conclusions of material flow maps to the cross-verification records of personnel spatiotemporal trajectories, forming a complete evidence loop tracing back from financial values ​​to physical facts.

[0058] The specific technical means for the data acquisition module to obtain material flow records and the positioning of this data in the intelligent agent reconstruction of R&D business scenarios are defined. Radio frequency identification (RFID) technology is identified as the core method for material data acquisition, and the role of the acquired data in constructing the physical activity flow is explained.

[0059] When collecting material flow data, the data acquisition module deploys RFID readers at key logistics nodes within the enterprise, such as raw material warehouses, semi-finished product buffer areas, production line workstations, R&D and trial production areas, and finished product warehouses. Every material entering the R&D or production process, including raw materials, work-in-process, experimental samples, and purchased test components, has an RFID tag affixed to its outer packaging or carrier. This tag stores a unique code pointing to the batch of the material, which is bound to the material throughout its entire lifecycle. When tagged materials pass within the reading range of any RFID reader via forklift, automated guided vehicle, or manual handling, the reader captures the coded information reflected back from the tag and simultaneously records the precise time of the reading operation. This time is generated by the reader's internal clock synchronized with the system's network time protocol server. The reader combines the read tag code and the reading time into a material flow event record, which is uploaded in real-time to the data acquisition module's front-end message queue via the network. The data acquisition module performs deduplication on the received material flow event records. When the same tag is read repeatedly by the same reader within a short period of time, only the first read record is retained. Subsequently, the event record is persistently stored and can be called by the intelligent agent for reconstructing business scenarios in the R&D field.

[0060] After acquiring records of material flow events, the intelligent agent used in the R&D business scenario reconstruction process treats these records as the foundational spatiotemporal coordinates for constructing the physical activity flow in the virtual simulation space. The agent parses each material flow event into a spatiotemporal node with three-dimensional information. The unique code of the material serves as the node's identity, the physical location of the reader / writer serves as the node's spatial coordinates, and the reading time node serves as the node's temporal coordinates. Following a unified timeline, the agent sequentially arranges spatiotemporal nodes belonging to different materials within the digital twin environment. When a specific material passes through multiple readers / writers in the time dimension, the agent reconstructs the material's complete movement trajectory within the workshop by connecting these spatiotemporal nodes. These trajectories are superimposed with equipment operation events and personnel activity trajectories, aligned with timestamps, forming the event flow skeleton driving the entire virtual simulation space. This provides a spatiotemporal reference for subsequent causal inference algorithms to analyze the driving relationship between material consumption and equipment behavior and personnel activities.

[0061] The downstream handling mechanism of the early warning output module was defined, and the composition and workflow of the risk closed-loop handling unit were clarified. This unit does not merely stop at the level of risk alerts, but establishes a communication connection with the early warning output module to form a complete control closed loop from risk identification to rectification and review.

[0062] Once the early warning output module completes risk calculations based on a high-confidence business, finance, and tax digital twin and determines that a certain risk point has met the early warning triggering conditions, the early warning output module pushes the risk information in the form of a structured data message to the message queue of the risk closed-loop handling unit through the system's internal message bus. This data message includes the risk point's type code, risk level, a unique identifier for the corresponding expense record in the business, finance, and tax digital twin that triggered the risk, the policy clause index on which the early warning was triggered, and a risk occurrence probability score.

[0063] Upon receiving the message, the built-in template engine of the risk closed-loop handling unit immediately initiates the automatic generation process of the risk warning form. The template engine pre-stores warning form template frames corresponding to different risk type codes. The engine extracts key fields from the message and fills them into the designated placeholder areas of the corresponding templates. The generated risk warning form's header clearly states the risk point description, explaining the specific attribution of expenses and the specific manifestations triggering the warning in natural language. For example, it may indicate that commissioned R&D expenses under a certain project were not calculated according to the policy-stipulated ratio for additional deduction. The warning form automatically generates rectification suggestions, driven by handling knowledge entries related to the risk type in the rule base, providing operational directions for the responsible party. For example, it may require re-verification of whether there is a relationship between the commissioning party and the commissioned party, and supplementary pricing explanation materials based on the arm's length principle. The warning form's footer specifies the rectification period. The initial default value of this period is determined by the risk level: general warnings correspond to a relatively ample working day period, important warnings correspond to a compressed working day period, and major warnings correspond to an emergency response period.

[0064] After generating an early warning notice, the risk closed-loop handling unit sends it to the terminal interface of the responsible person through the company's internal message push channel. Simultaneously, the system starts a background rectification tracking timer, which is linked to the rectification deadline specified in the early warning notice. Upon receiving the early warning notice, the responsible person can submit the rectification progress status in the system interface and upload rectification supporting documents. These documents can be scanned copies of corrected accounting vouchers, supplementary signed technical contract attachments, re-certified R&D personnel lists, or adjusted cost allocation calculation sheets. The risk closed-loop handling unit continuously monitors the supporting document records and status change events submitted by the responsible person, archiving the rectification progress records and attached documents one by one into a subdirectory under the early warning notice. When the timer is about to expire and the rectification status has not yet changed to "completed," the unit automatically sends a reminder notification to the responsible person that the deadline is approaching and simultaneously sends a notification message to their immediate supervisor, thereby ensuring the closed-loop progress and traceability of the handling process.

[0065] The policy dynamic adaptation unit is defined, which introduces the ability of the system to continuously perceive and internalize external policy changes. Its core lies in the automatic update link between the built-in natural language processing engine and the knowledge computing engine of the fiscal and tax regulations deconstruction and stress testing intelligent agent.

[0066] The operation of the policy dynamic adaptation unit begins with targeted monitoring of policy data sources. During system initialization, this unit pre-sets a list of official publishing platforms to be monitored, covering announcement pages from the State Taxation Administration, the Ministry of Finance, and the science and technology administrative departments. The unit's built-in targeted data collection component incrementally polls these addresses at preset time intervals. When a change in page content or the addition of a file link is detected, the collection component retrieves the title and summary information of the newly published file to a local temporary storage area and compares it with the list of processed files. If the comparison result indicates that the file has not been processed previously, the complete file download process is automatically triggered, storing the full text of the policy document in structured text format into the original policy database.

[0067] After the original policy text enters the parsing pipeline, the policy dynamic adaptation unit calls its built-in natural language processing engine for deep processing. This engine first performs sentence and word segmentation on the policy text, dividing the continuous policy provisions into the smallest semantic units suitable for syntactic analysis. Then, it initiates a named entity recognition process, specifically designed to extract specific entity types from the fiscal and tax policy text. These include the type of applicable subject, the name of the expense collection category, the proportion threshold, the start and end dates of implementation, and the items in the filing materials list. Simultaneously, the engine performs semantic relationship extraction, identifying the limiting and restricted relationships between the aforementioned entities in the text. For example, it can parse from the wording of the provisions information a rule change information indicating that a certain type of expense was removed from the scope of additional deductions after a certain date.

[0068] After extracting entities and relationships, the natural language processing engine reassembles these scattered semantic fragments according to a pre-defined structured rule template. For each identified rule change, the engine generates a corresponding structured rule update record. The record's fields include the change type (e.g., addition, modification, or repeal), the applicable fee category code, the updated threshold parameter or logical expression, and the identifier and specific clause location of the source policy document on which the update is based.

[0069] This unit pushes the generated structured rule update records directly to the knowledge computing engine mounted on the tax and finance regulation deconstruction and stress testing agent through the system's internal data synchronization interface. Upon receiving the update records, the knowledge computing engine performs corresponding modification operations on its own set of computable constraint models. For records with a change type of "modification," the engine locates the constraint nodes in the model corresponding to the expense category and old rule parameters, replacing their parameter values ​​or decision logic expressions. For records with a change type of "addition," the engine adds a constraint node at the corresponding level in the model and establishes mutually exclusive or superimposed logical connections with other nodes at the same level. For records with a change type of "abolishment," the engine marks the corresponding constraint node as disabled, preventing it from participating in subsequent stress test calculations. After the update operation is completed, the knowledge computing engine returns a successful update confirmation signal to the policy dynamic adaptation unit. At this point, in the new round of adversarial iteration, the tax and finance regulation deconstruction and stress testing agent automatically uses the updated policy constraint model to perform stress testing on the business flow map.

[0070] The system defines the data organization logic and storage format of the final output digital twin of business, finance and taxation, specifies that the unique code of the R&D project should be used as the primary key for data association, and puts forward specific requirements for the classification of expense categories and the immutability of data structure.

[0071] Once the adaptive adversarial evolution controller determines that the adversarial process has reached equilibrium and terminates the iteration, the system extracts the current version of the business flow map from memory and enters the solidification and encapsulation stage of the digital twin. First, based on the project initiation number in the R&D project management system, the system assigns a globally unique code to each ongoing or completed R&D project. This code serves as the top-level index key for the digital twin. Under this top-level index, the system establishes four primary aggregation partitions according to the expense categories stipulated in the R&D expense deduction policy, corresponding to personnel costs, direct input costs, depreciation costs, and intangible asset amortization, respectively. Each expense attribution record verified and retained during the adversarial game is assigned to its corresponding primary partition based on its business attributes.

[0072] Within each expense attribution record, the system establishes a bidirectional link path: a unique project code points to the record, and from that record, it points down to the original business voucher and the traceability evidence chain. Upward, the record uses the unique project code as a foreign key to ensure its attribution to the corresponding R&D project is unambiguous at the data model level. Downward, the record stores a locator pointing to the original business voucher. This locator can be the storage address of the corresponding accounting voucher in the Enterprise Resource Planning system or an index identifier of the corresponding equipment operation log in the Manufacturing Execution System. Simultaneously, the record also stores the storage path of the associated traceability evidence chain. This evidence package is the compliance proof record set jointly constructed by the R&D business scenario reconstruction agent and the financial and tax regulations deconstruction and stress testing agent during the adversarial iteration process.

[0073] To ensure the immutability of the data structure, the system calculates a hash digest value for each expense collection record when solidifying the digital twin. The data fields included in the calculation include the project attribution code, expense category code, amount, accounting period, associated original business voucher locator, and the hash values ​​of all contents in the associated traceability evidence chain. The system arranges all expense collection records within the same primary partition in chronological order and employs a chained hash structure, where the hash digest value of the previous record is used as additional input for the hash calculation of the next record. This ensures that any modification to the historical record will cause a mismatch in the hash digest values ​​of all downstream records. The chained hash root digest values ​​of the four primary partitions, along with the project's unique code, are written into the index header block of the R&D project in the digital twin.

[0074] The encapsulated digital twin is mounted as a read-only storage volume to the data service interface provided by the system. When external audits or tax inspections need to retrieve the cost collection details of a certain R&D project, they can access the project's dedicated read-only storage volume by entering the project's unique code. They can then view all collection records, from personnel costs, direct inputs, depreciation expenses to intangible asset amortization, step by step. They can also trace each record along a bidirectional link back to the original business vouchers and the entire history of adversarial verification, forming a closed-loop traceability from financial report values ​​to physical business activities and compliance proofs.

[0075] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0076] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A big data-based financial and tax risk early warning management system for the entire manufacturing R&D process, comprising a data acquisition module, a risk indicator model, and an early warning output module. The data acquisition module is used to acquire enterprise financial data, business data, and tax data. The risk indicator model is used to perform threshold comparisons on the data. The early warning output module is used to trigger an early warning when an exceedance is detected. Its features are, The system also includes: The R&D business scenario is reconstructed into an intelligent agent, with an industrial digital twin engine and causal inference algorithm built in. It is configured to actively pull multi-source heterogeneous raw data, including time-series data of equipment IoT, material flow records, and personnel workstation sensor data. It dynamically replays the physical activity flow within the R&D cycle in virtual space and generates the first version of the business flow mapping by inferring the time-series causal chain between material consumption, equipment occupancy and personnel activities. The financial and tax regulations deconstruction and stress testing intelligent agent has a built-in knowledge computing engine that transforms the financial and tax policy clauses into a computable constraint model. It is configured to receive the first version of the business flow mapping graph and perform simulated audit path search on the nodes and edges in the mapping graph. It exhaustively lists the logical vulnerabilities and evidence chain breaks that lead to non-compliance in the allocation and generates a contradiction and conflict report carrying risk coefficients and attack paths. An adaptive adversarial evolution controller is connected to the R&D business scenario reconstruction agent and the financial and tax regulations deconstruction and stress testing agent, respectively. It is configured to project the conflict report as a challenge input to the R&D business scenario reconstruction agent, drive it to introduce implicit evidence to adjust the parameters of the causal inference model to reconstruct the business logic, and submit the reconstructed business flow map to the financial and tax regulations deconstruction and stress testing agent again. This process is repeated for multiple rounds of adversarial iteration under non-cooperative game, until the financial and tax regulations deconstruction and stress testing agent can no longer generate an effective attack path. Finally, it outputs a high-confidence business and financial and tax digital twin that has converged after the adversarial game. The risk indicator model is specifically configured to perform risk calculations based on the high-confidence business, finance, and tax digital twin.

2. The big data-based financial and tax risk early warning management system for the entire manufacturing R&D process as described in claim 1, characterized in that: When the intelligent agent reconstructing the R&D business scenario infers the temporal causal chain, it specifically executes the transfer entropy algorithm to quantify the information transmission intensity between the fluctuation of equipment operating parameters and the subsequent material consumption pulse signal, so as to identify the nonlinear driving relationship between seemingly unrelated events, thereby constructing a business flow mapping diagram containing the event context.

3. The big data-based financial and tax risk early warning management system for the entire manufacturing R&D process as described in claim 2, characterized in that: When performing simulated audit path search, the tax and financial regulations deconstruction and stress testing agent specifically adopts the Monte Carlo tree search algorithm to conduct game theory in the state space composed of all possible cost allocation schemes. For each cost allocation node in the current business flow map, it systematically explores the reverse evidence chain that can prove its improper allocation.

4. The big data-based financial and tax risk early warning management system for the entire manufacturing R&D process as described in claim 3, characterized in that: When the intelligent agent reconstructing the R&D business scenario dynamically adjusts after receiving the conflict report, the implicit evidence it introduces includes retrieving historical equipment process parameter curves of the same production equipment during the same time period and using a dynamic time warping algorithm to measure similarity, in order to verify whether the equipment operation mode during that period is closer to R&D trial production or mass production, thereby responding to the doubts of the stress test with quantitative results.

5. The big data-based financial and tax risk early warning management system for the entire manufacturing R&D process as described in claim 4, characterized in that: The adaptive adversarial evolution controller has a built-in convergence judgment unit. This unit continuously monitors the conflict reports generated by the financial and tax regulations deconstruction and the stress test agent. When the risk coefficient of the reports generated in two consecutive iterations is lower than the preset threshold, or no new attack paths are generated, the adversarial relationship is considered to have reached Nash equilibrium. At this time, the output financial and tax digital twin is the final result.

6. The big data-based financial and tax risk early warning management system for the entire manufacturing R&D process as described in claim 5, characterized in that: Each data item in the final output of the business, finance and tax digital twin of the system is accompanied by a complete traceability evidence chain generated during the adversarial game to prove its compliance. The evidence chain clearly points to the stress test and business logic reconstruction records of a specific round.

7. The big data-based financial and tax risk early warning management system for the entire manufacturing R&D process as described in claim 6, characterized in that: When collecting material flow records, the data acquisition module specifically uses radio frequency identification technology to obtain the unique code, movement trajectory and time node of the material, and uses this data as the basic spatiotemporal coordinates for the construction of physical activity flow by the intelligent agent reconstructing the R&D business scenario.

8. The big data-based financial and tax risk early warning management system for the entire manufacturing R&D process as described in claim 7, characterized in that: The early warning output module is communicatively connected to the risk closed-loop handling unit. The risk closed-loop handling unit is used to automatically generate a risk early warning form containing a description of the risk point, rectification suggestions and rectification deadline after receiving the early warning information, and to track the rectification submission vouchers and rectification progress.

9. The big data-based financial and tax risk early warning management system for the entire manufacturing R&D process as described in claim 8, characterized in that: The system also includes a policy dynamic adaptation unit, which has a built-in natural language processing engine to capture and parse the rule elements in the latest fiscal and tax policy texts in real time, and automatically update the parsed structured rules to the knowledge computing engine of the fiscal and tax regulations deconstruction and stress testing agent.

10. The big data-based financial and tax risk early warning management system for the entire manufacturing R&D process as described in claim 9, characterized in that: The aforementioned business, finance, and tax digital twin uses the unique code of the R&D project as the primary key for data association. It organizes personnel costs, direct input costs, depreciation costs, and intangible asset amortization into a standardized data system with an immutable data structure that can be traced throughout the entire chain, allowing external auditors to access the data directly.

Images