A finance and tax data automatic accounting robot system based on a memory-computing integrated chip

Abstract

This invention discloses an automated accounting robot system for financial and tax data based on an in-memory computing chip, belonging to the field of automated financial and tax data accounting. It includes an in-memory computing edge computing node, a multi-version rule hot synchronization engine, a financial and tax invoice confidence verification unit, a semantic UI self-healing execution unit, a dynamic concurrent scheduling unit, an interpretable audit and evidence storage unit, a unified data adaptation middleware, a complex financial and tax business inference engine, and an abnormal task hierarchical transfer unit. This invention reduces the computational error of the in-memory computing unit by utilizing the hardware-level error correction module built into the in-memory computing chip, combined with temperature compensation and noise suppression algorithms, achieving high-precision financial and tax-level calculations. Furthermore, by providing a dedicated financial and tax-specific in-memory computing software stack and pre-compiled operator library, it shortens algorithm adaptation time. The unified data adaptation middleware supports one-click integration with mainstream ERP systems, e-commerce platforms, and manual ledgers, thereby shortening single-customer integration time and reducing development and integration costs.

Description

Technical Field

[0001] This invention relates to the field of automatic accounting technology for financial and tax data, and in particular to an automatic accounting robot system for financial and tax data based on an in-memory computing chip. Background Technology

[0002] With the advancement of digital transformation in finance and taxation, RPA (Robotic Process Automation) technology has been widely applied in financial and tax scenarios such as invoice recognition, bookkeeping and accounting, and tax declaration, significantly improving the efficiency of financial work.

[0003] However, existing tax and financial RPA systems generally adopt a centralized cloud architecture, which leads to problems such as systemic congestion during tax collection periods due to cloud rule dependencies, the dual contradiction between security and efficiency in cloud OCR, the defects of separate processing of interface and rule changes, the lack of automation in processing cross-period contracts, and insufficient emergency audit response capabilities. The specific manifestations of these problems are as follows: The systemic congestion caused by cloud rule dependence during the tax collection period refers to the fact that all existing financial and tax RPA systems store accounting rules on cloud servers. During peak tax collection periods, tens of thousands of robots call the cloud rule interface at the same time, causing the interface response time to soar, accounting efficiency to decrease, and there is a possibility that the cloud and local robot rules are out of sync, which can easily lead to batch accounting errors. The dual contradiction between security and efficiency in cloud-based OCR refers to the fact that existing systems need to upload the original invoice images to the cloud for recognition and verification. This not only poses a significant risk of leakage of core corporate financial and tax data, but also causes network latency to cause timeouts in the processing of a single invoice, making it impossible to meet the batch processing needs of accounting firms for tens of thousands of invoices per day. The defect of separating interface and rule changes refers to the fact that the existing semantic-driven UI can only solve the positioning problem of a single interface element. When the tax platform undergoes interface layout changes and accounting rule adjustments at the same time, the scripts need to be manually repaired and the rules updated separately, which takes a long time on average and is very easy to miss the filing deadline. The lack of automation in handling cross-period contracts refers to the fact that the existing multi-version rule system can only match tax rates by date, but cannot automatically identify cross-period contracts and execute the red-ink invoice reversal and re-invoice process. It still requires manual processing by finance personnel, resulting in low efficiency and uncontrollable error rate in handling cross-period contracts. Insufficient emergency audit response capability refers to the fact that the existing cloud cache only stores single voucher data and cannot quickly link to a complete set of audit materials such as contracts, invoices, and payment vouchers. If the tax authorities make a mandatory request to provide cross-border transaction information within a short period of time, manual retrieval will be time-consuming and may result in high fines. Summary of the Invention

[0004] The purpose of this invention is to solve the problems existing in the prior art by proposing an automatic accounting robot system for financial and tax data based on an in-memory computing chip.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A robotic system for automatic accounting of financial and tax data based on an in-memory computing chip includes an in-memory computing edge computing node, a multi-version rule hot synchronization engine, a financial and tax invoice confidence verification unit, a semantic UI self-healing execution unit, a dynamic concurrent scheduling unit, an interpretable audit and evidence storage unit, a unified data adaptation middleware, a financial and tax complex business reasoning engine, and an abnormal task hierarchical transfer unit. The storage-and-computing integrated edge computing node is deployed on the invoice collection terminal and the local edge server. It has a built-in financial and tax-specific storage-and-computing integrated chip. The chip integrates a bill inference acceleration core, a three-level cold and hot data partition storage unit, a hardware-level error correction module and a homomorphic encryption computing core, which are used to complete the image preprocessing, field extraction, and local caching and encryption computing of high-frequency audit data of non-standard bills locally. The multi-version rule hot synchronization engine communicates with the storage and computing integrated edge computing node to distribute and store structured rules, such as tax type, tax rate, effective time, applicable industry, and local implementation standards, in the form of copies to the edge node, without relying on the cloud rule interface when performing accounting. The financial and tax invoice confidence verification unit is connected to the storage and computing integrated edge computing node, and is used to perform weighted comprehensive scoring based on OCR field confidence, tax amount equation verification, subject semantic matching, and multi-system cross-comparison. The semantic UI self-healing execution unit is used to automatically relocate the operation target and repair the execution script through interface semantic matching when element displacement or style change occurs on the tax platform interface. At the same time, it requests interface adaptation rules from the multi-version rule hot synchronization engine. The dynamic concurrent scheduling unit is used to adaptively adjust the number of concurrent robots based on the real-time response latency of the tax platform and the task priority. The interpretable audit evidence storage unit is used to generate a chained decision log with hashes for automated decision-making and store it on the chain at fixed intervals, while recording the intermediate result hash value of each step of the calculation inside the storage and computing chip. The unified data adaptation middleware is used to connect with traditional financial software, ERP systems, e-commerce platforms and manual ledgers to achieve standardized conversion and incremental synchronization of multi-source heterogeneous data; The complex financial and tax business reasoning engine has a built-in financial and tax knowledge graph and rule reasoning module, which is used to handle complex business scenarios such as equity incentives, mergers and acquisitions, and R&D expense capitalization. The abnormal task classification and transfer unit is used to automatically assign unprocessable tasks to financial personnel with corresponding permissions based on the task risk level and abnormality type, and push real-time warnings. The multi-version rule hot synchronization engine supports automatically matching the old and new tax rates according to the contract signing time, and automatically executes the red-ink reversal and re-invoicing logic for cross-period contracts.

[0006] As a preferred embodiment, the hardware-level error correction module of the in-memory computing chip executes the following correction algorithm: ,in, The original calculation results of the storage unit. This is the temperature compensation coefficient. This is the difference between the chip's operating temperature and the reference temperature. This is the noise suppression coefficient. This is the quantification value of environmental noise.

[0007] As a preferred embodiment, the three-level hot and cold data partition storage unit of the in-memory computing chip is divided as follows: the most recent high-frequency audit vouchers are classified as hot data and stored in the chip's internal storage area, using hardware-level encryption; the recent accounting data are classified as warm data and stored on the local server, using transparent encryption; and the long-term archived data are classified as cold data and stored in low-cost object storage, using offline encryption, and a unified retrieval interface is provided.

[0008] As a preferred embodiment, the multi-version rule hot synchronization engine performs rule matching using the following calculation method: When the document's transaction date falls into the [number]th [year], The effective range of this rule is 0; otherwise, it is 0. When both the industry label and the region label of the document belong to the first category... The rule applies to any number of rules; otherwise, the value is 0. The final matching rule index is: The old rules only marked them as expired, but did not delete them, and retained complete accounting basis and local implementation records.

[0009] As a preferred embodiment, the financial and tax invoice confidence verification unit calculates the overall confidence level of the invoice using a fixed-weighted formula: ,when When the time comes, it will automatically switch to manual review. The review results will be used to optimize the OCR model and verification rules. At the same time, the built-in data quality pre-verification engine will pre-detect the clarity, completeness and logical consistency of the original documents. Unqualified documents will be automatically returned and prompt for correction.

[0010] As a preferred embodiment, the target matching score of the semantic UI self-healing execution unit is calculated using the following formula: It takes the highest-scoring target as the operation position, and automatically triggers relocation and script self-healing when the interface changes and detects that an element is invalid, without relying on manual modification.

[0011] As a preferred embodiment, the dynamic concurrency scheduling unit performs concurrency control: The priority coefficients for audit tasks are pre-set to 1.5, regular declarations to 1.0, and archiving tasks to 0.5. When the tax platform's response exceeds the threshold, the concurrent tasks are automatically reduced and the tasks are executed in a staggered manner.

[0012] As a preferred option, the interpretable audit evidence storage unit is authorized according to risk level: through Calculate the decision risk value, where, To normalize transaction amounts, Furthermore, a confidence level ≥ 0.98 indicates fully automatic execution, while a confidence level ≤ 0.3 indicates automatic execution. <0.7 indicates mandatory manual review. A score of ≥0.7 indicates that only supplementary suggestions are provided and final review is conducted manually.

[0013] As a preferred embodiment, the unified data adaptation middleware includes a multi-source data acquisition module, a standardized conversion engine, and an incremental synchronization module. The multi-source data acquisition module supports integration with mainstream ERP systems, e-commerce platforms, Excel ledgers, and manual vouchers. The standardized conversion engine converts data of different formats into unified structured data based on the financial and tax data meta-standard. The incremental synchronization module uses CDC technology to achieve real-time incremental data synchronization.

[0014] As a preferred embodiment, the complex financial and tax business reasoning engine includes a financial and tax knowledge graph, a rule reasoning module, and a manual annotation feedback module. The financial and tax knowledge graph contains accounting standards, tax law provisions, local implementation guidelines, and typical cases. The rule reasoning module supports IF-THEN rules and fuzzy reasoning to handle business scenarios in areas of policy ambiguity. The manual annotation feedback module automatically updates complex business cases processed manually to the knowledge graph, continuously optimizing reasoning capabilities.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention reduces the calculation error of the in-memory computing unit to below the threshold by using a hardware-level error correction module built into the in-memory computing chip, combined with temperature compensation and noise suppression algorithms, thereby achieving high-precision calculations for financial and tax purposes. Furthermore, by providing a dedicated in-memory computing software stack for financial and tax purposes and a pre-compiled operator library, it shortens the algorithm adaptation time and unifies the data adaptation middleware to support one-click integration with mainstream ERP, e-commerce platforms and manual ledgers, thereby shortening the integration time for a single customer and reducing development and integration costs.

[0016] 2. This invention achieves end-to-end traceability from the original invoice to the final declaration by recording the intermediate result hash value of each step of the calculation inside the storage and computing chip and storing it on the blockchain. The chip has a built-in homomorphic encryption computing core and memory isolation mechanism to achieve privacy calculation without data leaving the local machine, thereby reducing the risk of data leakage and achieving end-to-end traceability and high security.

[0017] 3. This invention sets up an automatic hierarchical transfer and real-time warning mechanism for abnormal tasks, and automatically identifies and returns unqualified documents through a data quality pre-verification engine, reducing subsequent error handling costs and thus improving the human-machine collaboration mechanism. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the financial and tax data accounting process of an automated financial and tax data accounting robot system based on an in-memory computing chip, as proposed in this invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0021] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0022] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0023] Example, refer to Figure 1 A system for automatic accounting of financial and tax data based on an in-memory computing chip includes an in-memory computing edge computing node, a multi-version rule hot synchronization engine, a financial and tax invoice confidence verification unit, a semantic UI self-healing execution unit, a dynamic concurrent scheduling unit, an interpretable audit and evidence storage unit, a unified data adaptation middleware, a financial and tax complex business reasoning engine, an abnormal task hierarchical transfer unit, and a multi-version rule hot synchronization engine. The in-store computing edge computing node is deployed on the invoice collection terminal and the local edge server. It has a built-in financial and tax-specific in-store computing chip. The chip integrates a bill inference acceleration core, a three-level cold and hot data partition storage unit, a hardware-level error correction module and a homomorphic encryption computing core. It is used to complete the image preprocessing, field extraction, and local caching and encryption computing of high-frequency audit data of non-standard bills locally. When performing accounting, the in-store computing edge computing node prioritizes the use of local rule copies and only requests a full update from the cloud when the local rules are missing or expired. Furthermore, the hardware-level error correction module of the in-memory computing chip executes the following correction algorithm: ,in, The result of the corrected in-store calculation unit (used for high-precision calculations of tax amount, amount, etc., with a value range of [0,1]); The original calculation results (within the range of [0,1]) of the memory computing unit (memristor cross array output) can be directly read from the memristor cross array output terminal of the memory computing chip and converted into digital quantities by ADC; This is the temperature compensation coefficient (value 0.0002 / ℃). The difference between the chip's operating temperature and the reference temperature (within the range of [−35∘C, +35∘C]) is collected in real time by the chip's built-in digital temperature sensor at a sampling frequency of 1Hz. This is the noise suppression coefficient (value 0.05), and and The chips are calibrated and cured in high and low temperature chambers and electromagnetic shielding rooms before leaving the factory, and support remote fine-tuning via the cloud. The environmental noise quantification value (within the range of [0,1]) is collected by the noise detection circuit built into the chip and updated every 10 seconds. After correction, the calculation error of the storage unit is ≤0.001%, which meets the accuracy requirements of financial and tax accounting. The multi-version rule hot synchronization engine communicates with the storage and computing integrated edge computing node to distribute and store the structured rules, such as tax type, tax rate, effective time, applicable industry, and local implementation standards, in the form of copies to the edge node. When performing accounting, it does not rely on the cloud rule interface. The multi-version rule hot synchronization engine supports automatically matching the old and new tax rates according to the contract signing time, and automatically executes the red ink reversal and re-invoicing logic for cross-period contracts. Furthermore, the three-level hot and cold data partition storage unit of the in-memory computing chip is divided as follows: the most recent (e.g., the last 3 years) high-frequency audit vouchers are classified as hot data and stored in the chip's internal storage area, using hardware-level encryption; the recent (e.g., the last 3-5 years) accounting data are classified as warm data and stored on the local server, using transparent encryption; and the long-term (more than 5 years) archived data are classified as cold data and stored in low-cost object storage, using offline encryption, and a unified retrieval interface is provided, with a hot data retrieval response latency of ≤50ms. Furthermore, the multi-version rule hot synchronization engine performs rule matching using the following calculation method: When the document's transaction date falls into the [number]th [year], The effective range of this rule is 0; otherwise, it is 0. When both the industry label and the region label of the document belong to the first category... The rule applies to any number of rules; otherwise, the value is 0. The final matching rule index is: The old rules are only marked as expired, not deleted, and the complete accounting basis and local implementation standards are retained. in, The final matching rule index points to the unique identifier of the tax and finance rule with the highest matching degree; For the first The time matching value of the rule is used to determine whether the document's business date falls within the effective period of the rule. This is done by comparing the document's business date with the first rule. The start and end fields of the rule are set to 1 if they match, and 0 otherwise. For the first The range matching value of the rule is used to simultaneously verify the applicability of both industry and region dimensions by comparing the industry label of the document with the first rule. The list of rules covers the scope of each rule, and compares the administrative division code of the document with the first rule. The list of administrative divisions that meet the rules is 1 if both rules are met, otherwise it is 0. For the first The priority of a rule, with higher values ​​indicating higher priority. It is used to resolve rule conflicts and is read from the structured rule table in the cloud rule configuration center. It is set by the administrator when adding / modifying a rule, with a default value of 1. The confidence verification unit for financial and tax invoices is connected to the storage and computing integrated edge computing node, and is used to perform weighted comprehensive scoring based on OCR field confidence, tax amount equation verification, subject semantic matching, and cross-comparison of multiple systems. Furthermore, the tax and financial invoice confidence verification unit calculates the overall confidence level of invoices using a fixed-weighted formula: ,when When the time comes, it will automatically switch to manual review. The review results will be used to optimize the OCR model and verification rules. At the same time, the built-in data quality pre-verification engine will pre-detect the clarity, completeness and logical consistency of the original documents. Unqualified documents will be automatically returned and prompt for correction. in, The overall confidence level of the invoice is determined by comprehensively evaluating the credibility of the invoice identification and verification results, deciding whether to automatically pass, and calculating the scores of four dimensions using a fixed-weighted formula. The average confidence score of the key OCR fields reflects the overall accuracy of document text recognition. It is obtained by calculating the arithmetic mean of the confidence scores of the eight core key fields (amount, tax rate, tax number, invoice number, invoice date, buyer's name, seller's name, and tax amount) in the OCR recognition output. The rule verification score reflects the degree to which the invoice passes the tax compliance rule verification. It is calculated based on the number and importance of the rules passed by the invoice: 1.0 for all rules passed; 0.5 points for incorrect tax amount equation; 0.3 points for incorrect invoice format; and points are deducted proportionally for other errors. The semantic matching score reflects the semantic similarity between the invoice content and the corresponding accounting subject. The cosine similarity between the invoice content text and the accounting subject text is calculated by a pre-trained BERT model specifically for finance and taxation (fine-tuned with 1 million pieces of finance and taxation text). The score is calculated based on the cross-validation of multiple systems, reflecting the consistency between invoice information and multi-source business data. It is calculated based on the number of matching items between invoice information and ERP orders, contracts, and bank statements: 1.0 for all matching items; 0.25 points are deducted for each missing matching item.

[0024] The semantic UI self-healing execution unit is used to automatically relocate the operation target and repair the execution script when element displacement or style change occurs on the tax platform interface. At the same time, it requests interface adaptation rules from the multi-version rule hot synchronization engine. Furthermore, the target matching score of the semantic UI self-healing execution unit is calculated using the following formula: It takes the highest-scoring target as the operation position. When the interface changes and detects that an element is invalid, it automatically triggers relocation and script self-healing without relying on manual modification. in, For the first The semantic similarity between the element text and the task instructions reflects the degree of matching between the element's function and the task requirements. The cosine similarity between the two is calculated by a pre-trained multimodal UI semantic model (fine-tuned from 500,000 pieces of tax platform interface text). For the first The matching degree between the element type and the required type is used to filter elements of non-target type. It compares the type of the element (button, input box, drop-down box, etc.) with the type required by the task. If they match, it is 1.0, otherwise it is 0.

[0025] The dynamic concurrent scheduling unit is used to adaptively adjust the number of concurrent robots based on the real-time response latency of the tax platform and task priority; Furthermore, the dynamic concurrency scheduling unit performs concurrency control: The priority coefficients for audit tasks are pre-set to 1.5, regular declarations to 1.0, and archiving tasks to 0.5. When the tax platform's response exceeds the threshold, the system automatically reduces concurrency and executes tasks at off-peak times. in, The current concurrency level is the number of robot tasks that the system currently allows to execute simultaneously, calculated using a dynamic concurrency control formula. The maximum number of concurrent tasks supported by the system hardware is the upper limit, which is preset based on the computing power of the edge node's CPU, memory, and in-memory computing chip. The default value for a single edge node is 150. The baseline concurrency is the default concurrency when the tax platform responds normally. It is preset based on historical operating data and hardware performance, with a default value of 80 for a single edge node. The target response time is the system's expected average response time for the tax platform interface, expressed in seconds. It is preset based on the timeliness requirements of tax declaration business, with a default value of 10 seconds. The real-time average response time of the tax platform is the sliding average of the interface response time of all tasks in the most recent minute, in seconds. It is calculated by monitoring the interface call logs of all tasks in real time and calculating the sliding window average. This is the task priority coefficient, which determines the weight of concurrent resources received by different types of tasks. The higher the value, the higher the priority. Preset values ​​are: Inspection task 1.5, Regular declaration 1.0, Data archiving 0.5.

[0026] The interpretable audit evidence storage unit is used to generate chained decision logs with hashes for automated decision-making and store them on the chain at fixed intervals, while also recording the hash values ​​of intermediate results of each step of the calculation inside the storage and computing chip. Furthermore, it can be explained that the audit evidence storage unit is authorized according to risk level: through Calculate the decision risk value, where, To normalize transaction amounts, A value <0.3 and a confidence level ≥0.98 indicate fully automatic execution; a value ≤0.3 indicates automatic execution. <0.7 indicates mandatory manual review. ≥0.7 indicates that only auxiliary suggestions are provided and final review is conducted manually. The on-chain storage period for all decision logs and intermediate calculation results hash values ​​of the storage unit is 10 minutes. in, The decision risk value, the degree of risk in automated decision-making, and the authorization level are determined by a weighted formula that calculates scores across three dimensions. To normalize transaction amounts, transaction amounts are mapped to the [0,1] range to prevent large transactions from excessively affecting the risk value. This is achieved through calculation. The benchmark amount of 100,000 yuan is the common risk threshold in the finance and taxation industry; The transaction amount is the actual transaction amount corresponding to the bill or business, in yuan, extracted from the bill recognition results or business system data. The tax risk coefficient represents the inherent tax risk level corresponding to this business type, with preset values ​​read from the tax knowledge graph: VAT special invoice 0.8, export tax refund business 0.9, ordinary invoice 0.3, and travel expense reimbursement 0.2. The historical error rate is the processing error rate of this type of business over the past 3 months, obtained by comparing the number of errors in this type of business with the total number of processing operations in the historical processing records of the statistical system.

[0027] The unified data adaptation middleware is used to connect with traditional financial software, ERP systems, e-commerce platforms and manual ledgers to achieve standardized conversion and incremental synchronization of multi-source heterogeneous data; Furthermore, the unified data adaptation middleware includes a multi-source data acquisition module, a standardized conversion engine, and an incremental synchronization module. The multi-source data acquisition module supports integration with six mainstream ERP systems, including Yonyou and Kingdee, three e-commerce platforms, including Taobao and JD.com, as well as Excel ledgers and manual vouchers. The standardized conversion engine converts data of different formats into unified structured data based on the financial and tax data meta standard. The incremental synchronization module uses CDC technology to achieve real-time incremental data synchronization with a synchronization delay of ≤5 minutes. The complex financial and tax business reasoning engine has a built-in financial and tax knowledge graph and rule reasoning module, which is used to handle complex business scenarios such as equity incentives, mergers and acquisitions, and R&D expense capitalization. Furthermore, the complex business reasoning engine for finance and taxation includes a finance and taxation knowledge graph, a rule reasoning module, and a manual annotation feedback module. The finance and taxation knowledge graph contains accounting standards, tax law provisions, local implementation guidelines, and typical cases. The rule reasoning module supports IF-THEN rules and fuzzy reasoning to handle business scenarios in areas of policy ambiguity. The manual annotation feedback module automatically updates complex business cases processed manually to the knowledge graph, continuously optimizing reasoning capabilities. The abnormal task classification and transfer unit is used to automatically assign unprocessable tasks to finance personnel with corresponding permissions based on the task risk level and abnormality type, and push real-time warnings. Furthermore, the abnormal task classification and transfer unit executes the following process: abnormal tasks are divided into Level 1 (system self-healing), Level 2 (handled by ordinary financial personnel), Level 3 (handled by financial supervisors), and Level 4 (handled by expert committee). Level 1 exceptions will automatically retry 3 times; if they fail, they will be upgraded to Level 2. If a Level 2 anomaly is not resolved within 1 hour, it will be escalated to Level 3. If a Level 3 anomaly is not addressed within 4 hours, it will be escalated to Level 4 and a warning will be sent to the responsible person.

[0028] The automatic accounting robot system for financial and tax data of this invention also includes a multi-tenant AI automatic configuration unit: based on the customer's historical financial data for 3 months, it automatically identifies the data source format, industry type, account mapping and tax rate strategy, generates a configuration template that can be directly used, without the need to write RPA scripts, provides a low-code configuration platform, and supports customers to customize business rules and report formats; This invention adopts an edge-cloud hybrid storage and computing architecture, deploying lightweight storage and computing edge chips in small and medium-sized enterprises to complete invoice recognition, local accounting, and data encryption; it provides computing power scheduling, rule updates, model training, and big data analysis services through the cloud; and it supports progressive deployment, allowing customers to gradually upgrade the computing power of edge nodes according to their needs.

[0029] The following is an example: On May 1, 2025, the value-added tax rate for manufacturing in a certain province was adjusted from 13% to 11%, and it was also clarified that small and micro enterprises in the province could enjoy a 1% local surtax reduction policy. The system performs parallel accounting according to the following process: The administrator adds a new rule R-VAT-2025-11 in the cloud rule configuration center, sets the effective start date to 2025-05-01, the effective end date to 2099-12-31, the applicable scope ["Manufacturing", "Wholesale and Retail"], the applicable region ["A certain province"], the local surtax reduction rate to 1%, and the priority to 1. The multi-version rule hot synchronization engine distributes the rule in the form of a copy to the local rule library of all storage and computing edge nodes within 5 minutes, completing the steps of rule cloud configuration and edge synchronization. When the robot processes documents, it directly calls the local rule copy of the storage-compute integrated edge node, without accessing the cloud rule interface, and executes the rule matching algorithm: ; Regarding the manufacturing contract signed on April 20, 2025, in a certain province, Matching the old tax rate of 13% with the original local surtax; for a small and micro enterprise contract signed on May 5, 2025, in a certain province, The steps of matching the new 11% tax rate with the 1% local surtax reduction are completed to fulfill the local rule matching and accounting requirements. When the system detects that the transaction date of the document is April 25, 2025 (before the policy takes effect) and the invoice date is May 2, 2025 (after the policy takes effect), it automatically identifies it as a cross-period contract, generates a red-ink invoice to reverse the original withheld tax of 1,300 yuan and local surtax of 130 yuan, and then generates a blue-ink invoice at a tax rate of 11% with a tax amount of 1,100 yuan and a local surtax of 110 yuan. At the same time, it updates the tax information in the contract ledger and completes the automatic processing steps for cross-period contracts. The old rule R-VAT-2025-13 will be marked as "expired" and fully retained in the local rule base, including records of the local enforcement guidelines at the time, to meet audit traceability requirements and complete the historical rule retention step.

[0030] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A robotic system for automatic accounting of financial and tax data based on an in-memory computing chip, characterized in that, It includes in-store computing edge computing nodes, multi-version rule hot synchronization engine, financial and tax invoice confidence verification unit, semantic UI self-healing execution unit, dynamic concurrency scheduling unit, interpretable audit and evidence storage unit, unified data adaptation middleware, financial and tax complex business reasoning engine, and abnormal task hierarchical flow unit. The storage-and-computing integrated edge computing node is deployed on the invoice collection terminal and the local edge server. It has a built-in financial and tax-specific storage-and-computing integrated chip. The chip integrates a bill inference acceleration core, a three-level cold and hot data partition storage unit, a hardware-level error correction module and a homomorphic encryption computing core, which are used to complete the image preprocessing, field extraction, and local caching and encryption computing of high-frequency audit data of non-standard bills locally. The multi-version rule hot synchronization engine communicates with the storage and computing integrated edge computing node to distribute and store structured rules, such as tax type, tax rate, effective time, applicable industry, and local implementation standards, in the form of copies to the edge node. When performing accounting, it does not rely on the cloud rule interface. The multi-version rule hot synchronization engine supports automatically matching the old and new tax rates according to the contract signing time and automatically executes the red ink reversal and re-invoicing logic for cross-period contracts. The financial and tax invoice confidence verification unit is connected to the storage and computing integrated edge computing node, and is used to perform weighted comprehensive scoring based on OCR field confidence, tax amount equation verification, subject semantic matching, and multi-system cross-comparison. The semantic UI self-healing execution unit is used to automatically relocate the operation target and repair the execution script through interface semantic matching when element displacement or style change occurs on the tax platform interface. At the same time, it requests interface adaptation rules from the multi-version rule hot synchronization engine. The dynamic concurrent scheduling unit is used to adaptively adjust the number of concurrent robots based on the real-time response latency of the tax platform and the task priority. The interpretable audit evidence storage unit is used to generate a chained decision log with hashes for automated decision-making and store it on the chain at fixed intervals, while recording the intermediate result hash value of each step of the calculation inside the storage and computing chip. The unified data adaptation middleware is used to connect with traditional financial software, ERP systems, e-commerce platforms and manual ledgers to achieve standardized conversion and incremental synchronization of multi-source heterogeneous data; The complex financial and tax business reasoning engine has a built-in financial and tax knowledge graph and rule reasoning module, which is used to handle complex business scenarios such as equity incentives, mergers and acquisitions, and R&D expense capitalization. The abnormal task classification and transfer unit is used to automatically assign unprocessable tasks to financial personnel with corresponding permissions based on the task risk level and abnormality type, and push real-time warnings.

2. The automated accounting robot system for financial and tax data based on an in-memory computing chip according to claim 1, characterized in that, The hardware-level error correction module of the in-memory computing chip executes the following correction algorithm: ,in, The original calculation results of the storage unit. This is the temperature compensation coefficient. This is the difference between the chip's operating temperature and the reference temperature. This is the noise suppression coefficient. This is the quantification value of environmental noise.

3. The automated accounting robot system for financial and tax data based on an in-memory computing chip according to claim 1, characterized in that, The three-level hot and cold data partition storage unit of the in-memory computing chip is divided as follows: the most recent high-frequency audit vouchers are classified as hot data and stored in the chip's internal storage area, using hardware-level encryption; recent accounting data are classified as warm data and stored on the local server, using transparent encryption; long-term archived data are classified as cold data and stored in low-cost object storage, using offline encryption, and a unified retrieval interface is provided.

4. The automated accounting robot system for financial and tax data based on an in-memory computing chip according to claim 1, characterized in that, The multi-version rule hot synchronization engine performs rule matching using the following calculation method: When the document's transaction date falls into the [number]th [year], The effective range of this rule is 0; otherwise, it is 0. When both the industry label and the region label of the document belong to the first category... The rule applies to any number of rules; otherwise, the value is 0. The final matching rule index is: The old rules only marked them as expired, but did not delete them, and retained complete accounting basis and local implementation records.

5. The automated accounting robot system for financial and tax data based on an in-memory computing chip according to claim 1, characterized in that, The tax and financial invoice confidence verification unit calculates the overall confidence level of the invoice using a fixed-weighted formula: ,when When the time comes, it will automatically switch to manual review. The review results will be used to optimize the OCR model and verification rules. At the same time, the built-in data quality pre-verification engine will pre-detect the clarity, completeness and logical consistency of the original documents. Unqualified documents will be automatically returned and prompt for correction.

6. The automated accounting robot system for financial and tax data based on an in-memory computing chip according to claim 1, characterized in that, The target matching score of the semantic UI self-healing execution unit is calculated using the following formula: The system takes the highest-scoring target as the operation position. When the interface changes and detects that an element is invalid, it automatically triggers relocation and script self-healing without relying on manual modification.

7. The automated accounting robot system for financial and tax data based on an in-memory computing chip according to claim 1, characterized in that, The dynamic concurrency scheduling unit performs concurrency control: The priority coefficients for audit tasks are pre-set to 1.5, regular declarations to 1.0, and archiving tasks to 0.

5. When the tax platform's response exceeds the threshold, the concurrent processing is automatically reduced and the tasks are executed in staggered shifts.

8. The automated accounting robot system for financial and tax data based on an in-memory computing chip according to claim 1, characterized in that, The interpretable audit evidence storage unit is authorized according to risk level: through Calculate the decision risk value, where, To normalize transaction amounts, Furthermore, a confidence level ≥ 0.98 indicates fully automatic execution. To enforce manual review, Only supplementary suggestions will be provided; final review will be conducted manually.

9. The automated accounting robot system for financial and tax data based on an in-memory computing chip according to claim 1, characterized in that, The unified data adaptation middleware includes a multi-source data acquisition module, a standardized conversion engine, and an incremental synchronization module. The multi-source data acquisition module supports integration with mainstream ERP systems, e-commerce platforms, Excel ledgers, and manual vouchers. The standardized conversion engine converts data of different formats into unified structured data based on the financial and tax data meta-standard. The incremental synchronization module uses CDC technology to achieve real-time incremental data synchronization.

10. The automated accounting robot system for financial and tax data based on an in-memory computing chip according to claim 1, characterized in that, The complex financial and tax business reasoning engine includes a financial and tax knowledge graph, a rule reasoning module, and a manual annotation feedback module. The financial and tax knowledge graph contains accounting standards, tax law provisions, local implementation guidelines, and typical cases. The rule reasoning module supports IF-THEN rules and fuzzy reasoning to handle business scenarios in areas of policy ambiguity. The manual annotation feedback module automatically updates complex business cases processed manually to the knowledge graph, continuously optimizing reasoning capabilities.

Images