A Customs Compliance Intelligent Audit System and Method

Abstract

This invention belongs to the field of customs auditing, and particularly relates to a customs compliance intelligent auditing system and method, comprising: a data acquisition module that generates an audit rule update table and a customs declaration business information table in real time through a multimodal extraction algorithm; a compliance auditing module that performs multi-dimensional compliance verification based on a horizontal business audit node network, a vertical business transmission node chain, and a declaration node network, combined with a dynamic rule priority index and a reinforcement learning algorithm; and an evaluation and optimization module that calculates the probability of business transmission risk, audit delay probability, and error probability in real time through a two-way risk assessment model, and feeds it back to the audit space to optimize rule weights and process paths. This invention, through a composite auditing architecture of "declaration clustering + vertical encrypted transmission + horizontal topology association," achieves dynamic adaptation of audit rules, accurate risk tracing, and intelligent process optimization, effectively solving the problems of inefficient integration of multi-source heterogeneous data and lagging auditing in complex business scenarios in cross-border trade.

Description

Technical Field

[0001] This invention belongs to the field of customs auditing, and in particular relates to an intelligent customs compliance auditing system and method. Background Technology

[0002] Currently, customs declaration compliance review faces systemic challenges such as low efficiency in integrating multi-source heterogeneous data, weak dynamic rule adaptation capabilities, and lagging traceability of hidden risks. Traditional review systems rely on manual input and static rule engines, making it difficult to handle complex scenarios such as cross-border e-commerce and multi-country transshipment, resulting in excessively long review times per shipment and a significantly increased rate of missed hidden violations. Although blockchain technology can improve data credibility, the insufficient throughput of public blockchains and the lack of deep collaboration with large-scale models in vertical domains limit their application value in real-time review scenarios. Existing knowledge graph technologies are mostly limited to static rule mapping and cannot achieve dynamic reasoning, while cross-source alignment and real-time indexing technologies for text, images, and IoT sensor data are not yet mature, causing biases in the extraction of key information and review delays. In addition, problems such as rule updates relying on manual configuration and the lack of causal tracing in risk warnings further restrict the accuracy and agility of automated review; therefore, this application proposes a customs compliance intelligent review system and method. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a customs compliance intelligent audit system and system, which includes a data acquisition module, a compliance audit module, and an evaluation and optimization module. The data acquisition module uses a multimodal extraction algorithm to synchronize customs review rules and declaration information in real time, generating a timestamp-aligned audit rule update table and a declaration business information table. The compliance audit module constructs an enhanced audit space, integrating a three-tiered architecture of a horizontal business audit node network, a vertical business transmission node chain, and a declaration node network. The horizontal business audit node network establishes the topological relationships between business audit nodes based on graph algorithms, while the vertical business transmission node chain uses distributed reinforcement learning and encryption algorithms to construct an encrypted customs declaration process. The invention employs a dense transmission link and a declaration node network that uses business type similarity clustering to achieve batch processing of similar declarations. These three elements work together, combining dynamic rule priority indexing and reinforcement learning strategies to complete multi-dimensional compliance verification. The evaluation and optimization module uses a two-way risk assessment model to calculate the probability of business transmission risks, the probability of review delays, and the probability of errors in real time, and feeds this information back to the review space to optimize rule weights and process paths. This invention breaks through the traditional single-link review mode through a composite review architecture of "horizontal topology-vertical transmission-declaration clustering," achieving dynamic rule adaptation, accurate risk tracing, and intelligent process optimization. It effectively solves the problems of inefficient integration of multi-source heterogeneous data and delayed review in complex business scenarios in cross-border trade.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A customs compliance intelligent audit system includes: a data acquisition module, a compliance audit module, and an evaluation and optimization module;

[0006] The data acquisition module, based on real-time collected review rules and customs declaration information combined with a multimodal extraction model, extracts the review rule update table and customs declaration business information table in real time.

[0007] The compliance review module allows users to obtain a completed customs declaration business information table based on the customs declaration classification information table, through a preset enhanced review space and a review rule update table. The enhanced review space includes a declaration node network, a vertical business transmission node chain, and a horizontal business review node network. This enhanced review space is constructed and trained using reinforcement learning algorithms, combining business information from the horizontal business review node network and the vertical business transmission node chain, preset review rule indexes, and the priorities corresponding to the rules and business reviews. The vertical business transmission node chain is constructed using distributed reinforcement learning algorithms and encryption algorithms, combining each declaration node, its corresponding business review node, and the customs declaration business information transmission path. The horizontal business review node network is constructed using graph algorithms, combining the correlation strength between all business review nodes. The declaration node network is constructed by combining the similarity of all business declaration nodes with their corresponding business declaration types. The declaration node network is connected to the horizontal business review node network through the vertical business transmission node chain.

[0008] The evaluation and optimization module, based on a preset two-way risk assessment model and combined with the real-time business review information corresponding to the horizontal business review node network and the vertical business transmission node chain, obtains the probability of business review transmission risk, the probability of review delay, and the probability of review error, and feeds them back to the enhanced review space to adjust the business transmission and review process.

[0009] Specifically, the data acquisition module includes a data acquisition unit, a multimodal extraction unit, and an enhancement and update unit;

[0010] The data acquisition unit is used to collect in real time all business declaration nodes' uploaded business customs declaration information, business customs declaration information whose review process is not completed, and updated business review rule information, and classify the collected data information according to data type;

[0011] The multimodal extraction unit obtains a customs declaration business information table and an audit rule update table based on the real-time collected information after classification and a multimodal feature extraction algorithm. It then performs a second audit update of incomplete declaration business based on the effective timestamp of the audit rule in the audit rule update table.

[0012] The enhanced update unit is used to update the review rule information corresponding to each node in the application node network and the horizontal business review node network according to the review rule update table and the preset enhanced update model, and to use the updated review rules to perform secondary review updates for incomplete application business reviews.

[0013] Specifically, the compliance review module includes an application clustering unit, an initial review unit, a transmission unit, and a secondary review unit;

[0014] The application clustering unit obtains a set of similar application nodes based on the business information corresponding to all business application nodes and a business similarity clustering model, and uses the cluster center corresponding to each subset of similar application nodes in the set of similar application nodes as the initial review node.

[0015] The initial review unit obtains the customs declaration business information table after initial review based on the first enhanced review model configured for each initial review node, combined with the customs declaration business information table and corresponding review rules in the corresponding similar declaration node subset.

[0016] The transmission unit, based on the customs declaration business information table after review corresponding to each initial review node, combined with the preset encryption code, the load information of the transmission channel corresponding to each vertical business transmission node chain, the review business type information and load information corresponding to all the business review nodes, and the initial review evaluation result information corresponding to each declaration node, transmits and distributes the customs declaration business information table after initial review to the corresponding business review node through a transmission path optimization algorithm combined with a distributed reinforcement allocation algorithm.

[0017] The secondary review unit obtains the customs declaration business information table after secondary review based on the priority of the review business corresponding to each business review node, the priority of the sub-business review process corresponding to each business, the index information between the review rules and each sub-business review process, and the second enhanced review model configured for each business review node.

[0018] Specifically, the evaluation and optimization module includes a first evaluation unit, a second evaluation unit, and an evaluation feedback unit;

[0019] The first evaluation unit is used to perform real-time evaluation through the first evaluation sub-model based on the clustering process corresponding to the application clustering unit, the initial review process corresponding to the initial review unit, and the supply-demand relationship stability coefficient, to obtain the first review risk probability; the first review risk probability is constructed from the clustering error probability and the initial false detection and false detection risk probability; the supply-demand relationship stability coefficient is constructed from the ratio of the transaction completion frequency of the supply and demand parties to the total transaction frequency;

[0020] The second evaluation unit is used to perform real-time evaluation based on the data transmission process corresponding to the transmission unit and the secondary review process and review anomaly coefficient corresponding to the secondary review unit, and obtain the second review risk probability through the second evaluation sub-model. The second review risk probability is constructed by combining the tampering risk probability and business allocation risk probability corresponding to the data transmission process and the risk probabilities corresponding to false detection, false omission and review delay corresponding to the secondary review with the Bayesian algorithm.

[0021] The audit anomaly coefficient is constructed from the historical anomaly frequency of the corresponding applicant and the total application frequency;

[0022] The evaluation feedback unit optimizes the first enhanced audit model, the second enhanced audit model, and the encrypted information by providing graded feedback based on the risk probability difference between the first audit risk probability and the second audit risk probability, combined with preset multi-level risk discrimination thresholds.

[0023] Specifically, the workflow of the evaluation feedback unit includes:

[0024] Preset first review threshold First review difference judgment threshold Second review difference threshold ,and When the probability of risk in the first review is less than If the initial audit result is positive, the customs declaration information table will be transmitted and distributed to the corresponding audit node. Otherwise, the first audit risk probability will be fed back to the first enhanced audit model to update and optimize the clustering process and the initial audit process until the first audit risk probability is less than 1%. until;

[0025] When the risk probability difference is less than If the second review is completed, a customs declaration information table will be output. If the risk probability difference is greater than or equal to... and less than If the second audit risk probability is not met, the data transmission process and the secondary audit process will be optimized and updated by feeding back the second audit risk probability to the transmission unit and the second enhanced audit model until the risk probability difference is less than 1. Until then;

[0026] When the risk probability difference is greater than or equal to If an encryption anomaly is detected, the business information corresponding to the business review node is compared with the business information of the corresponding application node to obtain comparison difference information. Simultaneously, the encryption encoding level in the transmission unit and the business information indicating encryption anomaly are adjusted using the obtained comparison difference information. The adjusted business information undergoes a second update review until the risk probability difference is less than [a certain value]. Until then.

[0027] Specifically, the process of obtaining the customs declaration business information form after initial review includes:

[0028] Each user who submits a business application is treated as a business application node. Based on the business type information of the user and the corresponding target customer information, a similarity algorithm is used to obtain the business similarity of different users who submit applications at the same timestamp.

[0029] Based on the similarity between business application nodes and their corresponding businesses, a network of application nodes is constructed using a graph algorithm.

[0030] Based on the real-time collection of business customs declaration information of different declaring users under the same timestamp by the declaration node network, the collected business customs declaration information is input into the multimodal extraction model to obtain the customs declaration business information table corresponding to different declaring users;

[0031] The customs declaration business information table includes a unique identifier, business type, and structured business process information after feature extraction;

[0032] Based on the business information and business similarity corresponding to the application node network, clustering is performed by combining clustering algorithm with time series algorithm to obtain a subset of similar application nodes corresponding to each similar business type, and the cluster center node of the subset of similar application nodes is used as the initial review node of the corresponding subset of similar application nodes.

[0033] Specifically, the process of obtaining the customs declaration business information form after the initial review also includes:

[0034] The first enhanced review model is constructed by using a federated algorithm framework based on the first enhanced review sub-model configured for each application node in each subset of similar application nodes and the first evaluation sub-model configured for the corresponding initial review node.

[0035] Based on the customs declaration business information table corresponding to each declaration node, the corresponding first enhanced audit sub-model is used for auditing, and the audit result information is transmitted to the corresponding initial audit node through the federated framework. The first evaluation sub-model is configured for auditing and evaluation. If the evaluation is passed, the customs declaration business information table of the initial audit is output.

[0036] If the review fails, the assessment results will be fed back to the corresponding declaration node through the federal framework for adjustment until the first review threshold is met, at which point the adjusted initial review completion customs declaration information table will be obtained.

[0037] Specifically, the process of obtaining the customs declaration information form after the second review includes:

[0038] Each department of the reviewing party for each business type is treated as a business review node. The review process of the corresponding business type under each business review node and the review priority of each review sub-node in the review process under the current business type review process are used to construct a business review process node chain corresponding to each business review node through a blockchain algorithm. Each review sub-node corresponds one-to-one with each review sub-process in the corresponding business type review process.

[0039] Based on the business review process node chain information corresponding to each business review node, the business overlap correlation degree corresponding to each business review node is obtained through the association algorithm.

[0040] Based on each business review node and its corresponding business overlap correlation, a horizontal business review node network is constructed using a graph algorithm.

[0041] Specifically, the process of obtaining the customs declaration information form after the second review also includes:

[0042] The horizontal business audit node network obtains the audit results and the comprehensive results of the corresponding business by means of the customs declaration business information table completed by the initial audit and the corresponding audit priority and initial audit evaluation results allocated by the vertical business transmission node chain, the audit rule index information corresponding to the audit sub-process stored by each audit sub-node under each business audit node, and the second enhanced audit sub-model configured by each audit sub-node.

[0043] Based on the comprehensive results of each business audit node and the corresponding initial business audit results, combined with the first audit difference judgment threshold, the second audit difference judgment threshold and the second evaluation sub-model, a second audit judgment is made. If the judgment is passed, a customs declaration business information table with the second audit completed is obtained.

[0044] Otherwise, the corresponding business, initial review process, transmission process and secondary review process will be adjusted in real time based on the results of the second review until all review sub-processes of each review business meet the review rules.

[0045] The second enhanced audit model is constructed by the second enhanced audit sub-model configured for each audit sub-node and the corresponding second evaluation sub-model through a federated framework.

[0046] A smart auditing method for customs compliance includes:

[0047] Based on real-time collected review rules and customs declaration information, combined with a multimodal extraction algorithm, an review rule update table and a customs declaration business information table are obtained.

[0048] Based on the customs declaration classification information table, the table is updated by combining the preset enhanced review space with the review rules to obtain the customs declaration business information table after review.

[0049] The enhanced review space comprises a horizontal business review node network and a vertical business transmission node chain. This enhanced review space is constructed and trained using reinforcement learning algorithms, combining business information from the horizontal business review node network and the vertical business transmission node chain, preset review rule indexes, and the priorities corresponding to the rules and business reviews. The vertical business transmission node chain is constructed using distributed reinforcement learning algorithms and encryption algorithms, combining each declaration node, its corresponding business review node, and the customs declaration business information transmission path. The horizontal business review node network is constructed using graph algorithms, combining the correlation strength between all business review nodes. The declaration node network is constructed by combining the similarity of all business declaration nodes with their corresponding business declaration types. The declaration node network is connected to the horizontal business review node network via the vertical business transmission node chain.

[0050] Based on a pre-defined two-way risk assessment model, combined with real-time business audit information corresponding to the horizontal business audit node network and the vertical business transmission node chain, the probability of business audit transmission risk, the probability of audit delay, and the probability of audit error are obtained and fed back to the enhanced audit space to adjust the business transmission and audit process.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] This invention addresses the shortcomings of existing technologies by constructing a dynamically adaptive customs review system through deep collaboration between multi-source data fusion and intelligent architecture. The multimodal extraction model of the data acquisition module integrates review rules and customs declaration information in real time, ensuring accurate alignment between rules and business data, and drives dynamic iteration of review rules through a reinforced update mechanism. The compliance review module integrates the topology of the declaration node network, the vertical business transmission node chain, and the horizontal business review node network. Based on a federated learning framework and distributed reinforcement learning algorithms, it achieves business similarity-driven clustering review, encrypted transmission path optimization, and collaborative decision-making among related review nodes. The horizontal business review node network models the business association strength between review nodes using graph algorithms, while the vertical business transmission node chain dynamically allocates review tasks using a load-aware strategy. The evaluation and optimization module's bidirectional risk assessment model analyzes transmission risks, latency, and error probabilities in real time, and forms a closed-loop control through a multi-level feedback mechanism that links encryption strategy adjustments, reinforces model parameter optimization, and review process reconstruction. This invention possesses real-time rule adaptation, cross-modal conflict resolution, and risk layering and tracing capabilities, while ensuring efficient collaboration of review tasks and end-to-end data security and reliability, achieving a systematic leap in review efficiency, accuracy, and security. Attached Figure Description

[0053] Figure 1 This is a module diagram of a customs compliance intelligent audit system according to the present invention;

[0054] Figure 2 This is a flowchart of a customs compliance intelligent audit system according to the present invention.

[0055] Figure 3 This is a flowchart of a customs compliance intelligent audit method according to the present invention. Detailed Implementation

[0056] Example 1:

[0057] Please see Figure 1 The present invention provides an embodiment of a customs compliance intelligent audit system, comprising: a data acquisition module, a compliance audit module, and an evaluation and optimization module;

[0058] The data acquisition module, based on real-time collected review rules and customs declaration information combined with a multimodal extraction model, extracts the review rule update table and customs declaration business information table in real time.

[0059] The compliance review module allows users to obtain a completed customs declaration information table based on the customs declaration classification information table, through a preset enhanced review space and review rule update table. The enhanced review space includes a declaration node network, a vertical business transmission node chain, and a horizontal business review node network. This enhanced review space is constructed and trained using reinforcement learning algorithms, combining business information from the horizontal and vertical business transmission node chains, preset review rule indexes, and the priorities corresponding to each rule and business review. The vertical business transmission node chain is constructed using distributed reinforcement learning algorithms and encryption algorithms, connecting each declaration node, its corresponding business review node, and the customs declaration information transmission path. The horizontal business review node network is constructed using graph algorithms, combining the correlation strength between all business review nodes. The declaration node network is constructed by combining the similarity of all business declaration nodes with their corresponding business declaration types. The declaration node network is connected to the horizontal business review node network via the vertical business transmission node chain.

[0060] The evaluation and optimization module, based on a preset two-way risk assessment model and combined with real-time business review information corresponding to the horizontal business review node network and the vertical business transmission node chain, obtains the probability of business review transmission risk, the probability of review delay, and the probability of review error, and feeds it back to the enhanced review space to adjust the business transmission and review process.

[0061] In this embodiment, please refer to Figure 2 The data acquisition module includes a data acquisition unit, a multimodal extraction unit, and an enhancement and update unit.

[0062] The data acquisition unit is used to collect real-time business customs declaration information uploaded by all business declaration nodes, business customs declaration information with incomplete review processes, and updated business review rule information, and to classify the collected data information according to data type.

[0063] Furthermore, in this embodiment, the business customs declaration information specifically includes, but is not limited to, unique identification information, such as declaration form ID and user ID; product details, such as HS code and value; logistics information, such as mode of transport and bill of lading number; transaction party information, such as importer and exporter information; and compliance documents, such as certificate of origin and quality inspection report.

[0064] Information on incomplete customs declarations during the review process includes, but is not limited to, recording the process status, such as pending preliminary review; reasons for suspension, such as missing data or rule conflicts; and operation logs, such as rejection records and user resubmission actions.

[0065] Updating audit rule information includes, but is not limited to, rule attributes such as ID and effective period; logical descriptions such as conditional expressions and associated fields; dynamic parameters such as exchange rate tables and risk lists; and original audit rule documentation and explanations.

[0066] Data types include, but are not limited to, structured fields, unstructured files, and semi-structured logs, supporting full data storage and parsing.

[0067] This system ensures compliance, efficiency, and traceability of customs declarations through standardized definitions and dynamic rule management, adapting to complex needs such as customs supervision and international trade.

[0068] The multimodal extraction unit, based on the real-time collected information after classification and combined with the multimodal feature extraction algorithm, obtains the customs declaration business information table and the audit rule update table, and performs secondary audit updates for incomplete declaration business starting from the effective timestamp of the audit rule in the audit rule update table.

[0069] Furthermore, the detailed process of obtaining the customs declaration business information table and the audit rule update table in this embodiment includes:

[0070] First, structured, semi-structured and unstructured data are received in real time through the reporting nodes. Regular expression validation, OCR text extraction, image recognition models and JSON parsing technology are used to classify, clean and standardize the format of the raw data.

[0071] It is important to note that in this embodiment, structured data is validated using format rules and currency conversion to ensure field integrity and consistency. Unstructured data utilizes an OCR engine to extract text features, combined with a YOLOv8 model to identify hazard symbols in images, and is converted into structured labels. The conditional expressions for parsing semi-structured data are executable triplet logic. This stage outputs standardized multi-type data, providing high-quality input for subsequent processing. Key implementation details include anomaly data marking and re-examination, an OCR retry mechanism, and image recognition confidence threshold control. Format rule validation is preferably performed using the declaration form ID format.

[0072] Second, based on predefined field templates, the cleaned structured data is directly mapped to the customs declaration business information table. At the same time, key information is extracted from unstructured documents through OCR and NLP technologies, and the ResNet model is used to identify brand trademarks in product images and associate them with the intellectual property database.

[0073] Furthermore, in this embodiment, the semi-structured rule data is decomposed into triple logical conditions and dynamically associated with external parameter interfaces. This step realizes feature fusion and semantic unification of multi-source data and outputs a standardized feature set. Attention should be paid to text semantic standardization, sensitive word library matching, and secure escaping processing in rule parsing.

[0074] Third, multimodal features are integrated using the declaration form ID as the primary key, and data credibility is ensured through integrity checks and logical verification. It should be noted that in this embodiment, cargo value verification uses asynchronous calculation to avoid blocking; state conflicts trigger an automatic freeze process, outputting a complete and consistent customs declaration information table to support subsequent review rule matching.

[0075] Fourth, parse the rule attributes and design a table structure to store logical conditions and dynamic parameters. Synchronize external parameters and manage rule status through scheduled tasks. Furthermore, this embodiment outputs a dynamically effective audit rule table. Attention must be paid to UTC timestamp storage, soft deletion marking, and the circuit breaker mechanism for multi-source parameter interfaces. Rule attributes include, but are not limited to, ID, type, and effective time; rule status management includes, but is not limited to, marking expired / pending rules.

[0076] Fifth, filter incomplete customs declarations based on the rule's effective time and field correlation, re-execute rule verification and update the status, triggering manual review and audit log recording in case of conflict; it should be noted that the updated business status and audit trajectory are output, including time window matching, snapshot comparison analysis and two-way conflict details push;

[0077] Sixth, business tables and rule tables are persistently stored in a relational database, and a composite index is established to optimize query performance; a message queue pushes rule change notifications and correction reminders, and a priority channel is set up to ensure emergency notifications; it should be noted that this embodiment uses blockchain to store key data change records, supports multi-institutional consensus auditing, and outputs queryable data tables, real-time notifications, and an immutable audit link, paying attention to index design, message priority classification, and blockchain node disaster recovery deployment.

[0078] The enhanced update unit is used to update the review rule information corresponding to each node in the application node network and the horizontal business review node network according to the review rule update table and the preset enhanced update model, and to use the updated review rules to perform secondary review updates for incomplete application business reviews.

[0079] In this embodiment, the compliance review module includes an application clustering unit, an initial review unit, a transmission unit, and a secondary review unit;

[0080] The application clustering unit is based on the business information corresponding to all business application nodes and the business similarity clustering model to obtain a set of similar application nodes, and uses the cluster center corresponding to each subset of similar application nodes in the set of similar application nodes as the initial review node.

[0081] The initial review unit obtains the customs declaration business information table after initial review based on the first enhanced review model configured for each initial review node, combined with the customs declaration business information table and corresponding review rules in the corresponding similar declaration node subset.

[0082] Furthermore, the process of obtaining the customs declaration business information form after the initial review includes:

[0083] Each user who submits a business application is treated as a business application node. Based on the business type information of the user and the corresponding target customer information, a similarity algorithm is used to obtain the business similarity of different users who submit applications at the same timestamp.

[0084] Based on the similarity between business application nodes and their corresponding businesses, a network of application nodes is constructed using a graph algorithm.

[0085] Based on the real-time collection of business customs declaration information of different declaring users under the same timestamp by the declaration node network, the collected business customs declaration information is input into the multimodal extraction model to obtain the customs declaration business information table corresponding to different declaring users;

[0086] The customs declaration business information table includes a unique identifier, business type, and structured business process information after feature extraction;

[0087] Based on the business information and business similarity corresponding to the application node network, clustering is performed by combining clustering algorithm with time series algorithm to obtain a subset of similar application nodes corresponding to each similar business type, and the cluster center node of the subset of similar application nodes is used as the initial review node of the corresponding subset of similar application nodes.

[0088] The first enhanced review model is constructed by using a federated algorithm framework based on the first enhanced review sub-model configured for each application node in each subset of similar application nodes and the first evaluation sub-model configured for the corresponding initial review node.

[0089] Based on the customs declaration business information table corresponding to each declaration node, the corresponding first enhanced audit sub-model is used for auditing, and the audit result information is transmitted to the corresponding initial audit node through the federated framework. The first evaluation sub-model is configured for auditing and evaluation. If the evaluation is passed, the customs declaration business information table of the initial audit is output.

[0090] If the review fails, the assessment results will be fed back to the corresponding declaration node through the federal framework for adjustment until the first review threshold is met, at which point the adjusted initial review completion customs declaration information table will be obtained.

[0091] For example, the process of obtaining the customs declaration business information form after the initial review needs further explanation:

[0092] First, the business relevance of the reporting nodes is quantified by a hybrid similarity algorithm to construct a node network graph. It should be noted that in this embodiment, a similarity matrix is ​​generated by comprehensively calculating the structured business type, target country, and product text features, combined with adjacency list storage optimization. The hybrid similarity algorithm in this embodiment is obtained by weighting through Jaccard classification feature matching, TF-IDF text vectorization, and time window.

[0093] Second, based on the DBSCAN algorithm combined with time-constrained clustering, highly similar nodes are divided into subsets, and nodes with high historical approval rates and strong business consistency are selected as the initial approval benchmark. By expanding time-series features (application frequency, weekly trends) and dynamically adjusting density thresholds, the problem of traffic fluctuations during peak and off-peak periods is addressed, outputting reliable subsets and initial nodes. Through a degradation replacement mechanism for cluster centers and adaptive expansion of the neighborhood range, low-quality data is avoided from being clustered incorrectly; in this embodiment, the degradation replacement mechanism is implemented by selecting the second-best similar nodes with a priority approval rate greater than a preset approval threshold.

[0094] It should be further explained that this embodiment selects the second-best similar nodes with a pass rate greater than a preset pass threshold for the degradation and replacement mechanism because of the core objectives and data quality requirements of the audit business. In the customs declaration audit scenario, the historical pass rate directly reflects the accuracy and reliability of the node's business processing and is a key indicator to ensure audit quality. If only similarity is relied upon, nodes with similar business characteristics but poor audit quality may be included, making the initial audit benchmark unreliable and affecting subsequent audit results. Setting a pass threshold to filter second-best similar nodes can ensure the business relevance between nodes, maintain the effectiveness of clustering, and ensure that the selected nodes have high audit capabilities. This improves the quality of the audit benchmark from the source, reduces the risk of audit errors caused by node quality issues, and ensures the efficiency and accuracy of the entire audit process.

[0095] Third, under the federal framework, the reporting node deploys a local sub-model to perform rule verification, and the initial review node aggregates global parameters to update the model, achieving collaborative optimization under privacy protection. It should be noted that the global parameter update model in this embodiment uses homomorphic encryption to transmit gradient parameters, combined with rule coverage monitoring to trigger incremental training, to ensure the real-time performance of the model and the completeness of the rules. Specific parameters include synchronization frequency control, secure multi-party computation protocol and early warning of uncovered rules, balancing efficiency and security.

[0096] Fourth, the local sub-model performs rule matching and data integrity verification. The evaluation model generates review suggestions by calculating risk levels and comparing them with historical consistency, while simultaneously triggering automated resubmission guidance and manual review mechanisms, forming a closed loop of "initial review-feedback-re-review". It is necessary to strengthen the dynamic loading of mandatory fields, switching verification rules and risk scoring algorithms according to business type to reduce missed and false detections. Furthermore, in this embodiment, rule matching is preferably HS code tax rate mapping; data integrity verification is preferably mandatory field detection.

[0097] The transmission unit, based on the customs declaration business information table after review corresponding to each initial review node, combined with the preset encryption code, the load information of the transmission channel corresponding to each vertical business transmission node chain, the review business type information and load information corresponding to all business review nodes, and the initial review evaluation result information corresponding to each declaration node, transmits and distributes the customs declaration business information table after initial review to the corresponding business review node through the transmission path optimization algorithm combined with the distributed reinforcement allocation algorithm.

[0098] Furthermore, the detailed implementation process of transmitting and allocating the customs declaration business information table after initial review to the corresponding business review node in this embodiment includes:

[0099] First, the encryption strength is adaptively selected based on the sensitivity of the product using a dynamic encryption level matching algorithm, and the transmission link status is evaluated in conjunction with a real-time channel load quantization algorithm. In this embodiment, the dynamic encryption level matching algorithm is preferably AES-128 / 256 and blockchain notarization; the load quantization algorithm is preferably constructed from a random forest pre-trained with bandwidth, latency, and error rate.

[0100] Second, a transmission task graph model is constructed based on the audit node capability matrix modeling and the Jaccard similarity algorithm to calculate the node matching degree. In this embodiment, the audit node capability matrix modeling is constructed from audit type, load, and accuracy. The transmission task graph model is constructed by combining the improved Dijkstra algorithm with path cost.

[0101] Third, distributed reinforcement learning is used to dynamically optimize the combination of nodes and channels to balance matching degree and load;

[0102] Fourth, data packet compression, SHA-256 hash verification, and blockchain notarization technologies are used to ensure transmission integrity, and an exception handling mechanism is used to trigger the selection of alternative paths.

[0103] Fifth, rely on real-time monitoring panels and dynamic parameter adjustment technology to iteratively optimize priority formulas and routing strategies.

[0104] This process significantly improves the overall intelligence level of the cross-border trade review system by constructing a multimodal data fusion framework and a dynamic rule collaboration mechanism. At the data governance level, a hybrid processing mode of structured feature mapping and unstructured semantic parsing is adopted to achieve multi-dimensional feature decoupling and reorganization of declaration information, effectively solving the problem of fragmented review elements caused by data silos in traditional customs declaration systems. Through cross-modal correlation verification of HS codes and logistics documents, a spatiotemporal consistency verification chain of commodity attributes and logistics trajectories is formed, strengthening the ability to identify abnormal declaration patterns. The timestamp-driven mechanism of the review rule update table, combined with the dynamic matching of rule effectiveness and the status of business under review, establishes a causal relationship model between rule versions and review progress, enabling historical declaration data to be automatically backtracked and reassessed as review rules change, eliminating the risk of review logic gaps caused by regulatory rule iterations. At the node collaboration level, a federated learning framework based on business similarity maps heterogeneous declaration nodes to homogeneous business clusters through feature space projection. Secure aggregation of model parameters within the cluster achieves review knowledge sharing, breaking through the limitations of single-node samples while ensuring data privacy, and enhancing the model's generalization ability to regional trade characteristics. The dynamic optimization algorithm for transmission paths constructs a multi-dimensional constraint-based transmission strategy through adaptive matching of encryption levels and real-time network topology awareness. It implements differentiated encryption strategies based on product sensitivity levels, ensuring core data security while avoiding resource waste caused by excessive encryption. Combined with joint modeling of channel load index and node processing capacity, a balance function between transmission cost and review efficiency is established, enabling elastic resource scheduling in high-concurrency scenarios. Deep integration of blockchain notarization technology creates an irreversible notarization chain between rule file hash values ​​and key review operations. Through a smart contract-driven automatic verification mechanism, the auditability and non-repudiation of cross-node data interactions are ensured. In terms of anomaly handling, an intelligent routing switching mechanism and a rapid backup node election strategy construct a fault-tolerant recovery system for transmission failures. Combined with real-time visual feedback from the transmission quality monitoring panel, a closed-loop optimization path decision-making iteration mechanism is formed. The time-series clustering algorithm incorporates historical application frequency and periodic fluctuation characteristics, enhancing the temporal correlation of business similarity calculations. This ensures that node clustering results reflect both the commonalities of business attributes and the dynamic patterns of application behavior. The homomorphic encrypted transmission of audit model parameters and the distributed gradient aggregation mechanism achieve a precise balance between data availability and privacy during federated learning, establishing a trusted computing environment for multi-entity collaborative auditing. The dynamic priority calculation model quantifies business attributes such as risk level and correction records into transmission weight factors, and uses a weighted decision tree to allocate transmission resources preferentially to key business operations, ensuring the timeliness of high-priority applications.The systematic integration of these technologies enables the customs declaration review system to form a positive reinforcement loop in data collection, rule adaptation, node collaboration, and secure transmission, building an intelligent review ecosystem with elastic scalability and continuous evolution characteristics, and providing full-chain technical support for compliance supervision in complex trade environments.

[0105] The secondary review unit obtains the customs declaration business information table after secondary review based on the priority of the review business corresponding to each business review node, the priority of the sub-business review process corresponding to each business, the index information between the review rules and each sub-business review process, and the second enhanced review model configured for each business review node.

[0106] In this embodiment, the evaluation and optimization module includes a first evaluation unit, a second evaluation unit, and an evaluation feedback unit;

[0107] The first evaluation unit is used to conduct real-time evaluation based on the clustering process corresponding to the application clustering unit, the initial review process corresponding to the initial review unit, and the supply-demand relationship stability coefficient, through the first evaluation sub-model to obtain the first review risk probability. The first review risk probability is constructed from the clustering error probability and the initial false detection and false detection risk probability. The supply-demand relationship stability coefficient is constructed from the ratio of the transaction completion frequency of the supply and demand parties to the total transaction frequency.

[0108] The second evaluation unit is used to perform real-time evaluation based on the data transmission process corresponding to the transmission unit and the secondary review process and review anomaly coefficient corresponding to the secondary review unit, using the second evaluation sub-model to obtain the second review risk probability. The second review risk probability is constructed by combining the tampering risk probability and business allocation risk probability corresponding to the data transmission process and the risk probabilities corresponding to false detection, missed detection and review delay corresponding to the secondary review with a Bayesian algorithm. It should be further noted that one specific implementation method of the second review risk probability is as follows:

[0109] The probability of data transmission tampering risk is obtained based on blockchain evidence storage technology and SHA-256 hash verification algorithm, and it is used as the prior input of Bayesian fusion model;

[0110] The transmission data feature deviation score is obtained based on the LSTM-autoencoder timing anomaly detection algorithm and used as a conditional variable for tampering risk.

[0111] The probability of business allocation risk is obtained by modeling the Jaccard similarity algorithm and the capability matrix of the review nodes, and then input into the Bayesian model;

[0112] The probability of false detection and false detection risk is obtained based on the historical confusion matrix and time decay algorithm of the second enhanced review model, and is used as a priori input.

[0113] The probability of review delay risk is obtained based on the Weibull distribution model and the maximum likelihood estimation algorithm, and Bayesian fusion is incorporated.

[0114] The audit anomaly coefficients were obtained based on the time series clustering algorithm (K-Means, K=5) and the BERT pre-trained model, and used as Bayesian conditional variables.

[0115] The Metropolis-Hastings algorithm based on Markov chain Monte Carlo (MCMC) performs posterior sampling on the Bayesian model and iterates 10,000 times to obtain the fused second audit risk probability.

[0116] The channel load score is output based on the random forest algorithm, and the review priority score is obtained based on the weighted decision tree algorithm. The product of the two is used as a dynamic weight adjustment factor.

[0117] Based on the HSL color space mapping algorithm, the risk probability is converted into a visual heatmap, which is then pushed to the monitoring panel via the WebSocket protocol.

[0118] When the probability of risk exceeds a preset threshold, a backup node switch is triggered based on an intelligent routing algorithm, forming a complete closed loop of data collection, risk quantification, model fusion, and dynamic prevention and control.

[0119] The review anomaly coefficient is constructed by combining the historical anomaly frequency of the corresponding applicant user with the total application frequency;

[0120] The evaluation feedback unit optimizes the first enhanced audit model, the second enhanced audit model, and the encrypted code information by providing graded feedback based on the risk probability difference between the first audit risk probability and the second audit risk probability, combined with preset multi-level risk discrimination thresholds.

[0121] Furthermore, the workflow for evaluating the feedback unit includes:

[0122] Preset first review threshold First review difference judgment threshold Second review difference threshold ,and When the probability of risk in the first review is less than If the initial audit result is positive, the customs declaration information table will be transmitted and distributed to the corresponding audit node. Otherwise, the first audit risk probability will be fed back to the first enhanced audit model to update and optimize the clustering process and the initial audit process until the first audit risk probability is less than 1%. until;

[0123] When the risk probability difference is less than If the second review is completed, a customs declaration information table will be output. If the risk probability difference is greater than or equal to... and less than If the second audit risk probability is not met, it will be fed back to the transmission unit and the second enhanced audit model to optimize and update the data transmission process and the secondary audit process until the risk probability difference is less than 1. Until then;

[0124] When the risk probability difference is greater than or equal to If an encryption anomaly is detected, the business information corresponding to the business review node is compared with the business information of the corresponding application node to obtain the comparison difference information. Simultaneously, the encryption encoding level in the transmission unit and the business information indicating encryption anomaly are adjusted using the obtained comparison difference information. The adjusted business information undergoes a second update review until the risk probability difference is less than [a certain value]. Until then.

[0125] Specifically, the process of obtaining the customs declaration information form after the second review includes:

[0126] Each department of the reviewing party is treated as a business review node. The review process of the corresponding business type under each business review node and the review priority of each review sub-node in the current business type review process are used to construct a business review process node chain corresponding to each business review node through blockchain algorithm. Each review sub-node corresponds one-to-one with each review sub-process in the corresponding business type review process.

[0127] Based on the business review process node chain information corresponding to each business review node, the business overlap correlation degree corresponding to each business review node is obtained through the association algorithm.

[0128] Based on each business review node and its corresponding business overlap correlation, a horizontal business review node network is constructed using a graph algorithm.

[0129] The horizontal business audit node network obtains the audit results and the comprehensive results of the corresponding business by means of the customs declaration business information table completed by the initial audit and the corresponding audit priority and initial audit evaluation results allocated by the vertical business transmission node chain, the audit rule index information corresponding to the audit sub-process stored by each audit sub-node under each business audit node, and the second enhanced audit sub-model configured by each audit sub-node.

[0130] Based on the comprehensive results of each business audit node and the corresponding initial business audit results, combined with the first audit difference judgment threshold, the second audit difference judgment threshold and the second evaluation sub-model, a second audit judgment is made. If the judgment is passed, a customs declaration business information table with the second audit completed is obtained.

[0131] Otherwise, the corresponding business, initial review process, transmission process and secondary review process will be adjusted in real time based on the results of the second review until all review sub-processes of each review business meet the review rules.

[0132] The second enhanced audit model is constructed through a federated framework by configuring the second enhanced audit sub-model and the corresponding second evaluation sub-model for each audit sub-node.

[0133] Furthermore, in this embodiment, the process of obtaining the customs declaration business information form after the second review includes:

[0134] First, based on the business review process node chain, a recursive decomposition algorithm is used to break down complex business processes into the smallest executable atomic sub-processes, each acting as an independent unit. Then, using blockchain notarization technology, hash pointers are used to chain these sub-processes together. Key data from each sub-process is hashed using the SHA-256 hash algorithm to generate a unique hash value, pointing to the hash value of the previous sub-process, forming an immutable node relationship chain. This ensures the review process is traceable and the data is authentic and reliable. Through recursive decomposition and hash chain construction, complex business processes are transformed into a traceable, tamper-proof atomic sub-process chain, providing a foundation for subsequent collaborative review. Key data includes, but is not limited to, operation records and timestamps.

[0135] Second, based on the business review rule fields, the Jaccard similarity algorithm is used to calculate the overlap and correlation between rules. Specifically, the ratio of the number of elements in the intersection to the number of elements in the union of rule fields is used to obtain a quantified similarity value. Rules with similarity values ​​higher than a preset threshold are considered as associated rules. Edge connections are established for these rules using graph theory, gradually constructing a horizontal network of business review nodes to achieve the associative integration of review rules. Business review rule fields include, but are not limited to, HS codes and country of origin.

[0136] Third, based on graph algorithms, highly correlated nodes in the horizontal business review node network are identified and connected to form a subgraph; using dynamic routing configuration algorithms, combined with real-time load data of review sub-nodes, review tasks are dynamically allocated to idle or low-load nodes to achieve load balancing across business reviews.

[0137] Simultaneously, a failover mechanism is set up so that when a node fails, the routing path is automatically switched through a backup path search algorithm to ensure the continuous execution of the audit tasks. This embodiment constructs a subgraph using a graph algorithm, and combines dynamic routing with the set failover mechanism to achieve dynamic balanced distribution and reliable execution of audit tasks based on node load. The graph algorithm is preferably a depth-first search algorithm; the dynamic routing configuration algorithm is preferably an improved shortest path first algorithm; the real-time load data of the audit sub-nodes includes, but is not limited to, task backlog, processing rate, etc.; the set failover mechanism specifically means: automatically switching the routing path when a node fails.

[0138] Fourth, based on the federated learning algorithm, local review models are deployed at each review sub-node, and each review sub-node trains the model on local data. During training, local gradient parameters are encrypted and aggregated using a secure aggregation protocol, and the encrypted parameters are uploaded to the central server. The central server uses a weighted average algorithm to fuse the parameters of each node, generate a global model, and feed it back to each sub-node to achieve dynamic model updates and improve the model's review capabilities while protecting data privacy. The Secure Boost algorithm is preferred for the secure aggregation protocol, and the FedAvg algorithm is preferred for the weighted average algorithm.

[0139] Fifth, based on a multimodal result aggregation algorithm, the audit results output by each audit sub-node are fused and processed. Appropriate weights are assigned according to the credibility of different modal results, and a comprehensive result is calculated. This comprehensive result is compared with a preset audit result threshold. When the difference exceeds the threshold, a manual review process is triggered; if it is within the threshold range, the process is automatically optimized to achieve intelligent processing of audit results. The preferred multimodal result aggregation algorithm is a weighted voting algorithm. Audit results include, but are not limited to, textual conclusions, risk scores, and image recognition results.

[0140] Sixth, based on a two-way risk assessment model, the deviation probability is calculated by comprehensively considering the differences between the initial and secondary audit results. The data encryption level and routing strategy are dynamically adjusted according to the magnitude of the deviation probability. When the deviation probability is high, the encryption level is increased to ensure data security, while the routing path is optimized, selecting more reliable nodes for task transmission and auditing to reduce risk. The preferred two-way risk assessment model is a Bayesian risk assessment model. Dynamic adjustment of the data encryption level is implemented using the AES-128 or AES-256 algorithm. The routing strategy involves adjusting the weight parameters of the shortest path algorithm based on the adjusted encryption level. This process is implemented by those skilled in the art using existing technology.

[0141] Seventh, a reinforcement learning algorithm is introduced, with review efficiency and risk control as the objective functions, to construct a reward mechanism. Through continuous trial and error, the reinforcement learning algorithm learns the optimal load allocation and path selection strategies under different node loads and business scenarios. Based on real-time node status and task attributes, the allocation scheme and routing path are dynamically adjusted to achieve a balance between review efficiency and risk control. Review efficiency includes, but is not limited to, task processing time and throughput; risk control includes review error rate and data leakage risk; node status includes, but is not limited to, load, review capacity, and risk level; task attributes include, but are not limited to, urgency and complexity. This process is implemented by those skilled in the art using existing technologies.

[0142] This process achieves intelligent end-to-end review from business decomposition to result output by ensuring traceability of the node chain, constructing a horizontal business review node network, iterative optimization of the federated model, and a closed-loop feedback mechanism, thereby improving the accuracy, security, and efficiency of customs declaration processing.

[0143] This process, through the integration of multimodal review processes and dynamic collaboration mechanisms, constructs an intelligent customs declaration review system with elastic scalability and continuous optimization. Based on blockchain-based review node chain technology, the rule logic and execution paths of atomic review sub-processes are stored in a chain, ensuring the immutability and full-chain traceability of each operation, thereby enhancing the transparency and credibility of the review process. The horizontal business review node network, through business association quantification and graph network modeling, achieves intelligent routing and load balancing of cross-departmental review resources. When the processing capacity of a single node is limited, tasks can be dynamically allocated to collaborative nodes based on association weights, effectively improving review throughput in high-concurrency scenarios. The federated learning framework-driven sub-node model collaborative optimization mechanism, through encrypted parameter aggregation and distributed model updates, achieves knowledge sharing of review rule features and enhanced model generalization capabilities while protecting the data privacy of each node. For example, the ability to identify forgery patterns in cross-border certificates of origin can be synchronously migrated across nodes, significantly reducing the probability of missed detections of regional risks.

[0144] The dynamic priority-driven transmission encryption mechanism adaptively matches differentiated encryption strategies and transmission paths through multi-dimensional analysis of product attributes and risk levels. This ensures the security of sensitive data while avoiding resource waste caused by excessive encryption in low-risk businesses. The bidirectional risk assessment model constructs a risk difference-driven hierarchical feedback system by probabilistically modeling multiple risk factors such as clustering errors, rule conflicts, and transmission leakage. When encryption anomalies or gaps in audit logic are detected, it automatically triggers transmission path rerouting, encryption level upgrades, and model parameter retraining, forming a complete control chain from risk perception to closed-loop optimization. The smart contract-enabled automated verification process embeds audit rule logic into on-chain executable code, achieving seamless execution of rule matching and result judgment, reducing response delays and operational errors caused by manual intervention. Furthermore, business overlap correlation analysis technology, through historical data mining and rule pattern matching, identifies common characteristics and dependencies in cross-business review processes, providing a logical basis for collaborative review. For example, optimization strategies for the origin verification process can be horizontally reused to related business nodes, improving overall review consistency. A multi-level threshold mechanism, combined with real-time risk assessment results, enables dynamic resilience adjustments to the review process. In low-risk scenarios, rapid release is achieved, while in medium- and high-risk scenarios, multiple filtering methods such as model iteration, path switching, and manual review are used to achieve a precise balance between efficiency and security. This systemic interoperability design based on technological means enables the customs declaration review system to adapt to changes in review rules, business fluctuations, and security threats, continuously optimizing the intelligent customs compliance review system.

[0145] Example 2:

[0146] Please see Figure 3 Another embodiment of the present invention provides: a customs compliance intelligent audit method, comprising:

[0147] Based on real-time collected review rules and customs declaration information, combined with a multimodal extraction algorithm, an review rule update table and a customs declaration business information table are obtained.

[0148] Based on the customs declaration classification information table, the table is updated by combining the preset enhanced review space with the review rules to obtain the customs declaration business information table after the review is completed.

[0149] The enhanced audit space comprises a horizontal business audit node network and a vertical business transmission node chain. This enhanced audit space is constructed and trained using reinforcement learning algorithms, combining business information from the horizontal and vertical business audit node networks and the vertical business transmission node chain, preset audit rule indexes, and the priorities corresponding to each rule and business audit. The vertical business transmission node chain is constructed using distributed reinforcement learning algorithms and encryption algorithms, connecting each declaration node with its corresponding business audit node and the customs declaration business information transmission path. The horizontal business audit node network is constructed using graph algorithms, combining the correlation strength between all business audit nodes. The declaration node network is constructed by combining the similarity of all business declaration nodes with their corresponding business declaration types. The declaration node network is connected to the horizontal business audit node network via the vertical business transmission node chain.

[0150] Based on a pre-set two-way risk assessment model, combined with real-time business audit information corresponding to the horizontal business audit node network and the vertical business transmission node chain, the probability of business audit transmission risk, the probability of audit delay, and the probability of audit error are obtained and fed back to the enhanced audit space to adjust the business transmission and audit process.

[0151] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the claims. All of these variations are within the protection scope of the present invention.

Claims

1. A customs compliance intelligent audit system, characterized in that, include: Data acquisition module, compliance audit module, evaluation and optimization module; The data acquisition module, based on real-time collected review rules and customs declaration information combined with a multimodal extraction model, extracts the review rule update table and customs declaration business information table in real time. The compliance review module allows users to obtain a completed customs declaration business information table by combining a pre-set enhanced review space with an updated review rule table based on the customs declaration classification information table. The enhanced review space includes a declaration node network, a vertical business transmission node chain, and a horizontal business review node network. The enhanced review space is constructed and trained using a reinforcement learning algorithm, combining business information from the horizontal business review node network and the vertical business transmission node chain, a pre-set review rule index, and the priorities corresponding to the rules and business reviews. The evaluation and optimization module, based on a preset two-way risk assessment model and combined with the real-time business review information corresponding to the horizontal business review node network and the vertical business transmission node chain, obtains the probability of business review transmission risk, the probability of review delay, and the probability of review error, and feeds them back to the enhanced review space to adjust the business transmission and review process. The compliance review module includes an application clustering unit, an initial review unit, a transmission unit, and a secondary review unit. The evaluation and optimization module includes a first evaluation unit, a second evaluation unit, and an evaluation feedback unit; The first evaluation unit is used to perform real-time evaluation through the first evaluation sub-model based on the clustering process corresponding to the application clustering unit, the initial review process corresponding to the initial review unit, and the supply-demand relationship stability coefficient, to obtain the first review risk probability; the first review risk probability is constructed from the clustering error probability and the initial false detection and false detection risk probability; the supply-demand relationship stability coefficient is constructed from the ratio of the transaction completion frequency of the supply and demand parties to the total transaction frequency; The second evaluation unit is used to perform real-time evaluation through the second evaluation sub-model based on the data transmission process corresponding to the transmission unit and the secondary review process and review anomaly coefficient corresponding to the secondary review unit, so as to obtain the second review risk probability. The second audit risk probability is constructed by combining the tampering risk probability corresponding to the data transmission process, the business allocation risk probability, and the risk probabilities corresponding to false detection, false omission, and audit delay in the secondary audit with a Bayesian algorithm; The audit anomaly coefficient is constructed from the historical anomaly frequency of the corresponding applicant and the total application frequency; The evaluation feedback unit optimizes the first enhanced audit model, the second enhanced audit model, and the encrypted information by providing graded feedback based on the risk probability difference between the first audit risk probability and the second audit risk probability, combined with preset multi-level risk discrimination thresholds.

2. The customs compliance intelligent audit system as described in claim 1, characterized in that, The vertical business transmission node chain is constructed by combining distributed reinforcement learning algorithms and encryption algorithms with each declaration node, its corresponding business review node, and the customs declaration business information transmission path; the horizontal business review node network is constructed by combining the correlation strength between all business review nodes using graph algorithms; the declaration node network is constructed by combining the similarity of corresponding business declaration types with all business declaration nodes; the declaration node network is connected to the horizontal business review node network through the vertical business transmission node chain. The data acquisition module includes a data acquisition unit, a multimodal extraction unit, and an enhancement and update unit; The data acquisition unit is used to collect in real time all business declaration nodes' uploaded business customs declaration information, business customs declaration information whose review process is not completed, and updated business review rule information, and classify the collected data information according to data type; The multimodal extraction unit obtains a customs declaration business information table and an audit rule update table based on the real-time collected information after classification and a multimodal feature extraction algorithm. It then performs a second audit update of incomplete declaration business based on the effective timestamp of the audit rule in the audit rule update table. The enhanced update unit is used to update the review rule information corresponding to each node in the application node network and the horizontal business review node network according to the review rule update table and the preset enhanced update model, and to use the updated review rules to perform secondary review updates for incomplete application business reviews.

3. The customs compliance intelligent audit system as described in claim 2, characterized in that, The application clustering unit obtains a set of similar application nodes based on the business information corresponding to all business application nodes and a business similarity clustering model, and uses the cluster center corresponding to each subset of similar application nodes in the set of similar application nodes as the initial review node. The initial review unit obtains the customs declaration business information table after initial review based on the first enhanced review model configured for each initial review node, combined with the customs declaration business information table and corresponding review rules in the corresponding similar declaration node subset. The transmission unit, based on the customs declaration business information table after review corresponding to each initial review node, combined with the preset encryption code, the load information of the transmission channel corresponding to each vertical business transmission node chain, the review business type information and load information corresponding to all the business review nodes, and the initial review evaluation result information corresponding to each declaration node, transmits and distributes the customs declaration business information table after initial review to the corresponding business review node through a transmission path optimization algorithm combined with a distributed reinforcement allocation algorithm. The secondary review unit obtains the customs declaration business information table after secondary review based on the priority of the review business corresponding to each business review node, the priority of the sub-business review process corresponding to each business, the index information between the review rules and each sub-business review process, and the second enhanced review model configured for each business review node.

4. The customs compliance intelligent audit system as described in claim 3, characterized in that, The workflow of the evaluation feedback unit includes: Preset first review threshold First review difference judgment threshold Second review difference threshold ,and When the probability of risk in the first review is less than If the initial audit result is positive, the customs declaration information table will be transmitted and distributed to the corresponding audit node. Otherwise, the first audit risk probability will be fed back to the first enhanced audit model to update and optimize the clustering process and the initial audit process until the first audit risk probability is less than 1%. until; When the risk probability difference is less than If the second review is completed, a customs declaration information table will be output. If the risk probability difference is greater than or equal to... and less than If the second audit risk probability is not met, the data transmission process and the secondary audit process will be optimized and updated by feeding back the second audit risk probability to the transmission unit and the second enhanced audit model until the risk probability difference is less than 1. Until then; When the risk probability difference is greater than or equal to If an encryption anomaly is detected, the business information corresponding to the business review node is compared with the business information of the corresponding application node to obtain comparison difference information. Simultaneously, the encryption encoding level in the transmission unit and the business information indicating encryption anomaly are adjusted using the obtained comparison difference information. The adjusted business information undergoes a second update review until the risk probability difference is less than [a certain value]. Until then.

5. The customs compliance intelligent audit system as described in claim 4, characterized in that, The process of obtaining the customs declaration business information form after the initial review includes: Each user who submits a business application is treated as a business application node. Based on the business type information of the user and the corresponding target customer information, a similarity algorithm is used to obtain the business similarity of different users who submit applications at the same timestamp. Based on the similarity between business application nodes and their corresponding businesses, a network of application nodes is constructed using a graph algorithm. Based on the real-time collection of business customs declaration information of different declaring users under the same timestamp by the declaration node network, the collected business customs declaration information is input into the multimodal extraction model to obtain the customs declaration business information table corresponding to different declaring users; The customs declaration business information table includes a unique identifier, business type, and structured business process information after feature extraction; Based on the business information and business similarity corresponding to the application node network, clustering is performed by combining clustering algorithm with time series algorithm to obtain a subset of similar application nodes corresponding to each similar business type, and the cluster center node of the subset of similar application nodes is used as the initial review node of the corresponding subset of similar application nodes.

6. The customs compliance intelligent audit system as described in claim 5, characterized in that, The process of obtaining the customs declaration business information form after the initial review also includes: The first enhanced review model is constructed by using a federated algorithm framework based on the first enhanced review sub-model configured for each application node in each subset of similar application nodes and the first evaluation sub-model configured for the corresponding initial review node. Based on the customs declaration business information table corresponding to each declaration node, the corresponding first enhanced audit sub-model is used for auditing, and the audit result information is transmitted to the corresponding initial audit node through the federated framework. The first evaluation sub-model is configured for auditing and evaluation. If the evaluation is passed, the customs declaration business information table of the initial audit is output. If the review fails, the assessment results will be fed back to the corresponding declaration node through the federal framework for adjustment until the first review threshold is met, at which point the adjusted initial review completion customs declaration information table will be obtained.

7. The customs compliance intelligent audit system as described in claim 6, characterized in that, The process of obtaining the customs declaration information form after the second review includes: Each department of the reviewing party for each business type is treated as a business review node. The review process of the corresponding business type under each business review node and the review priority of each review sub-node in the review process under the current business type review process are used to construct a business review process node chain corresponding to each business review node through a blockchain algorithm. Each review sub-node corresponds one-to-one with each review sub-process in the corresponding business type review process. Based on the business review process node chain information corresponding to each business review node, the business overlap correlation degree corresponding to each business review node is obtained through the association algorithm. Based on each business review node and its corresponding business overlap correlation, a horizontal business review node network is constructed using a graph algorithm.

8. The customs compliance intelligent audit system as described in claim 7, characterized in that, The process of obtaining the customs declaration information form after the second review also includes: The horizontal business audit node network obtains the audit results and the comprehensive results of the corresponding business by means of the customs declaration business information table completed by the initial audit and the corresponding audit priority and initial audit evaluation results allocated by the vertical business transmission node chain, the audit rule index information corresponding to the audit sub-process stored by each audit sub-node under each business audit node, and the second enhanced audit sub-model configured by each audit sub-node. Based on the comprehensive results of each business audit node and the corresponding initial business audit results, combined with the first audit difference judgment threshold, the second audit difference judgment threshold and the second evaluation sub-model, a second audit judgment is made. If the judgment is passed, a customs declaration business information table with the second audit completed is obtained. Otherwise, the corresponding business, initial review process, transmission process and secondary review process will be adjusted in real time based on the results of the second review until all review sub-processes of each review business meet the review rules. The second enhanced audit model is constructed by the second enhanced audit sub-model configured for each audit sub-node and the corresponding second evaluation sub-model through a federated framework.

9. A customs compliance intelligent audit method, implemented based on the customs compliance intelligent audit system according to any one of claims 1-8, characterized in that, include: Based on real-time collected review rules and customs declaration information, combined with a multimodal extraction algorithm, an review rule update table and a customs declaration business information table are obtained. Based on the customs declaration classification information table, the table is updated by combining the preset enhanced review space with the review rules to obtain the customs declaration business information table after review. The enhanced audit space includes a horizontal business audit node network and a vertical business transmission node chain; The enhanced audit space is constructed and trained by combining the business information of the horizontal business audit node network and the vertical business transmission node chain, the preset audit rule index, and the priority of each rule and business audit with the reinforcement learning algorithm. The vertical business transmission node chain is constructed by combining distributed reinforcement learning algorithms and encryption algorithms with each declaration node, its corresponding business review node, and the customs declaration business information transmission path; the horizontal business review node network is constructed by combining the correlation strength between all business review nodes using graph algorithms; the declaration node network is constructed by combining the similarity of corresponding business declaration types with all business declaration nodes; the declaration node network is connected to the horizontal business review node network through the vertical business transmission node chain. Based on a pre-defined two-way risk assessment model, combined with real-time business audit information corresponding to the horizontal business audit node network and the vertical business transmission node chain, the probability of business audit transmission risk, the probability of audit delay, and the probability of audit error are obtained and fed back to the enhanced audit space to adjust the business transmission and audit process.

Images