A Cross-Border Electronic Customs Declaration Inspection Early Warning Method and System Based on Dynamic Risk Learning

Abstract

This invention provides a method and system for cross-border electronic customs declaration inspection and early warning based on dynamic risk learning. The method includes: acquiring historical and real-time declaration data, whereby historical declaration data includes structured and semi-structured fields, and real-time declaration data includes status markers, declaration timestamps, and company numbers; extracting static features from the structured and semi-structured fields, and extracting dynamic features from the status markers, declaration timestamps, and company numbers; constructing a dual-channel deep forest model, and inputting the static and dynamic features into the model to obtain static and dynamic risk probabilities; obtaining the risk probability based on the static and dynamic risk probabilities, and executing corresponding strategies based on the risk probabilities. This invention, by comprehensively extracting static and dynamic features and constructing a dual-channel deep forest model for risk assessment, enables comprehensive and dynamic analysis of customs declaration data.

Description

Technical Field

[0001] This invention relates to the field of cross-border electronic declaration detection technology, and in particular to a cross-border electronic declaration inspection and early warning method and system based on dynamic risk learning. Background Technology

[0002] In the context of globalized trade, cross-border electronic customs declaration, as a crucial link in international trade, directly impacts the smooth flow of trade through its efficiency and accuracy. With the rapid development of cross-border e-commerce, the volume of customs declaration data has increased dramatically, making traditional manual review methods insufficient to meet the demands of efficient customs clearance. Therefore, automated and intelligent customs declaration inspection and early warning methods have become a research hotspot.

[0003] Currently, cross-border electronic customs declaration inspection mainly relies on rule-based systems and simple statistical analysis methods. These methods typically use preset rules or thresholds to determine whether customs declaration data contains anomalies. For example, for key indicators such as value and weight, the system sets a fixed threshold range; if the data exceeds this range, it is considered an anomaly and an alert is issued. In addition, some advanced methods have introduced machine learning models to identify anomaly patterns by training on historical data.

[0004] Traditional methods often limit feature extraction to structured data, such as numerical features like value and weight, while neglecting valuable information in semi-structured or unstructured data (such as declared product names and descriptions). This limits the model's ability to identify complex anomaly patterns. Furthermore, existing methods rarely consider the dynamic changes in customs declaration behavior, such as abrupt changes in declaration frequency and the randomness of declaration time intervals. These dynamic features are crucial for the timely detection of abnormal declaration behavior. Summary of the Invention

[0005] The purpose of this invention is to provide a cross-border electronic customs declaration inspection and early warning method and system based on dynamic risk learning. It aims to achieve comprehensive and dynamic analysis of customs declaration data by comprehensively extracting static and dynamic features and constructing a dual-channel deep forest model for risk assessment.

[0006] In a first aspect, the present invention provides a method for cross-border electronic customs declaration inspection and early warning based on dynamic risk learning. The method includes: acquiring historical declaration data and real-time declaration data, wherein the historical declaration data includes structured fields and semi-structured fields, and the real-time declaration data includes status markers, declaration timestamps, and enterprise numbers; extracting static features from the structured fields and semi-structured fields, and extracting dynamic features from the status markers, declaration timestamps, and enterprise numbers; constructing a dual-channel deep forest model, and inputting the static features and the dynamic features into the dual-channel deep forest model respectively to obtain static risk probabilities and dynamic risk probabilities; obtaining risk probabilities based on the static risk probabilities and the dynamic risk probabilities, and executing corresponding strategies based on the risk probabilities.

[0007] Furthermore, the steps of extracting static features from the structured and semi-structured fields and extracting dynamic features from the status marker, declaration timestamp, and enterprise number include: extracting numerical features of value, weight, and country of origin code from the structured fields, the numerical features including the value dispersion coefficient and weight skewness coefficient; extracting semantic features of the declared product name and description text from the semi-structured fields, the semantic features including TF-IDF word vectors and commodity HS code embedding; and concatenating all the numerical and semantic features to obtain static features.

[0008] Furthermore, the steps of extracting static features from the structured and semi-structured fields and extracting dynamic features from the status marker, declaration timestamp, and enterprise number further include: extracting time-series behavioral features and state transition features from the real-time declaration data, and concatenating all time-series behavioral features and state transition features to obtain dynamic features; the state transition features include state dwell time and state loop count, and the time-series behavioral features include declaration frequency mutation index and declaration interval entropy.

[0009] Furthermore, the method includes: calculating the value dispersion coefficient according to the following formula: ,in, The coefficient of variation is the value of the goods. The standard deviation of the declared value of goods by the same enterprise within 30 days This is the average value of goods declared by the same company over 30 days.

[0010] The weight skewness coefficient is calculated using the following formula: ;in, The weight skewness coefficient, Let n be the weight of the goods in the i-th declaration form, and n be the total number of declaration forms. The average weight of the goods in all declarations. This represents the standard deviation of the cargo weight for all declarations.

[0011] The TF-IDF word vectors are calculated using the following formula: ;in, For terms t In the document d TF-IDF value in For terms t The frequency of occurrence in historical declaration data For included terms t The number of documents, This represents the total number of documents in the historical declaration data.

[0012] The 5-bit HS code H is converted into a numerical vector and represented using a hybrid one-hot encoding and word embedding.

[0013] Furthermore, the method includes: calculating the reporting frequency mutation index according to the following formula: ;in, The mutation index is the frequency of applications. The number of times a company submits an application within the time period t. This represents the average number of submissions made by the company within the daily time period t. This represents the standard deviation of the number of declarations made by a company within a daily time period t.

[0014] The declaration interval entropy is calculated using the following formula: ;in, To report the interval entropy, Let k be the probability density of the i-th reporting time interval, and k be the number of time intervals.

[0015] The state dwell time is calculated using the following formula: ;in, The state dwell time, All are empirical coefficients. This refers to the average time from application submission to release in the past. This represents the standard deviation of the time taken from application to release in the historical application process.

[0016] The statistical reporting process progresses from submitted to revised and back to submitted frequency.

[0017] Further, the step of obtaining the risk probability based on the static risk probability and the dynamic risk probability includes: calculating the risk probability according to the following formula: ;in, For risk probability, These are static risk probability and dynamic risk probability, respectively. All are weighting coefficients.

[0018] Furthermore, the step of executing the corresponding strategy based on the risk probability includes: if the risk probability is greater than or equal to the first threshold, the application is suspended and a risk report containing abnormal feature tracing is generated; if the risk probability is less than the first threshold but greater than or equal to the second threshold, a manual review mark is added; if the risk probability is less than the second threshold, the application is approved.

[0019] Secondly, this invention provides a cross-border electronic customs declaration inspection and early warning system based on dynamic risk learning. The system includes: a data acquisition module for acquiring historical declaration data and real-time declaration data, wherein the historical declaration data includes structured fields and semi-structured fields, and the real-time declaration data includes status markers, declaration timestamps, and enterprise numbers; a feature extraction module for extracting static features from the structured fields and semi-structured fields, and extracting dynamic features from the status markers, declaration timestamps, and enterprise numbers; a risk probability prediction module for constructing a dual-channel deep forest model, and inputting the static features and the dynamic features into the dual-channel deep forest model respectively to obtain static risk probabilities and dynamic risk probabilities; and an early warning execution module for obtaining risk probabilities based on the static risk probabilities and the dynamic risk probabilities, and executing corresponding strategies based on the risk probabilities.

[0020] Thirdly, the present invention provides a storage medium that stores one or more programs, which, when executed by a processor, implement the above-described cross-border electronic customs declaration inspection and early warning method based on dynamic risk learning.

[0021] Fourthly, the present invention provides an electronic device, the electronic device comprising a memory and a processor, wherein: the memory is used to store a computer program; and the processor is used to execute the computer program stored in the memory to implement the above-mentioned cross-border electronic customs declaration inspection and early warning method based on dynamic risk learning.

[0022] Compared with the prior art, the present invention has the following advantages: 1. By introducing dynamic features, the present invention can adjust the risk threshold in real time, adapt to the dynamic changes of different enterprises, different commodities and different time periods, and improve the accuracy and flexibility of early warning.

[0023] 2. This invention not only extracts numerical features from structured data, but also makes full use of semantic features from semi-structured data, achieving comprehensive analysis of customs declaration data.

[0024] 3. By extracting temporal behavior features and state transition features, this invention can monitor the dynamic changes of customs declaration behavior in real time, promptly detect abnormal declaration behavior, and prevent potential risks. Attached Figure Description

[0025] Figure 1This is a flowchart of a cross-border electronic customs declaration inspection and early warning method based on dynamic risk learning, proposed in an embodiment of the present invention.

[0026] Figure 2 This is a schematic diagram of the structure of a cross-border electronic customs declaration inspection and early warning system based on dynamic risk learning, as proposed in an embodiment of the present invention.

[0027] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but does not exclude other elements or objects.

[0029] like Figure 1 As shown, this invention proposes a cross-border electronic customs declaration inspection and early warning method based on dynamic risk learning. The method includes steps S101 to S104, wherein: step S101: acquire historical declaration data and real-time declaration data, the historical declaration data includes structured fields and semi-structured fields, and the real-time declaration data includes status flags, declaration timestamps and enterprise numbers.

[0030] It should be noted that this step first obtains the company's historical declaration data over a period of time through the data system. This data includes structured fields such as value, weight, and country of origin code, as well as semi-structured fields such as declared product name and description text. At the same time, real-time declaration data is obtained, which includes status markers (such as declared, released, etc.), declaration timestamps (accurate to a specific time point), and company number (used to identify company A).

[0031] Step S102: Extract static features from the structured and semi-structured fields, and extract dynamic features from the status marker, declaration timestamp, and enterprise number.

[0032] It should be noted that in this step, firstly, numerical features of value, weight, and country of origin code are extracted from the structured fields. The numerical features include the value dispersion coefficient and the weight skewness coefficient. Semantic features of the declared product name and description text are extracted from the semi-structured fields. The semantic features include TF-IDF word vectors and commodity HS code embedding. Then, all the numerical features and semantic features are concatenated to obtain static features.

[0033] Specifically, the coefficient of variation of the value of goods is calculated according to the following formula: ,in, The coefficient of variation is the value of the goods. The standard deviation of the declared value of goods by the same enterprise within 30 days This is the average value of goods declared by the same company over 30 days.

[0034] The weight skewness coefficient is calculated using the following formula: ;in, The weight skewness coefficient, Let n be the weight of the goods in the i-th declaration form, and n be the total number of declaration forms. The average weight of the goods in all declarations. This represents the standard deviation of the cargo weight for all declarations.

[0035] The TF-IDF word vectors are calculated using the following formula: ;in, For terms t In the document d TF-IDF value in For terms t The frequency of occurrence in historical declaration data For included terms t The number of documents, This represents the total number of documents in the historical declaration data.

[0036] The 5-bit HS code H is converted into a numerical vector and represented using a hybrid one-hot encoding and word embedding.

[0037] For example, taking a batch of customs declaration data for imported electronic products as an example, numerical features of value, weight, and country of origin are extracted from the structured fields. When calculating the value dispersion coefficient, assuming the standard deviation of the declared value for the same company within 30 days is 5000 yuan and the average is 50000 yuan, the value dispersion coefficient is calculated to be 0.1. For the weight skewness coefficient, assuming there are 100 declarations, the weight of goods in the i-th declaration is different for each declaration, the average weight of goods in all declarations is 100 kg, and the standard deviation is 10 kg, the weight skewness coefficient is calculated using a formula. In the semi-structured fields, semantic features are extracted from the declared product name and description text using the TF-IDF method. For example, the term "smartphone" appears 0.05 times in historical declaration data, the number of documents containing this term is 200, and the total number of documents in historical declaration data is 1000. The TF-IDF value of this term is calculated using a formula. Simultaneously, the 5-bit HS code is converted into a numerical vector and represented using a hybrid one-hot encoding and word embedding. Finally, all numerical and semantic features are concatenated to obtain static features.

[0038] In summary, by extracting numerical features (such as the value dispersion coefficient and weight skewness coefficient) from structured fields for value, weight, and country of origin codes, and semantic features (such as TF-IDF word vectors and HS code embeddings) from semi-structured fields for declared product names and descriptions, and concatenating these features, static features are obtained. This extraction method can fully utilize the structured and semi-structured information in historical declaration data, characterizing the static features of customs declaration data from multiple perspectives, both numerical and semantic. This provides rich evidence for subsequent accurate assessment of static risks and helps improve the accuracy of risk judgment.

[0039] In addition, in some embodiments, it is also necessary to extract time-series behavioral features and state transition features from the real-time reporting data, and connect all the time-series behavioral features and state transition features to obtain dynamic features; the state transition features include state dwell time and state loop count, and the time-series behavioral features include reporting frequency mutation index and reporting interval entropy.

[0040] Specifically, the frequency mutation index of applications is calculated according to the following formula: ;in, The mutation index is the frequency of applications. The number of times a company submits an application within the time period t. This represents the average number of submissions made by the company within the daily time period t. This represents the standard deviation of the number of declarations made by a company within a daily time period t.

[0041] The declaration interval entropy is calculated using the following formula: ;in, To report the interval entropy, Let k be the probability density of the i-th reporting time interval, and k be the number of time intervals.

[0042] The state dwell time is calculated using the following formula: ;in, The state dwell time, All are empirical coefficients. This refers to the average time from application submission to release in the past. This represents the standard deviation of the time taken from application to release in the historical application process.

[0043] The statistical reporting process progresses from submitted to revised and back to submitted frequency.

[0044] For example, taking a company's customs declaration data over a period of time as an example, in real-time declaration data, we extract time-series behavioral features and state transition features. For the declaration frequency mutation index in the time-series behavioral features, assuming the company makes 10 declarations within time period t (e.g., one day), the average number of declarations within that time period is 5, and the standard deviation is 2, the declaration frequency mutation index is calculated using a formula. Regarding the declaration interval entropy, assuming the declaration time interval is divided into multiple intervals, such as 0-10 minutes, 10-30 minutes, etc., the probability density of each interval is calculated. Assuming the probability density of the i-th interval is 0.2, and k is 5, the declaration interval entropy is calculated using a formula. In the state transition features, regarding the state dwell time, assuming the average time from declared to released is 24 hours, the standard deviation is 5 hours, and the empirical coefficients α and β are 1.5 and 0.5, the state dwell time is calculated using a formula. Simultaneously, the frequency of the declaration process from declared to under modification and back to declared is statistically analyzed. Finally, all the time-series behavioral features and state transition features are concatenated to obtain dynamic features.

[0045] Step S103: Construct a dual-channel deep forest model, and input the static features and the dynamic features into the dual-channel deep forest model respectively to obtain the static risk probability and the dynamic risk probability.

[0046] It should be noted that the dual-channel deep forest model is an innovative model based on the deep forest (gcForest) architecture. It constructs two independent feature processing channels (static feature channel and dynamic feature channel), extracts and fuses multi-dimensional features respectively, and finally comprehensively evaluates the risk probability.

[0047] Specifically, the risk probability is calculated using the following formula: ;in, For risk probability, These are static risk probability and dynamic risk probability, respectively. All are weighting coefficients.

[0048] By assigning weight coefficients to static and dynamic risk probabilities and calculating the overall risk probability according to a formula, this method can reasonably balance the impact of static and dynamic risk factors on the overall risk, adjust the importance of the two according to the actual situation, and make the risk probability calculation more accurate and in line with the actual needs of cross-border electronic customs declaration risk assessment.

[0049] Step S104: Obtain the risk probability based on the static risk probability and the dynamic risk probability, and execute the corresponding strategy based on the risk probability.

[0050] In some embodiments, the specific strategy is as follows: if the risk probability is greater than or equal to a first threshold, the application is suspended and a risk report containing the source of the abnormal features is generated; if the risk probability is less than the first threshold but greater than or equal to a second threshold, a manual review mark is added; if the risk probability is less than the second threshold, the application is approved.

[0051] For example, if the calculated risk probability is 0.4, the first threshold is set to 0.35, and the second threshold is set to 0.2. Since 0.4 ≥ 0.35, the declaration is suspended, and a risk report containing the source of the abnormal characteristics is generated. The report may indicate specific abnormal characteristics such as abnormal fluctuations in the value of the goods or abnormal declaration time intervals. If the risk probability is 0.25, which is less than the first threshold of 0.35 but greater than the second threshold of 0.2, a manual review mark is added, prompting customs officials to conduct a manual review of the declaration. If the risk probability is 0.15, which is less than the second threshold of 0.2, the declaration is released directly.

[0052] In summary, different strategies are implemented based on different risk probability thresholds, achieving tiered risk management. When the risk probability is high (greater than or equal to the first threshold), the declaration is suspended and a risk report containing traceability of abnormal characteristics is generated. This promptly prevents potentially risky declarations and provides a basis for subsequent processing through detailed risk reports. When the risk probability is at a medium level (less than the first threshold but greater than or equal to the second threshold), a manual review flag is added. This ensures effective monitoring of potential risks while avoiding excessive intervention in normal declarations. When the risk probability is low (less than the second threshold), the declaration is directly released, improving customs clearance efficiency. This tiered management strategy optimizes the cross-border electronic customs declaration process and improves overall operational efficiency while ensuring security.

[0053] In summary, the cross-border electronic customs declaration inspection and early warning method based on dynamic risk learning proposed in this invention addresses the challenge of traditional cross-border electronic customs declaration inspections, which often rely on fixed thresholds and single features for anomaly detection, making them ill-suited to the complex and ever-changing trade environment. This invention extracts features from historical and real-time declaration data across multiple dimensions, encompassing not only numerical features such as value and weight in structured fields but also semantic features in semi-structured fields, such as TF-IDF word vectors and commodity HS code embeddings. This allows the model to more comprehensively understand the inherent meaning of the customs declaration data. Furthermore, this invention innovatively introduces dynamic features, such as the declaration frequency mutation index, declaration interval entropy, and state dwell time. These features reflect the dynamic changes in customs declaration behavior in real time, effectively capturing subtle clues of abnormal declarations. By constructing a dual-channel deep forest model and inputting static and dynamic features into the model for training, the model learns complex patterns in the data and accurately outputs static and dynamic risk probabilities. Finally, the risk probability is obtained by combining these two probabilities, and a corresponding early warning strategy is implemented. This process not only improves the accuracy of early warnings but also enhances the model's generalization ability, enabling it to cope with various new and covert anomaly patterns. Therefore, this invention, through a dynamic risk learning method, achieves intelligent and comprehensive early warning for cross-border electronic customs declaration inspections, providing a solid guarantee for the security and efficiency of international trade.

[0054] like Figure 2 As shown, an embodiment of the present invention also proposes a cross-border electronic customs declaration inspection and early warning system based on dynamic risk learning. The system includes: a data acquisition module 10, used to acquire historical declaration data and real-time declaration data. The historical declaration data includes structured fields and semi-structured fields, and the real-time declaration data includes status markers, declaration timestamps and enterprise numbers.

[0055] The feature extraction module 20 is used to extract static features from the structured fields and semi-structured fields, and to extract dynamic features from the status marker, declaration timestamp and enterprise number.

[0056] The risk probability prediction module 30 is used to construct a dual-channel deep forest model and input the static features and the dynamic features into the dual-channel deep forest model to obtain the static risk probability and the dynamic risk probability.

[0057] The early warning execution module 40 is used to obtain the risk probability based on the static risk probability and the dynamic risk probability, and execute the corresponding strategy based on the risk probability.

[0058] In another aspect, the present invention also proposes a storage medium on which one or more programs are stored, which, when executed by a processor, implement the above-described cross-border electronic customs declaration inspection and early warning method based on dynamic risk learning.

[0059] In another aspect, the present invention also proposes an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so as to realize the above-mentioned cross-border electronic customs declaration inspection and early warning method based on dynamic risk learning.

[0060] Those skilled in the art will understand that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain stored, communicated, propagated, or transmitted programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0061] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0062] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0063] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the invention as set forth in the claims. Furthermore, the invention described herein may have other embodiments and can be implemented or carried out in various ways.

Claims

1. A cross-border electronic customs inspection early warning method based on dynamic risk learning, characterized in that, The method only consists of the following steps: obtaining historical declaration data and real-time declaration data, the historical declaration data including structured fields and semi-structured fields, and the real-time declaration data including state marks, declaration time stamps and enterprise numbers; the real-time declaration data is obtained by real-time collection through a customs system server; static features are extracted from the structured fields and semi-structured fields, and dynamic features are extracted from the state marks, declaration time stamps and enterprise numbers; numerical features of cargo value, weight and origin coding are extracted from the structured fields, the numerical features including cargo value dispersion coefficient and weight skewness coefficient, and semantic features of declared commodity names and description texts are extracted from the semi-structured fields, the semantic features including TF-IDF word vectors and commodity HS coding embedding, all the numerical features and semantic features are connected to obtain static features; The value dispersion coefficient is calculated according to the following formula: ; wherein, is the value dispersion coefficient, is the standard deviation of the declared value of the same enterprise within 30 days, is the average value of the declared value of the same enterprise within 30 days; The weight skewness coefficient is calculated according to the following formula: ; wherein, is the weight skewness coefficient, is the weight of the goods in the ith declaration form, and n is the total number of declaration forms, is the average value of the weight of the goods of all declaration forms, is the standard deviation of the weight of the goods of all declaration forms; The TF-IDF word vector is calculated according to the following formula: ; wherein, is the value of the word item t in the document d, is the frequency of occurrence of the word item t in the historical filing data, is the number of documents containing the word item t, is the total number of documents in the historical filing data;​ 5-bit HS coding H is converted into a numerical vector, and a one-hot coding and word embedding hybrid representation is adopted; time sequence behavior features and state transition features are extracted from the real-time declaration data, and all the time sequence behavior features and state transition features are connected to obtain dynamic features; the state transition features include state residence time and state loop number, and the time sequence behavior features include declaration frequency mutation index and declaration interval entropy; The declaration frequency mutation index is calculated according to the following formula: ; wherein, is the declaration frequency mutation index, is the number of declarations of the enterprise in the t time period, is the average number of declarations of the enterprise in the t time period of each day, is the standard deviation of the number of declarations of the enterprise in the t time period of each day. The declaration interval entropy is calculated according to the following formula: ; where H is the declaration interval entropy, is the probability density of the ith declaration time interval, and k is the number of time intervals. The state residence time is calculated according to the following formula: ; wherein, is the state residence time, are empirical coefficients, is the average time consumption from declared to released in historical declaration process, is the standard deviation of time consumption from declared to released in historical declaration process; the frequency of the declaration process from having been declared to being modified and then to having been declared is counted; a double-channel deep forest model is constructed, and the static features and the dynamic features are input into the double-channel deep forest model to obtain static risk probability and dynamic risk probability; risk probability is obtained according to the static risk probability and the dynamic risk probability, and a corresponding strategy is executed according to the risk probability.

2. The method of claim 1, wherein, The step of obtaining risk probability according to the static risk probability and the dynamic risk probability includes: The risk probability is calculated according to the following formula: ; wherein, is the risk probability, are respectively the static risk probability, the dynamic risk probability, are all weight coefficients. 3.The method of claim 2, wherein, The step of executing a corresponding strategy according to risk probability includes: if the risk probability is greater than or equal to a first threshold value, declaration suspension is executed, and a risk report containing abnormal feature tracing is generated; if the risk probability is less than the first threshold value and greater than or equal to a second threshold value, an artificial review mark is added; if the risk probability is less than the second threshold value, the declaration is released.

4. A cross-border electronic customs inspection early warning system based on dynamic risk learning, characterized in that, The system is used to implement the method of claim 1, and only consists of the following modules: a data acquisition module for obtaining historical declaration data and real-time declaration data, the historical declaration data including structured fields and semi-structured fields, and the real-time declaration data including state marks, declaration time stamps and enterprise numbers; a feature extraction module for extracting static features from the structured fields and semi-structured fields, and extracting dynamic features from the state marks, declaration time stamps and enterprise numbers; numerical features of cargo value, weight and origin coding are extracted from the structured fields, the numerical features including cargo value dispersion coefficient and weight skewness coefficient, and semantic features of declared commodity names and description texts are extracted from the semi-structured fields, the semantic features including TF-IDF word vectors and commodity HS coding embedding; all the numerical features and semantic features are connected to obtain static features; The value dispersion coefficient is calculated according to the following formula: ; wherein, is the value dispersion coefficient, is the standard deviation of the declared value of the same enterprise within 30 days, is the average value of the declared value of the same enterprise within 30 days; The weight skewness coefficient is calculated according to the following formula: ; wherein, is the weight skewness coefficient, is the weight of the goods in the ith declaration form, and n is the total number of declaration forms, is the average value of the weight of the goods of all declaration forms, is the standard deviation of the weight of the goods of all declaration forms; The TF-IDF word vectors are calculated using the following formula: ;in, For terms t In the document d The TF-IDF value in For terms t The frequency of occurrence in historical declaration data For included terms t The number of documents, This represents the total number of documents in the historical declaration data; The HS code H in the 5th position is converted into a numerical vector using a one-hot encoding and a word embedding mixed representation; extracting time sequence behavior features and state transition features from the real-time declaration data, connecting all the time sequence behavior features and the state transition features to obtain dynamic features, the state transition features including state residence time and state loop times, and the time sequence behavior features including declaration frequency mutation index and declaration interval entropy; The declaration frequency mutation index is calculated according to the following formula: ; wherein, is the declaration frequency mutation index, is the number of declarations of the enterprise in the t time period, is the average number of declarations of the enterprise in the t time period per day, is the standard deviation of the number of declarations of the enterprise in the t time period per day; The declaration interval entropy is calculated according to the following formula: ; is the declaration interval entropy, is the probability density of the ith declaration time interval interval, and k is the number of time interval intervals. The state residence time is calculated according to the following formula: ; wherein, is the state residence time, are empirical coefficients, is the average time consumption from declared to released in the historical declaration process, is the standard deviation of the time consumption from declared to released in the historical declaration process; counting the frequency of the declaration process from having declared to modifying to having declared again; a risk probability prediction module configured to construct a double-channel deep forest model, input the static features and the dynamic features into the double-channel deep forest model respectively, and obtain static risk probability and dynamic risk probability; an early warning execution module configured to obtain risk probability according to the static risk probability and the dynamic risk probability, and execute corresponding strategies according to the risk probability.

5. A storage medium, characterized by The storage medium stores one or more programs, which are executed by the processor to implement the cross-border electronic declaration inspection early warning method based on dynamic risk learning according to any one of claims 1-3. 6.An electronic device, comprising a memory and a processor, wherein: the memory is configured to store a computer program; the processor is configured to execute the computer program stored on the memory to implement the cross-border electronic declaration inspection early warning method based on dynamic risk learning according to any one of claims 1-3.

Images