Multi-modal data fusion-based import and export risk real-time early warning method and system

Abstract

The invention discloses an import and export risk real-time early warning method and system based on multi-modal data fusion, and relates to the technical field of data analysis, and the method comprises the steps: obtaining multi-modal data in real time, and carrying out the preprocessing; designing a dynamic weight distribution algorithm, and endowing different types of data with differentiated weights to realize dynamic fusion according to quantitative indexes of timeliness, authority and target transaction association degree; fusion data features are extracted through a streaming computation engine, real-time risk calculation is carried out based on an Apache Flink cluster, and an association network is constructed in combination with a six-dimensional knowledge graph; on the basis of the extracted features and the association network, a CEP rule engine quickly identifies and calculates potential risk events and accurately locates risk points by relying on a predefined multi-class composite risk mode; and setting a grading early warning mechanism, enabling the identified risk mode to correspond to a corresponding early warning grade, generating a grading early warning signal, and synchronously prompting a dynamic disposal suggestion. According to the method, the risk identification accuracy can be improved, and the early warning response delay is shortened.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data analysis, in particular to an import and export risk real-time early warning method and system based on multi-modal data fusion. BACKGROUND

[0002] The existing technology for judging the import and export risks of enterprises has the following defects:

[0003] First, the data dimension is single. The traditional system relies on customs declaration data and does not integrate multi-modal data such as logistics, exchange rate, public opinion, etc., which leads to the inability to identify the risk of broken capital chain caused by exchange rate fluctuations and the tendency to overlook the policy transmission effect of social media public opinion.

[0004] Second, the response delay is serious. Based on the T+1 data batch processing mode, the average early warning delay is more than 4 hours, which cannot respond to sudden risks.

[0005] Third, the false positive rate is high. The rule engine relies on human experience, and the misjudgment rate of complex risks reaches 32%.

[0006] The current industry demand pain points mainly manifest as follows:

[0007] At the enterprise level, according to the statistics of China Trade Promotion Association, more than 20 billion US dollars of export order loss was caused by the lack of risk early warning in 2023.

[0008] At the technical level, the existing technology still has problems such as low efficiency of multi-source data fusion and insufficient real-time performance. SUMMARY

[0009] In view of the problems of data dimension limitation, insufficient response timeliness, high misjudgment rate and lack of decision support in the prior art, the present application provides an import and export risk real-time early warning method and system based on multi-modal data fusion to improve the risk identification accuracy, shorten the response delay and reduce the false positive rate.

[0010] In the first aspect, the present application provides an import and export risk real-time early warning method based on multi-modal data fusion, and the technical solutions adopted to solve the above technical problems are as follows:

[0011] An import and export risk real-time early warning method based on multi-modal data fusion includes the following steps:

[0012] S1, real-time acquisition of multi-modal data, including customs declaration data, logistics time series data, international freight rate data, RMB exchange rate and public opinion text data;

[0013] S2, preprocessing the acquired multi-modal data, including standardization of structured data, extraction of key entities of unstructured text and extraction of time series data trend features;

[0014] S3, design a dynamic weight allocation algorithm based on reinforcement learning, assign different weights to different types of data according to the quantitative indicators of multi-modal data in timeliness, authority and correlation with target transactions, and realize accurate dynamic fusion of multi-modal data;

[0015] S4, feature extraction of the fused real-time data by the streaming computing engine, focusing on risk feature capture at a second-level granularity; rely on the distributed computing cluster built by Apache Flink to carry out real-time risk calculation, and combine the six-dimensional knowledge graph to build a "enterprise-goods-port-freight rate-exchange rate-policy" correlation network, providing a more comprehensive and deep analysis basis for real-time risk calculation;

[0016] S5, based on the features extracted in step S4 and the correlation network built, the CEP rule engine relies on pre-defined multi-class composite risk patterns to quickly identify and calculate potential risk events for multi-modal fusion data after dynamic weight allocation, accurately positioning risk points;

[0017] S6, set up a three-level early warning mechanism, establish an intelligent decision output mechanism, according to the preset warning level trigger condition, correspond the specific risk patterns identified by the CEP rule engine to the corresponding warning level, generate a graded warning signal, and simultaneously prompt a dynamic disposal suggestion, realize the visualization of risk and the operability of disposal.

[0018] Optionally, the specific operation of step S1 of acquiring multi-modal data includes:

[0019] (1a) interface with the customs data interface, receive raw data in real time according to the data transmission protocol, extract structured information containing HS code, declared value and proof of origin field through pre-set parsing rules, ensure the matching of data fields and business needs, and form a standardized customs declaration data set;

[0020] (1b) access the ship AIS positioning system and truck GPS trajectory monitoring platform, establish a continuous data receiving channel, clean and associate the obtained original time series data containing geographic location and timestamp, generate dynamic indicators including transportation delay rate and port congestion index, and build a complete logistics time series database;

[0021] (1c) integrate the data interfaces of the Baltic freight index release channel and the China-Europe block train freight information platform, regularly grab real-time freight data, establish a storage and update mechanism combined with historical data, and form an international freight data set;

[0022] (1d) Through the specified open API interface, the real-time exchange rate data and the forward exchange rate curve of RMB are obtained at a set frequency, the data is format-verified and timestamp-aligned, the accuracy and timeliness of the data are ensured, and the exchange rate database supporting the risk prediction of cross-border settlement is established;

[0023] (1e) The discussion content related to compliance policies is collected from public channels through web crawler technology, the public opinion text data is obtained, the BERT model is used to perform sentiment analysis and keyword extraction on the obtained public opinion text data, and a structured public opinion analysis data set is formed.

[0024] Further optionally, the preprocessing operation of the multi-modal data in step S2 specifically includes:

[0025] (2a) For the three types of structured data of customs declaration data, international freight data and RMB exchange rate, regular expression matching is used to correct the field format that does not conform to the standard, and the rule engine is used to unify the data standard, so as to ensure that the formats of various structured data are consistent and the field meanings are unified;

[0026] (2b) For the public opinion text type unstructured data, named entity recognition technology is used to deeply analyze the text content and extract the key entity information contained therein, so as to convert the unstructured text into structured entity data that can be used for analysis;

[0027] (2c) For the logistics time series data, a sliding window with a window size of 5 minutes and a sliding step of 1 minute is set, the continuous logistics time series data is intercepted and calculated through the sliding window, and the logistics delay trend characteristics in different time periods are extracted to clearly reflect the change rule and trend of logistics delay.

[0028] Further optionally, step S3 is specifically implemented as follows: different weights are given to different types of data to realize precise and dynamic fusion of multi-modal data:

[0029] S3.1, taking the preprocessed multi-modal data as the processing object, the performance of each type of data in timeliness, authority and association degree with the target transaction is quantified respectively, and timeliness coefficient, authority coefficient and association degree coefficient that can be input into the dynamic weight distribution algorithm are formed;

[0030] S3.2, a dynamic weight distribution algorithm based on reinforcement learning is constructed, the timeliness coefficient, authority coefficient and association degree coefficient of each type of data are taken as input parameters of the dynamic weight distribution algorithm, and through the intelligent decision mechanism of reinforcement learning, a dynamic mapping relationship between data features and weight distribution is established;

[0031] S3.3, different weights are given to different types of data by a dynamic weight allocation algorithm, and each type of multi-modal data is further weighted and fused;

[0032] S3.4, the performance of the fused data in actual business application is monitored and fed back in real time through the reinforcement learning mechanism, and when it is found that the weight allocation of a certain type of data does not match the actual business value, the parameters and decision logic of the dynamic weight allocation algorithm are adjusted to optimize the weight allocation result.

[0033] Further optionally, the steps S4 specifically include:

[0034] S4.1, after the multi-modal data is preprocessed and dynamically fused, the real-time data after fusion is feature extracted through a streaming computing engine, focusing on risk feature capture at a second-level granularity;

[0035] S4.2, a distributed computing cluster built on Apache Flink is used to perform real-time risk calculation on the associated subjects that need to be monitored in the business scenario;

[0036] S4.3, a knowledge graph is built around the six core entity types of "enterprise-goods-port-freight rate-exchange rate-policy", and the key relationships and attributes between entities are clarified; the knowledge graph converts the scattered information in the multi-modal data into a structured associated network, providing deep logical support between entities for risk calculation, complementing the correlation quantification indicators of multi-modal data, and enhancing the comprehensiveness and accuracy of risk analysis.

[0037] In a second aspect, the application provides an import and export risk real-time early warning system based on multi-modal data fusion, which solves the above technical problems by adopting the following technical solutions:

[0038] An import and export risk real-time early warning system based on multi-modal data fusion includes:

[0039] A multi-modal data acquisition module is used to acquire multi-modal data in real time, including customs declaration data, logistics time series data, international freight rate data, RMB exchange rate and public opinion text data;

[0040] A data preprocessing module is used to preprocess the acquired multi-modal data, including structured data standardization, unstructured text key entity extraction and time series data trend feature extraction;

[0041] A data weighted fusion module is used to design a dynamic weight allocation algorithm based on reinforcement learning, and different weights are given to different types of data according to the quantitative indicators of multi-modal data in timeliness, authority and correlation with target transactions, to realize accurate dynamic fusion of multi-modal data;

[0042] The feature extraction and network construction module is used for, on the one hand, extracting features from the fused real-time data by the stream computing engine, focusing on capturing risk features at a second-level granularity; and on the other hand, developing real-time risk calculation based on a distributed computing cluster constructed by Apache Flink, and combining a six-dimensional knowledge graph to construct a correlation network of "enterprise-goods-port-freight rate-exchange rate-policy", thereby providing a more comprehensive and deep analysis basis for real-time risk calculation;

[0043] The risk identification and positioning module is used for, based on the features extracted by the feature extraction and network construction module and the correlation network constructed, using a CEP rule engine to rely on a plurality of predefined composite risk patterns to quickly identify and calculate potential risk events from the multi-modal fusion data after dynamic weight distribution, and accurately locate risk points;

[0044] The early warning and suggestion generation module is used for setting up a three-level early warning mechanism, establishing an intelligent decision output mechanism, and according to a preset early warning level triggering condition, corresponding the specific risk pattern identified by the CEP rule engine to the corresponding early warning level, generating a graded early warning signal, and synchronously prompting a dynamic disposal suggestion, thereby realizing visual presentation and operable disposal of risks.

[0045] Optionally, the multi-modal data acquisition module involved obtains the customs declaration data, logistics time series data, international freight rate data, RMB exchange rate and public opinion text data through the following operations:

[0046] (1a) connecting a customs data interface, receiving raw data in real time according to a data transmission protocol, extracting structured information containing HS code, declared value and proof of origin field through a preset parsing rule, ensuring the matching of data fields and business needs, and forming a standardized customs declaration data set;

[0047] (1b) accessing a ship AIS positioning system and a truck GPS trajectory monitoring platform, establishing a continuous data receiving channel, cleaning and correlation calculating the obtained raw time series data containing geographical position and time stamp, generating dynamic indicators containing transportation delay rate and port congestion index, and constructing a complete logistics time series database;

[0048] (1c) integrating data interfaces of the Baltic freight index release channel and the China-Europe block train freight information platform, regularly grabbing real-time freight data, combining historical data to establish a storage and update mechanism, and forming an international freight rate data set;

[0049] (1d) acquiring real-time RMB exchange rate data and forward exchange rate curve at a set frequency through a specified open API interface, performing format verification and time stamp alignment processing on the data, ensuring the accuracy and timeliness of the data, and establishing an exchange rate database supporting cross-border settlement risk prediction;

[0050] (1e) Collecting discussion content related to compliance policies from public channels through web crawler technology, obtaining public opinion text data, using BERT model to perform sentiment analysis and keyword extraction on the obtained public opinion text data, and forming a structured public opinion analysis dataset.

[0051] Further optionally, the data preprocessing module involved performs the following preprocessing operations on the obtained customs declaration data, logistics timing data, international freight data, RMB exchange rate and public opinion text data:

[0052] (2a) For the three types of structured data of customs declaration data, international freight data and RMB exchange rate, regular expression matching is used to correct the field format that does not conform to the standard, and the rule engine is used to unify the data standard, to ensure that the formats of various structured data are consistent and the field meanings are unified;

[0053] (2b) For public opinion text unstructured data, named entity recognition technology is used to deeply analyze the text content and extract key entity information contained therein, so as to convert unstructured text into structured entity data that can be used for analysis;

[0054] (2c) For logistics timing data, a sliding window with a window size of 5 minutes and a sliding step of 1 minute is set, and the continuous logistics timing data is intercepted and calculated through the sliding window to extract the logistics delay trend characteristics in different time periods, so as to clearly reflect the change rule and trend of logistics delay.

[0055] Further optionally, the data weighting fusion module involved specifically includes:

[0056] The data feature quantization unit is used to quantize the performance of each type of data in terms of timeliness, authority and target transaction association degree based on the preprocessed multi-modal data, to form timeliness coefficient, authority coefficient and association degree coefficient that can be input into the dynamic weight allocation algorithm;

[0057] The dynamic weight intelligent allocation unit is used to construct a dynamic weight allocation algorithm based on reinforcement learning, taking the timeliness coefficient, authority coefficient and association degree coefficient of each type of data as the core input parameters of the dynamic weight allocation algorithm, and establishing a dynamic mapping relationship between data features and weight allocation through the intelligent decision mechanism of reinforcement learning;

[0058] The dynamic weighted fusion calculation unit is used to calculate the differential weight for different types of data through the dynamic weight allocation algorithm, and further to perform weighted fusion on various multi-modal data;

[0059] The reinforcement learning weight optimization unit is used to monitor and provide feedback on the performance of fused data in actual business applications in real time through the reinforcement learning mechanism. When it is found that the weight distribution of a certain type of data does not match the actual business value, the parameters and decision logic of the dynamic weight distribution algorithm are readjusted to optimize the weight distribution results.

[0060] Optionally, the feature extraction and network construction modules involved specifically include:

[0061] The feature extraction unit is used to extract features from the fused real-time data after preprocessing and dynamic fusion of multimodal data through a streaming computing engine, focusing on capturing risk features at a granularity of seconds.

[0062] The risk calculation unit is used to perform real-time risk calculations on the associated entities that need to monitor risks in business scenarios using a distributed computing cluster built on Apache Flink.

[0063] The graph construction unit is used to construct a knowledge graph around the six core entity types of "enterprise-commodity-port-freight rate-exchange rate-policy" to clarify the key relationships and attributes between entities; the constructed knowledge graph transforms the scattered information in multimodal data into a structured association network, providing deep logical support between entities for risk calculation, complementing the quantitative correlation indicators of multimodal data, and enhancing the comprehensiveness and accuracy of risk analysis.

[0064] The present invention provides a real-time import and export risk warning method and system based on multimodal data fusion, which has the following beneficial effects compared with the prior art:

[0065] 1. This invention can break through the reliance on a single data source and achieve cross-modal integration of customs, logistics, exchange rate, and public opinion data; shorten risk warning delays from hours to within 30 seconds; reduce false alarm rates through composite risk pattern recognition, and provide dynamic disposal recommendations, providing real-time intelligent risk control support for cross-border e-commerce platforms, foreign trade supply chain companies, and cross-border financial institutions;

[0066] 2. The present invention is particularly suitable for the following scenarios: (a) cross-border e-commerce platforms: real-time monitoring of commodity export compliance risks and logistics anomalies; (b) foreign trade supply chain enterprises: dynamic assessment of order fulfillment risks and capital chain security; (c) cross-border financial institutions: real-time credit assessment for letter of credit opening and cross-border settlement. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Attachment Figure 1 is a flow chart of a method according to embodiment 1 of the present invention;

[0068] Attachment Figure 2 This is a module connection block diagram of the second embodiment of the present invention. DETAILED DESCRIPTION

[0069] In order to make the technical solutions, the technical problems solved and the technical effects of the present application clearer, the technical solutions of the present application are described clearly and completely below in combination with specific embodiments.

[0070] Embodiment one:

[0071] Referring to the accompanying drawings, Figure 1 The present embodiment proposes a real-time early warning method for import and export risks based on multi-modal data fusion, which includes the following steps:

[0072] S1, real-time acquisition of multi-modal data, including customs declaration data, logistics time series data, international freight rate data, RMB exchange rate and public opinion text data; the specific operation for realizing this process includes:

[0073] (1a) Docking with the customs data interface, real-time receiving of raw data according to the data transmission protocol, extracting the structured information containing HS code, declared value and proof of origin field through the preset analysis rule, ensuring the matching of data field and business demand, forming a standardized customs declaration data set;

[0074] (1b) Access to ship AIS positioning system and truck GPS track monitoring platform, establish a continuous data receiving channel, clean and associate the obtained raw time series data containing geographic location and time stamp, generate dynamic indicators containing transportation delay rate and port congestion index, and build a complete logistics time series database;

[0075] (1c) Integrate the data interface of the Baltic freight index (FBX) release channel and the data interface of the China-Europe train freight information platform, regularly grab real-time freight data, combine historical data to establish a storage and update mechanism, and form an international freight data set;

[0076] (1d) Through the specified open API interface, get the real-time exchange rate data and forward exchange rate curve of RMB according to the set frequency, perform format verification and time stamp alignment processing on the data to ensure the accuracy and timeliness of the data, and establish an exchange rate database supporting cross-border settlement risk prediction;

[0077] (1e) Collect the discussion content related to compliance policy from the public channel through the network crawler technology, obtain the public opinion text data, use the BERT model to perform sentiment analysis and keyword extraction on the obtained public opinion text data, and form a structured public opinion analysis data set.

[0078] S2, preprocessing of the obtained multi-modal data, including standardization of structured data, extraction of key entities from unstructured text and extraction of time series data trend features; the specific operation for realizing this process includes:

[0079] (2a) For the three types of structured data: customs declaration data, international freight data, and RMB exchange rate, regular expression matching is used to correct the field formats that do not conform to the standard, and a rule engine is used to unify the data standards, ensuring that the formats of various structured data are consistent and the meanings of the fields are unified;

[0080] (2b) For public opinion text unstructured data, named entity recognition technology is used to deeply analyze the text content and extract key entity information, converting unstructured text into structured entity data that can be used for analysis;

[0081] (2c) For logistics time series data, a sliding window with a window size of 5 minutes and a sliding step of 1 minute is set up to intercept and calculate continuous logistics time series data, extracting logistics delay trend features in different time periods to clearly reflect the change pattern and trend of logistics delay.

[0082] S3, design a dynamic weight allocation algorithm based on reinforcement learning, assign different weights to different types of data according to the quantitative indicators of multi-modal data in timeliness, authority, and relevance to target transactions, and realize precise dynamic fusion of multi-modal data; The specific operation to achieve this process includes:

[0083] S3.1, taking the pre-processed multi-modal data as the processing object, quantifying the performance of each type of data in timeliness, authority, and relevance to target transactions, forming timeliness coefficients, authority coefficients, and relevance coefficients that can be input into the dynamic weight allocation algorithm;

[0084] S3.2, construct a dynamic weight allocation algorithm based on reinforcement learning, use the timeliness coefficients, authority coefficients, and relevance coefficients of various data as input parameters of the dynamic weight allocation algorithm, and through the intelligent decision-making mechanism of reinforcement learning, establish a dynamic mapping relationship between data features and weight allocation;

[0085] S3.3, calculate the different weights for different types of data through the dynamic weight allocation algorithm, and further weight and fuse various multi-modal data;

[0086] S3.4, through the reinforcement learning mechanism, real-time monitoring and feedback of the performance of the fused data in actual business applications, readjust the parameters and decision logic of the dynamic weight allocation algorithm when it is found that the weight allocation of a certain type of data does not match the actual business value, and optimize the weight allocation result.

[0087] S4. Feature extraction is performed on the integrated real-time data through a streaming computing engine, focusing on capturing risk features at a granularity of seconds. Real-time risk calculation is performed using a distributed computing cluster built on Apache Flink. A six-dimensional knowledge graph is used to construct a "company-commodity-port-freight-rate-exchange-policy" association network, providing a more comprehensive and in-depth analysis basis for real-time risk calculation. The specific operations to achieve this process include:

[0088] S4.1. After preprocessing and dynamic fusion of multimodal data, feature extraction is performed on the fused real-time data using a streaming computing engine, focusing on capturing risk features at a granularity of seconds.

[0089] S4.2, a distributed computing cluster built on Apache Flink, performs real-time risk calculations for associated entities that need to monitor risks in business scenarios;

[0090] S4.3. Construct a knowledge graph around the six core entity types of "enterprise-commodity-port-freight rate-exchange rate-policy" to clarify the key relationships and attributes between entities. For example, the relationship of "commodity-transportation via-port" includes attributes such as estimated time and actual delay rate. The relationship of "port-controlled by-policies and regulations" is annotated with a policy sensitivity score, and the relationship of "exchange rate-impact-freight rate" is quantified by the cointegration coefficient. The knowledge graph transforms the scattered information in multimodal data into a structured association network, providing deep logical support between entities for risk calculation, complementing the quantitative correlation indicators of multimodal data, and enhancing the comprehensiveness and accuracy of risk analysis.

[0091] S5. Based on the features extracted in step S4 and the association network constructed, the CEP (Complex Event Processing) rule engine relies on predefined multiple types of composite risk patterns to quickly identify and calculate potential risk events on the multimodal fusion data after dynamic weight assignment, and accurately locate risk points.

[0092] S6. Establish a three-level warning mechanism (red, yellow, and green), and establish an intelligent decision-making output mechanism. According to the preset warning level trigger conditions, the specific risk patterns identified by the CEP (Complex Event Processing) rule engine are mapped to the corresponding warning level, generating graded warning signals and simultaneously prompting dynamic disposal suggestions (such as alternative port recommendations and exchange rate lock-in period calculations) to achieve visual presentation of risks and operational disposal.

[0093] Example 2:

[0094] Reference Attachment Figure 2 This embodiment proposes a real-time early warning system for import and export risks based on multimodal data fusion, which includes:

[0095] a multi-modal data acquisition module, configured to acquire multi-modal data in real time, including customs declaration data, logistics timing data, international freight data, RMB exchange rate data, and public opinion text data;

[0096] a data preprocessing module, configured to preprocess the acquired multi-modal data, including standardization of structured data, key entity extraction of unstructured text, and trend feature extraction of timing data;

[0097] a data weighted fusion module, configured to design a dynamic weight distribution algorithm based on reinforcement learning, assign different weights to different types of data according to quantitative indicators of multi-modal data in terms of timeliness, authority, and relevance to target transactions, and realize precise dynamic fusion of multi-modal data;

[0098] a feature extraction and network construction module, configured to perform feature extraction on the fused real-time data through a streaming computing engine, focus on risk feature capture at a second-level granularity, and rely on a distributed computing cluster constructed based on Apache Flink to perform real-time risk calculation, and combine a six-dimensional knowledge graph to construct a "enterprise-goods-port-freight-rate-exchange rate-policy" correlation network, to provide a more comprehensive and deep analysis basis for real-time risk calculation;

[0099] a risk identification and positioning module, configured to perform rapid identification and calculation of potential risk events based on the features extracted by the feature extraction and network construction module and the correlation network constructed by the feature extraction and network construction module, and rely on a pre-defined multi-class composite risk pattern to perform rapid identification and calculation of potential risk events based on the multi-modal fusion data after dynamic weight distribution, and accurately locate risk points;

[0100] a warning and suggestion generation module, configured to set up a three-level warning mechanism, establish an intelligent decision output mechanism, trigger conditions according to pre-set warning levels, correspond specific risk patterns identified by the CEP (Complex Event Processing) rule engine to corresponding warning levels, generate graded warning signals, and simultaneously prompt dynamic disposal suggestions, to realize visual presentation and operable disposal of risks.

[0101] In this embodiment, the multi-modal data acquisition module acquires customs declaration data, logistics timing data, international freight data, RMB exchange rate data, and public opinion text data through the following operations:

[0102] (1a) Access the customs data interface in real time to receive raw data according to the data transmission protocol, extract structured information containing HS code, declared value and proof of origin field through preset parsing rules, ensure the matching of data fields and business needs, and form a standardized customs declaration data set;

[0103] (1b) Access the ship AIS positioning system and truck GPS trajectory monitoring platform to establish a continuous data receiving channel, clean and correlate the obtained raw time series data containing geographic location and timestamp, generate dynamic indicators including transportation delay rate and port congestion index, and build a complete logistics time series database;

[0104] (1c) Integrate the data interfaces of the Baltic Freight Index release channel and the China-Europe train freight information platform to regularly capture real-time freight data, combine historical data to establish a storage and update mechanism, and form an international freight data set;

[0105] (1d) Through the specified open API interface, obtain the real-time exchange rate data and forward exchange rate curve of the RMB at a set frequency, perform format verification and timestamp alignment processing on the data to ensure its accuracy and timeliness, and establish an exchange rate database to support cross-border settlement risk prediction;

[0106] (1e) Collect discussion content related to compliance policies from public channels through web crawling technology to obtain public opinion text data, use the BERT model to perform sentiment analysis and keyword extraction on the obtained public opinion text data, and form a structured public opinion analysis data set.

[0107] In this embodiment, the data preprocessing module involved in the customs declaration data, logistics time series data, international freight data, RMB exchange rate and public opinion text data performs the following preprocessing operations:

[0108] (2a) For the three types of structured data of customs declaration data, international freight data and RMB exchange rate, use regular expression matching to correct the field format that does not conform to the standard, and unify the data standard through the rule engine to ensure that the formats of various structured data are consistent and the field meanings are unified;

[0109] (2b) For public opinion text unstructured data, use named entity recognition technology to deeply analyze the text content and extract key entity information contained therein, and convert unstructured text into structured entity data that can be used for analysis;

[0110] (2c) For logistics time series data, set the window size to 5 minutes and the sliding step to 1 minute, and use the sliding window to intercept and calculate the continuous logistics time series data to extract the logistics delay trend characteristics in different time periods, so as to clearly reflect the change rule and trend of logistics delay.

[0111] In this embodiment, the data weighting fusion module specifically includes:

[0112] The data feature quantization unit is configured to take the preprocessed multi-modal data as a processing object, quantize the performance of each type of data in terms of timeliness, authority and association degree with the target transaction, and form timeliness coefficients, authority coefficients and association degree coefficients that can be input into a dynamic weight distribution algorithm.

[0113] The dynamic weight intelligent distribution unit is configured to construct a dynamic weight distribution algorithm based on reinforcement learning, take the timeliness coefficients, authority coefficients and association degree coefficients of various types of data as core input parameters of the dynamic weight distribution algorithm, and establish a dynamic mapping relationship between data features and weight distribution through an intelligent decision mechanism of reinforcement learning.

[0114] The dynamic weighting fusion calculation unit is configured to calculate different weights for different types of data through the dynamic weight distribution algorithm, and further perform weighted fusion on various multi-modal data.

[0115] The reinforcement learning weight optimization unit is configured to monitor and feedback the performance of the fused data in actual business applications in real time through the reinforcement learning mechanism, readjust the parameters and decision logic of the dynamic weight distribution algorithm when it is found that the weight distribution of a certain type of data does not match the actual business value, and optimize the weight distribution result.

[0116] In this embodiment, the feature extraction and network construction module specifically includes:

[0117] The feature extraction unit is configured to perform feature extraction on the fused real-time data through a streaming computing engine after the multi-modal data is preprocessed and dynamically fused, and focus on capturing risk features at a second-level granularity.

[0118] The risk calculation unit is configured to perform real-time risk calculation on associated subjects that need to be monitored in a business scenario relying on a distributed computing cluster constructed based on Apache Flink.

[0119] The graph construction unit is configured to construct a knowledge graph around six core entity types of "enterprise-goods-port-freight rate-exchange rate-policy", clearly define the key relationships and attributes between entities, for example, the relationship of "goods-transport via-port" includes attributes such as expected time consumption and actual delay rate, the relationship of "port-controlled by-policies and regulations" labels the policy sensitivity score, and the relationship of "exchange rate-affects-freight rate" quantifies the correlation strength through the cointegration relationship coefficient. The constructed knowledge graph converts the scattered information in multi-modal data into a structured associated network, provides deep logical support between entities for risk calculation, and complements the correlation quantification indicators of multi-modal data, thereby enhancing the comprehensiveness and accuracy of risk analysis.

[0120] In conclusion, the import and export risk real-time early warning method and system based on multi-modal data fusion of the application realizes the breakthrough improvement of cross-border trade risk early warning through multi-modal data dynamic fusion and real-time calculation technology: the composite risk identification accuracy is improved, the early warning response time is further shortened and the false positive rate is reduced; based on the six-dimensional knowledge graph of "enterprise-goods-port-freight rate-exchange rate-policy" and the CEP rule engine, the coupling risk of exchange rate fluctuation and logistics delay can be accurately identified, and real-time disposal suggestions (such as port switching and exchange rate locking) are provided for cross-border e-commerce, foreign trade enterprises and financial institutions.

[0121] The above application specific examples have described the principles and implementation modes of the application in detail, and these examples are only used to help understand the core technical content of the application. Based on the above specific embodiments of the application, any improvement and modification of the application made by the person skilled in the art without departing from the principles of the application shall fall within the patent protection scope of the application.

Claims

1. A real-time early warning method for import and export risks based on multimodal data fusion, characterized in that: The steps include: S1. Real-time acquisition of multimodal data, including customs declaration data, logistics time series data, international freight rate data, RMB exchange rate and public opinion text data; S2. Preprocess the acquired multimodal data, including structured data standardization, unstructured text key entity extraction, and time series data trend feature extraction; S3. Design a dynamic weight allocation algorithm based on reinforcement learning. This algorithm assigns differentiated weights to different types of data based on the data's quantitative indicators of timeliness, authority, and relevance to target transactions, enabling accurate dynamic fusion of multimodal data. S4: Use the streaming computing engine to extract features from the fused real-time data, focusing on capturing risk features at a granularity of seconds. Real-time risk calculations are performed using a distributed computing cluster built on Apache Flink. A six-dimensional knowledge graph is also used to construct a "company-commodity-port-freight-rate-exchange-policy" association network, providing a more comprehensive and in-depth analysis basis for real-time risk calculations. S5. Based on the features extracted in step S4 and the association network constructed, the CEP rule engine uses predefined multi-class composite risk models to quickly identify and calculate potential risk events on the multimodal fusion data after dynamic weight assignment, accurately locating risk points. S6. Establish a three-level warning mechanism of red, yellow and green, and establish an intelligent decision-making output mechanism. According to the preset warning level trigger conditions, the specific risk patterns identified by the CEP rule engine are mapped to the corresponding warning levels, and graded warning signals are generated. Dynamic disposal suggestions are also prompted simultaneously to achieve visual presentation and operational disposal of risks.

2. The real-time early warning method for import and export risks based on multimodal data fusion according to claim 1 is characterized in that: The specific operations of obtaining multimodal data in step S1 include: (1a) Connect to the customs data interface, receive raw data in real time according to the data transmission protocol, extract structured information including HS code, declared value and certificate of origin fields through preset parsing rules, ensure the matching of data fields with business requirements, and form a standardized customs declaration data set; (1b) Connect to the ship AIS positioning system and the truck GPS trajectory monitoring platform, establish a continuous data reception channel, clean and correlate the acquired raw time series data containing geographic location and timestamps, generate dynamic indicators including transportation delay rate and port congestion index, and build a complete logistics time series database; (1c) Integrate the data interface of the Baltic Ocean Freight Index publishing channel and the China-Europe Railway Express freight rate information platform, regularly capture real-time freight rate data, and establish a storage and update mechanism based on historical data to form an international freight rate data set; (1d) Obtain real-time RMB exchange rate data and forward exchange rate curves at a set frequency through designated open API interfaces, perform format verification and timestamp alignment on the data to ensure data accuracy and timeliness, and establish an exchange rate database to support cross-border settlement risk prediction; (1e) Using web crawler technology, we collect discussion content related to compliance policies from public channels to obtain public opinion text data. We use the BERT model to perform sentiment analysis and keyword extraction on the obtained public opinion text data to form a structured public opinion analysis dataset.

3. The real-time early warning method for import and export risks based on multimodal data fusion according to claim 2 is characterized in that: The pre-processing operation performed on the multimodal data in step S2 specifically includes: (2a) For three types of structured data, namely customs declaration data, international freight rate data, and RMB exchange rate data, regular expressions are used to match and correct non-compliant field formats. At the same time, data standards are unified through a rule engine to ensure that the formats of various structured data types are consistent and the meanings of the fields are unified; (2b) For unstructured data such as public opinion texts, named entity recognition technology is used to deeply analyze the text content, extract the key entity information contained therein, and convert the unstructured text into structured entity data that can be used for analysis; (2c) For logistics time series data, a sliding window with a window size of 5 minutes and a sliding step size of 1 minute is set. The continuous logistics time series data is intercepted and calculated through this sliding window to extract the logistics delay trend characteristics in different time periods to clearly reflect the changing laws and trends of logistics delays.

4. The real-time early warning method for import and export risks based on multimodal data fusion according to claim 3 is characterized in that: The step S3 specifically performs the following operations to assign differentiated weights to different types of data, thereby achieving accurate dynamic fusion of multimodal data: S3.

1. Quantify the timeliness, authority, and relevance of each type of data to the target transaction using the preprocessed multimodal data as the processing object, generating timeliness coefficients, authority coefficients, and relevance coefficients that can be input into the dynamic weight allocation algorithm. S3.

2. Construct a dynamic weight allocation algorithm based on reinforcement learning. Use the timeliness coefficient, authority coefficient, and relevance coefficient of various types of data as input parameters of the dynamic weight allocation algorithm. Through the intelligent decision-making mechanism of reinforcement learning, establish a dynamic mapping relationship between data features and weight allocation. S3.

3. Using a dynamic weight allocation algorithm, different weights are assigned to different types of data, and further weighted fusion of various multimodal data is performed; S3.

4. Use a reinforcement learning mechanism to monitor and provide feedback on the performance of the fused data in actual business applications in real time. When it is found that the weight distribution of a certain type of data does not match the actual business value, readjust the parameters and decision logic of the dynamic weight distribution algorithm to optimize the weight distribution results.

5. The real-time early warning method for import and export risks based on multimodal data fusion according to claim 4 is characterized in that: The step S4 specifically includes: S4.

1. After preprocessing and dynamic fusion of multimodal data, feature extraction is performed on the fused real-time data using a streaming computing engine, focusing on capturing risk features at a granularity of seconds. S4.2, a distributed computing cluster built on Apache Flink, performs real-time risk calculations for associated entities that need to monitor risks in business scenarios; S4.

3. Construct a knowledge graph around the six core entity types of "enterprise-commodity-port-freight rate-exchange rate-policy" to clarify the key relationships and attributes between entities; the knowledge graph transforms the scattered information in multimodal data into a structured association network, providing deep logical support between entities for risk calculation, complementing the quantitative correlation indicators of multimodal data, and enhancing the comprehensiveness and accuracy of risk analysis.

6. A real-time early warning system for import and export risks based on multimodal data fusion, characterized by: It includes: Multimodal data acquisition module, used to obtain multimodal data in real time, including customs declaration data, logistics time series data, international freight rate data, RMB exchange rate and public opinion text data; The data preprocessing module is used to preprocess the acquired multimodal data, including structured data standardization, unstructured text key entity extraction, and time series data trend feature extraction; The data weighted fusion module is used to design a dynamic weight allocation algorithm based on reinforcement learning. Based on the quantitative indicators of multimodal data in terms of timeliness, authority, and relevance to target transactions, it assigns differentiated weights to different types of data, achieving accurate dynamic fusion of multimodal data. The feature extraction and network construction module is used to extract features from fused real-time data through a streaming computing engine, focusing on capturing risk features at a granularity of seconds. On the other hand, it is used to perform real-time risk calculations on a distributed computing cluster built with Apache Flink. It also combines a six-dimensional knowledge graph to build a "enterprise-commodity-port-freight-rate-exchange-policy" association network, providing a more comprehensive and in-depth analysis basis for real-time risk calculations. The risk identification and positioning module is used to quickly identify and calculate potential risk events based on the features extracted and the association network constructed by the feature extraction and network construction modules. The CEP rule engine relies on predefined multi-class composite risk models to dynamically assign weights to multimodal fusion data, accurately locating risk points. The early warning and suggestion generation module is used to establish a three-level early warning mechanism of red, yellow and green, and to establish an intelligent decision-making output mechanism. According to the preset early warning level trigger conditions, the specific risk patterns identified by the CEP rule engine are mapped to the corresponding early warning levels, and graded early warning signals are generated. Dynamic disposal suggestions are also prompted simultaneously to achieve visual presentation and operational disposal of risks.

7. The real-time early warning system for import and export risks based on multimodal data fusion according to claim 6 is characterized in that: The multimodal data acquisition module specifically obtains customs declaration data, logistics time series data, international freight rate data, RMB exchange rate and public opinion text data through the following operations: (1a) Connect to the customs data interface, receive raw data in real time according to the data transmission protocol, extract structured information including HS code, declared value and certificate of origin fields through preset parsing rules, ensure the matching of data fields with business requirements, and form a standardized customs declaration data set; (1b) Connect to the ship AIS positioning system and the truck GPS trajectory monitoring platform, establish a continuous data reception channel, clean and associate the acquired raw time series data containing geographic location and timestamps, generate dynamic indicators including transportation delay rate and port congestion index, and build a complete logistics time series database; (1c) Integrate the data interface of the Baltic Ocean Freight Index publishing channel and the China-Europe Railway Express freight rate information platform, regularly capture real-time freight rate data, and establish a storage and update mechanism based on historical data to form an international freight rate data set; (1d) Obtain real-time RMB exchange rate data and forward exchange rate curves at a set frequency through designated open API interfaces, perform format verification and timestamp alignment on the data to ensure data accuracy and timeliness, and establish an exchange rate database to support cross-border settlement risk prediction; (1e) Using web crawler technology, we collect discussion content related to compliance policies from public channels to obtain public opinion text data. We use the BERT model to perform sentiment analysis and keyword extraction on the obtained public opinion text data to form a structured public opinion analysis dataset.

8. The real-time early warning system for import and export risks based on multimodal data fusion according to claim 7 is characterized in that: The data preprocessing module specifically performs the following preprocessing operations on the acquired customs declaration data, logistics time series data, international freight rate data, RMB exchange rate and public opinion text data: (2a) For three types of structured data, namely customs declaration data, international freight rate data, and RMB exchange rate data, regular expressions are used to match and correct non-compliant field formats. At the same time, data standards are unified through a rule engine to ensure that the formats of various structured data types are consistent and the meanings of the fields are unified; (2b) For unstructured data such as public opinion texts, named entity recognition technology is used to deeply analyze the text content, extract the key entity information contained therein, and convert the unstructured text into structured entity data that can be used for analysis; (2c) For logistics time series data, a sliding window with a window size of 5 minutes and a sliding step size of 1 minute is set. The continuous logistics time series data is intercepted and calculated through this sliding window to extract the logistics delay trend characteristics in different time periods to clearly reflect the changing laws and trends of logistics delays.

9. The real-time early warning system for import and export risks based on multimodal data fusion according to claim 8 is characterized in that: The data weighted fusion module specifically includes: The data feature quantification unit is used to process pre-processed multimodal data and quantify the timeliness, authority, and relevance of each type of data to the target transaction, generating timeliness coefficients, authority coefficients, and relevance coefficients that can be input into the dynamic weight allocation algorithm. The dynamic weight intelligent allocation unit is used to build a dynamic weight allocation algorithm based on reinforcement learning. The timeliness coefficient, authority coefficient, and relevance coefficient of various types of data are used as the core input parameters of the dynamic weight allocation algorithm. Through the intelligent decision-making mechanism of reinforcement learning, a dynamic mapping relationship between data features and weight allocation is established; Dynamic weighted fusion calculation unit, used to assign differentiated weights to different types of data through dynamic weight allocation algorithm, and further perform weighted fusion of various types of multimodal data; The reinforcement learning weight optimization unit is used to monitor and provide feedback on the performance of fused data in actual business applications in real time through the reinforcement learning mechanism. When it is found that the weight distribution of a certain type of data does not match the actual business value, the parameters and decision logic of the dynamic weight distribution algorithm are readjusted to optimize the weight distribution results.

10. The real-time early warning system for import and export risks based on multimodal data fusion according to claim 9 is characterized in that: The feature extraction and network construction module specifically includes: The feature extraction unit is used to extract features from the fused real-time data after preprocessing and dynamic fusion of multimodal data through a streaming computing engine, focusing on capturing risk features at a granularity of seconds. The risk calculation unit is used to perform real-time risk calculations on the associated entities that need to monitor risks in business scenarios using a distributed computing cluster built on Apache Flink. The graph construction unit is used to construct a knowledge graph around the six core entity types of "enterprise-commodity-port-freight rate-exchange rate-policy" to clarify the key relationships and attributes between entities; the constructed knowledge graph transforms the scattered information in multimodal data into a structured association network, providing deep logical support between entities for risk calculation, complementing the quantitative correlation indicators of multimodal data, and enhancing the comprehensiveness and accuracy of risk analysis.

Images