Intelligent Identification and Management System for Foreign-Related Data Based on Modulus-Digital Resonance
By adopting a modular-digital resonance closed-loop architecture, the problem of two-way interaction between data and models in the foreign-related data management system is solved, realizing the full utilization of data value and the continuous adaptability of the model, and improving the system's recognition accuracy and autonomy.
Patent Information
- Application Number
- CN202611342976.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-01
- Publication Date
- 2026-11-17
AI Technical Summary
In existing intelligent identification and management systems for foreign-related data, the lack of continuous two-way interaction and collaborative evolution between data and models leads to the decay of data value and a decrease in model recognition accuracy, making it unable to adapt to the dynamically changing foreign-related environment.
It adopts a closed-loop architecture based on analog-digital resonance, including a dual-stream perception and weakness discovery module, an adaptive data forging factory, a rhythm-synchronized online model evolution module, and a collaborative and mutually reinforcing closed-loop controller. Through real-time identification, data augmentation, and model updates, it forms a resonance between data governance and model iteration.
It achieves full utilization of data value and continuous adaptability of the model, maintains high recognition accuracy, reduces reliance on expert intervention, and enables the system to operate autonomously for a long time.
Smart Images

Figure CN123115902A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an intelligent identification and management system for foreign-related data based on modular-digital resonance, belonging to the field of intelligent identification and management technology for foreign-related data. Background Technology
[0002] In the field of intelligent identification and management of foreign-related data, existing systems generally adopt a one-way processing paradigm of static models and static data: data is first collected from cross-border business systems and international communication interfaces, and then labeled once by manual or semi-automatic tools to form a fixed training set. The classification model is then trained offline and deployed online. After that, fixed features are extracted and model inference is performed on the real-time data stream, and the judgment result is output and management actions are executed. Under this paradigm, there is a lack of continuous two-way interaction and co-evolution between data and models. On the one hand, high-value feedback data generated during the production process, such as boundary cases, new data patterns, and manual review and correction records, are not effectively collected and utilized, leading to data value decay and waste. On the other hand, dynamic changes in the international data environment, such as updates to the sanctioned entity list, minor adjustments to international settlement code rules, and the emergence of new cross-border fraud texts, cause the recognition accuracy of fixed models to continuously decline, resulting in missed and false alarms, and the models gradually become rigid and unsuitable. In addition, existing systems lack a two-way calibration mechanism between data and models. Models cannot perceive their own weaknesses, and data governance tools cannot prioritize high-value data to address model shortcomings. The two operate in isolation, making it difficult to form a synergistic and mutually reinforcing closed loop. Therefore, there is an urgent need for a closed-loop system mechanism that enables real-time feedback and resonance between the data governance process and the model iteration process, in order to fundamentally solve the pain points of data not being usable and models being unsuitable. To address the aforementioned technical challenges, a smart identification and management system for foreign-related data based on modular-digital resonance is proposed. Summary of the Invention
[0003] In view of this, the present invention provides an intelligent identification and management system for foreign-related data based on modulus resonance, in order to solve or alleviate the technical problems existing in the prior art, and at least provide a beneficial option.
[0004] The technical solution of the present invention is implemented as follows: a foreign-related data intelligent identification and management system based on modulus-digital resonance, including a dual-stream sensing and weakness discovery module, an adaptive data forging workshop, a rhythm-synchronized online model evolution module, and a collaborative and mutually reinforcing closed-loop controller deployed in a computing device; the computing device includes a processor, a memory, and a data communication interface; The dual-stream perception and weakness detection module is used to identify real-time foreign-related data streams using the current foreign-related data identification model, output classification results, record the correction operations of the classification results during manual review to obtain deviation correction samples, and determine target samples based on cognitive uncertainty and the deviation correction samples through the built-in uncertainty and deviation estimator, and send the target samples to the adaptive data forging workshop. The adaptive data forging workshop is used to perform data augmentation and annotation on the target sample, generate resonance data targeting the weaknesses of the current foreign-related data identification model, and send the resonance data to the rhythm-synchronized online model evolution module. The rhythm-synchronized online evolution module is used to perform incremental learning or small-sample rapid adaptation on the current foreign-related data identification model using the resonance data to obtain an updated foreign-related data identification model, and then transmits the updated foreign-related data identification model back to the dual-stream perception and weakness detection module. The collaborative and mutually reinforcing closed-loop controller is used to monitor the performance measurement data of the updated foreign-related data identification model, and dynamically adjust the operating parameters of the dual-flow perception and weakness discovery module, the adaptive data forging workshop, and the rhythm-synchronized online model evolution module according to the performance measurement data, so that the data governance process and the model iteration process form a closed loop of resonance.
[0005] More preferably, the dual-stream sensing and weakness detection module includes a data stream processing unit, an uncertainty and bias estimator, and a bias signal recording unit; the data stream processing unit, the uncertainty and bias estimator, and the bias signal recording unit are all deployed in the processor and memory of the computing device, and interact with each other through the data communication interface; The data stream processing unit is used to receive the real-time foreign-related data stream entering the system in real time, input the real-time foreign-related data stream into the current foreign-related data identification model for identification, and obtain the classification result. The uncertainty and bias estimator is used to estimate the cognitive uncertainty of the classification result, obtain the cognitive uncertainty, and identify data whose cognitive uncertainty is higher than a preset threshold as the first target sample; The deviation signal recording unit is used to record the correction operation of the classification result during manual review, use the data carrying the correction operation as the deviation correction sample, and determine the deviation correction sample as the second target sample; The data stream processing unit is further configured to send the first target sample and the second target sample as the target samples to the adaptive data forging plant.
[0006] More preferably, the uncertainty and bias estimator is a Bayesian neural network or an ensemble learning model; the cognitive uncertainty is prediction variance, prediction entropy, or ensemble divergence; and the preset threshold is dynamically adjusted by the collaborative and mutually reinforcing closed-loop controller based on the performance measurement data.
[0007] More preferably, the adaptive data forging workshop includes a seed receiving unit, a data augmentation and mutation unit, and a quality purification and labeling unit; the seed receiving unit, the data augmentation and mutation unit, and the quality purification and labeling unit are all deployed in the processor and memory of the computing device, and interact with each other through the data communication interface; The seed receiving unit is used to receive the target sample sent by the dual-stream sensing and vulnerability discovery module, and to use the target sample as a seed sample. The data augmentation and mutation unit is used to perform at least one of the following transformations on the seed sample: synonym replacement, encoding format mutation, and entity obfuscation, which conform to the logic of foreign-related data, to obtain an augmented sample. The quality purification and labeling unit is used to automatically clean and accurately label the enhanced samples by combining the rule engine and the manual review interface to obtain the resonance data, and send the resonance data to the rhythm-synchronized online model evolution module.
[0008] Further preferably, the data augmentation and mutation unit has a built-in foreign-related data logic rule library; the foreign-related data logic rule library is stored in the memory of the computing device; the data augmentation and mutation unit selects a transformation strategy from the foreign-related data logic rule library according to the data type corresponding to the seed sample, and the first weakness type characterized by cognitive uncertainty exceeding a preset threshold and the second weakness type characterized by the bias correction sample, and performs corresponding data augmentation and mutation on the seed sample, so that the augmented sample maintains the semantic and compliance characteristics of foreign-related data; the first weakness type and the second weakness type together constitute the weakness type of the current foreign-related data identification model.
[0009] More preferably, the rhythm-synchronized online model evolution module includes a resonance data receiving unit, a continuous learning pipeline, and a model parameter update unit; the resonance data receiving unit, the continuous learning pipeline, and the model parameter update unit are all deployed in the processor and memory of the computing device, and interact with each other through the data communication interface; The resonance data receiving unit is used to receive the resonance data produced by the adaptive data forging plant; The continuous learning pipeline is used to trigger incremental learning or rapid small-sample adaptation of the current foreign-related data identification model after the resonance data receiving unit receives the resonance data. The model parameter update unit is used to update the parameters of the current foreign-related data identification model using the resonance data under the constraint of forgetting old knowledge, so as to obtain the updated foreign-related data identification model, and then send the updated foreign-related data identification model back to the dual-stream perception and weakness detection module.
[0010] Further preferably, the continuous learning pipeline includes a memory replay buffer; the memory replay buffer is deployed in the memory of the computing device and is used to store historical representative samples; when the incremental learning or small sample fast adaptation is triggered, the continuous learning pipeline samples historical representative samples from the memory replay buffer, and after mixing the historical representative samples with the resonance data, updates the parameters of the current foreign-related data identification model; or, when the incremental learning or small sample fast adaptation is triggered, the continuous learning pipeline uses elastic weight consolidation or knowledge distillation to update the parameters of the current foreign-related data identification model.
[0011] More preferably, the collaborative and mutually reinforcing closed-loop controller includes a performance measurement unit and a policy scheduling unit; both the performance measurement unit and the policy scheduling unit are deployed in the processor and memory of the computing device, and interact with each other through the data communication interface; The performance measurement unit is used to monitor in real time the changes in the rare false alarm rate and the high-frequency false alarm rate of the updated foreign-related data identification model, and to obtain the performance measurement data. The strategy scheduling unit is used to generate strategy scheduling instructions based on the performance measurement data, and send the strategy scheduling instructions to the dual-stream perception and weakness detection module, the adaptive data forging workshop, and the rhythm-synchronized online model evolution module, respectively, to dynamically adjust the sensitivity of the dual-stream perception and weakness detection module, the forging intensity of the adaptive data forging workshop, and the evolution frequency of the rhythm-synchronized online model evolution module.
[0012] In a further preferred embodiment, under the constraints of system computing resources, storage resources, and manual review interface resources, the strategy scheduling unit aims to maximize the ratio of model performance improvement to resource consumption by dynamically optimizing the sensitivity of the dual-stream perception and weakness detection module, the forging intensity of the adaptive data forging workshop, and the evolution frequency of the rhythm-synchronized online model evolution module, so that the data governance process and the model iteration process are synchronized in time.
[0013] Further preferably, it also includes a foreign-related data access interface and a recognition result management execution interface; both the foreign-related data access interface and the recognition result management execution interface are deployed in the computing device and interact with each other through the data communication interface; The foreign-related data access interface is used to connect the cross-border business system and the international communication interface to collect the real-time foreign-related data stream; The identification result management execution interface is used to send the classification result output by the dual-stream perception and vulnerability discovery module to an external management execution module, so that the management execution module can perform at least one management action among blocking, alarming, or archiving.
[0014] The embodiments of the present invention have the following advantages due to the adoption of the above technical solutions: This invention addresses the fundamental pain points of existing technologies—namely, the inability to utilize data and the unsuitability of models—by constructing a modular-digital resonance closed-loop architecture comprised of dual-stream sensing and weakness detection, adaptive data forging, rhythm-synchronized online model evolution, and a collaborative and mutually reinforcing closed-loop controller. Driven by model cognitive uncertainty and human correction bias, it automatically identifies model weaknesses and targets the forging of high-value resonance data, ensuring that each data point is used to compensate for the model's current shortcomings, fully activating data value and avoiding the data waste caused by traditional full-scale annotation. Immediately after resonance data is generated, it triggers incremental model learning or small-sample adaptation, under the constraint of forgetting old knowledge. The system rapidly absorbs new knowledge, enabling it to continuously adapt to dynamic changes such as sanctions list updates and rule fine-tuning, maintaining high accuracy and overcoming post-deployment performance drift. The collaborative and mutually reinforcing closed-loop controller monitors model performance in real time and dynamically adjusts data forging intensity and model evolution frequency, ensuring that the data quality stream and model accuracy stream are strictly synchronized in time. This avoids blind investment of computing power and manpower in non-critical aspects, significantly improving the overall ROI of the system. At the same time, the entire resonance loop operates automatically, greatly reducing reliance on frequent intervention from data scientists and domain experts, and achieving long-term, high-level autonomous operation of the foreign-related data identification and management system.
[0015] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a diagram illustrating the overall closed-loop architecture of the intelligent identification and management system for foreign-related data based on modulus-digital resonance, as described in this invention. Figure 2This is a schematic diagram illustrating the working principle of the dual-flow sensing and weakness detection module in this invention. Figure 3 This is a flowchart of the adaptive data forging workshop in this invention; Figure 4 This is a schematic diagram of the online evolution module for rhythm synchronization in this invention; Figure 5 This is a comparison curve of the performance (F1-score) drift of the system of the present invention and the static recognition system under long-term operation. Detailed Implementation
[0018] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.
[0019] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0020] like Figure 1-4 As shown, this embodiment of the invention provides an intelligent identification and management system for foreign-related data based on modulus-digital resonance. Deployed on a computing device equipped with a processor, memory, and data communication interface, the system comprises four core modules: a dual-stream perception and weakness detection module, an adaptive data forging workshop, a rhythm-synchronized online model evolution module, and a collaborative and mutually reinforcing closed-loop controller. It also includes a foreign-related data access interface and a recognition result management and execution interface. The dual-stream perception and weakness detection module, the adaptive data forging workshop, the rhythm-synchronized online model evolution module, and the collaborative and mutually reinforcing closed-loop controller are all stored in the memory, executed by the processor, and interact with each other through the data communication interface. This system is applied to the business scenario of cross-border import and export trade customs declaration. The business scenario background is that the cross-border trade customs declaration system continuously generates real-time foreign-related data streams such as customs declaration messages, trade document texts, and cross-border communication messages. It needs to identify foreign-related risk data containing export control entities, sanctioned countries, and GDPR-sensitive personal information. During business operation, newly released export control lists, new trade term variations, and new document coding formats will continuously emerge. Traditional static models will gradually experience omissions and false alarms as business iterates. The foreign-related data access interface connects to the external cross-border trade customs declaration business system and the international communication interface, continuously collecting real-time foreign-related data streams consisting of customs declaration messages, trade document texts, and cross-border communication messages. The collected real-time foreign-related data streams are sent to the dual-flow perception and vulnerability discovery module. The identification result management execution interface forwards the classification results output by the dual-flow perception and vulnerability discovery module to the external management execution module, which then performs at least one management action based on the classification results: blocking, alarm, and archiving. When a customs declaration message that hits an export control entity is identified, the customs declaration process is blocked. When a trade document suspected of involving sanctioned country risks is identified, manual review and alarm are issued. Ordinary compliant cross-border business messages are archived and retained. The dual-stream sensing and vulnerability discovery module includes a data stream processing unit, an uncertainty and bias estimator, and a bias signal recording unit, all of which are deployed in the processor and memory of the computing device. The data stream processing unit receives real-time foreign-related data streams such as customs declarations and trade documents from the foreign-related data access interface, inputs the real-time foreign-related data streams into the current foreign-related data identification model to perform inference and identification, and outputs the corresponding classification results. The classification results include export control data, GDPR-related personal data, and non-foreign-related risk data category labels, and output the corresponding confidence information. The uncertainty and bias estimator is implemented using a Bayesian neural network or an ensemble learning model to estimate the cognitive uncertainty of the customs declaration classification results output by the model. Cognitive uncertainty can be calculated using prediction variance, prediction entropy, or ensemble divergence. If a new alias for a controlled entity appears in a customs declaration, and the model output class confidence is between 0.4 and 0.6, and the prediction entropy is higher than a preset threshold, the customs declaration data is identified as the first target sample. This type of sample represents the model's knowledge blind spot. The model's wavering judgment on this type of new alias sample is a manifestation of the model's weakness, namely the first type of weakness. The deviation signal recording unit records the correction operations made by the operators in the manual review process to the classification results of customs declaration documents. The original model judged the customs declaration containing the alias of sanctioned enterprises as non-foreign-related risk data. The manual review corrected it to be export control data. The customs declaration with the correction operation record is the deviation correction sample. The deviation correction sample serves as the second target sample. This type of sample directly reflects that the model has made a judgment error and represents the second weakness type. The first weakness type and the second weakness type together constitute the weakness type of the current foreign-related data identification model. In this scenario, the first weakness type is the blind spot of newly added control entity alias recognition, and the second weakness type is the recognition error of known sanctioned entity variants. The data stream processing unit aggregates the first target sample and the second target sample, and sends them together as the target sample to the adaptive data forging factory. The preset threshold used by the uncertainty and bias estimator is not a fixed constant. The preset threshold is dynamically adjusted by the collaborative and mutually reinforcing closed-loop controller based on performance measurement data. In this customs declaration business scenario, when the system detects an increase in the rare underreporting rate, which means that many newly added controlled entity documents are not identified, the threshold is raised to filter out fewer but more problematic customs declaration document samples. When the overall recognition efficiency of the model is high, the threshold is appropriately lowered to explore more potential boundary customs declaration document samples. The adaptive data forging factory includes a seed receiving unit, a data augmentation and mutation unit, and a quality cleaning and labeling unit. Each unit is deployed in the processor and memory of the computing device and completes data interaction with each other. The seed receiving unit receives target samples of customs declaration documents issued by the dual-stream sensing and vulnerability discovery module, and uses the target samples as seed samples. The data augmentation and mutation unit reads the foreign-related data logic rule base stored in the memory, combines the data type of the seed customs declaration sample with the first weakness type of the newly added control entity alias recognition blind spot and the second weakness type of the recognition error of known sanction entity variants, and selects the corresponding transformation strategy from the foreign-related data logic rule base to perform transformation operations on the seed customs declaration sample. The transformation method includes at least one of synonym replacement, encoding format mutation, and entity obfuscation to generate an enhanced customs declaration sample. The transformation process strictly follows the foreign-related data logic of customs declaration business to ensure that the enhanced sample retains the business semantics and compliance characteristics of the original customs declaration document and will not generate invalid samples that violate customs declaration logic. Synonym replacement can replace the export control enterprise in the customs declaration with the export control entity that frequently appears in the business message. Encoding format mutation can change the date encoding format and country code writing format of the customs declaration. Entity obfuscation can generate alias variants for the name of the control enterprise. The quality purification and annotation unit, combined with a rule engine and a manual review interface, automatically cleans the generated enhanced customs declaration sample, removing invalid samples with noise, logical conflicts, and those that do not conform to the customs declaration format. Afterward, precise annotation is completed to generate resonance data. In this scenario, the resonance data is a batch of simulated customs declaration datasets containing aliases and variants of controlled entities, specifically designed to address the weaknesses of current foreign-related data recognition models in customs declaration scenarios. The resonance data is then sent to the rhythm-synchronized online model evolution module. The collaborative and mutually reinforcing closed-loop controller can issue control commands to the adaptive data forging workshop to dynamically adjust the forging intensity. The forging intensity characterizes the number of enhanced customs declaration samples generated in a single batch and the complexity of the transformation strategy. When the false negative rate increases, the forging intensity is increased to generate more batches of controlled entity variant customs declaration samples. The rhythm-synchronized online evolution module for the model includes a resonant data receiving unit, a continuous learning pipeline, and a model parameter update unit, all of which are deployed on the processor and memory of the computing device. The resonance data receiving unit receives resonance data composed of control entity variant customs declaration documents output by the adaptive data forging workshop. Once the resonance data reception is completed, the continuous learning pipeline is immediately triggered to start the incremental learning or small sample rapid adaptation process of the current foreign-related data recognition model. The continuous learning pipeline is configured with a memory replay buffer, which is stored in memory and internally holds historically representative customs declaration samples, such as classic export control customs declaration samples. When the model performs incremental learning or rapid adaptation to small samples, there are two optional execution paths: First, the continuous learning pipeline samples historically representative customs declaration samples from the memory replay buffer, mixes the historically representative customs declaration samples with the newly received resonant customs declaration dataset, and uses the mixed dataset for model parameter updates; Second, it does not sample historical samples and uses elastic weight consolidation or knowledge distillation techniques to update parameters. Both methods serve to suppress catastrophic forgetting, constraining the model to retain its ability to recognize old version sanctions list customs declarations as much as possible while learning new knowledge such as newly added aliases of controlled entities, and avoiding degradation of recognition performance in old scenarios after updating the model. Under the constraint of forgetting old knowledge, the model parameter update unit uses resonant data or a mixture of resonant data and historical representative customs declaration samples to update the parameters of the current foreign-related data identification model, resulting in an updated foreign-related data identification model. The updated foreign-related data identification model is then fed back to the dual-flow perception and weakness discovery module to replace the original model for real-time data flow inference of customs declarations and trade documents. At the same time, the performance data of the updated model is reported to the collaborative and mutually reinforcing closed-loop controller. The collaborative and mutually reinforcing closed-loop controller can dynamically adjust the evolution frequency of the model online evolution module with synchronized rhythm, that is, control the frequency of incremental learning triggered by the model. During peak customs declaration periods, the evolution frequency is reduced to avoid consuming too many computing resources, while during off-peak periods, the evolution frequency can be appropriately increased. The collaborative and mutually reinforcing closed-loop controller, including a performance measurement unit and a policy scheduling unit, is deployed in the processor and memory of the computing device; The performance measurement unit collects and monitors the updated performance indicators of the foreign-related data identification model in real time. In this customs declaration business scenario, it focuses on monitoring the changes in the rare underreporting rate corresponding to the underreporting of newly added control alias customs declaration documents and the high-frequency false reporting rate corresponding to the misjudgment of regular customs declaration documents as control risks, and generates performance measurement data. The strategy scheduling unit receives performance measurement data, generates strategy scheduling instructions, and issues them to the dual-stream perception and weakness detection module, the adaptive data forging workshop, and the rhythm-synchronized online model evolution module, respectively; and dynamically adjusts the sensitivity of the dual-stream perception and weakness detection module (i.e., the aforementioned uncertainty preset threshold), the forging intensity of the adaptive data forging workshop, and the evolution frequency of the rhythm-synchronized online model evolution module. The parameter optimization process of the strategy scheduling unit is subject to the practical constraints of system computing resources, storage resources, and manual review interface resources. The optimization goal is to maximize the ratio of model performance improvement to resource consumption; that is, to use the least computing power, storage, and human resources for manual review of customs declaration documents to achieve the greatest improvement in model recognition performance in customs declaration scenarios. Through this optimization and control, the data governance process and model iteration process of customs declaration business are synchronized in the time dimension, forming a complete analog-digital resonance closed loop.
[0021] like Figure 5 The figure shows a comparison curve of the F1-score performance drift of the system of the present invention and the traditional static recognition system over a long period of time. The traditional static recognition system uses historical customs declaration datasets for offline training and does not update the model after completion. As the export control entity list is updated iteratively, new entity aliases and new document formats are constantly emerging, and the F1-score continues to decline as the operating cycle progresses. However, the modular resonance closed-loop system of the present invention relies on the closed-loop continuous evolution mechanism to continuously generate corresponding resonance data iterative models for customs declaration scenarios. The F1-score always remains at a high level with small fluctuations and does not experience significant performance drift. This embodiment addresses the pain points of existing static paradigms in cross-border customs declaration scenarios, such as data unusability and model inapplicability. It uses the model's cognitive uncertainty arising from the customs document recognition process and the manual correction bias of customs declaration personnel as driving signals to automatically locate the model's weaknesses in recognizing controlled entity aliases and other content. It then targets and generates resonant data for the customs declaration scenario, fully leveraging the value of customs declaration feedback data from the production environment. The continuous learning pipeline assurance model learns new controlled entity knowledge while avoiding forgetting the ability to recognize old sanctions lists. A collaborative and mutually reinforcing closed-loop controller combines the computing power of the customs declaration system and the workload of manual positions to dynamically optimize parameters across the entire chain, achieving synchronized data governance and model iteration in the customs declaration scenario. The entire closed loop can operate autonomously, reducing reliance on frequent parameter tuning by algorithm experts and enabling the long-term autonomous operation of the customs declaration business foreign-related data recognition and management system.
[0022] The working process of this invention is as follows: S1: The foreign-related data access interface collects real-time foreign-related data streams generated by cross-border business and sends them to the dual-stream perception and vulnerability discovery module. S2: The data stream processing unit inside the dual-stream perception and weakness discovery module calls the current foreign-related data recognition model to complete the data stream inference recognition and output the classification result; the uncertainty and bias estimator outputs cognitive uncertainty to screen the first target sample; the bias signal recording unit collects the manual review and correction records to obtain the second target sample, and summarizes the target sample to output to the adaptive data forging workshop; S3: The Adaptive Data Forging Factory uses the target sample as the seed sample, calls the foreign-related data logic rule library to complete the data augmentation and mutation to obtain the augmented sample, and outputs the resonance data after cleaning and precise labeling. S4: The resonant data is fed into the rhythm-synchronized online model evolution module, triggering a continuous learning pipeline. Under the anti-forgetting constraint, the model parameters are updated to obtain the updated foreign-related data recognition model. The updated model is then sent back to the dual-stream perception and weakness discovery module for online inference. S5: The collaborative and mutually reinforcing closed-loop controller continuously monitors the updated model performance measurement data, dynamically optimizes based on resource constraints, and issues control commands to adjust the operating parameters of each module. S6: The classification results output by the dual-stream perception and vulnerability discovery module are delivered to the external management execution module via the recognition result management execution interface to perform blocking, alarm, and archiving business actions; S7: Repeat S2-S6 in a loop to form a positive feedback loop that improves data quality, enhances model capabilities, discovers new target samples, re-forges resonant data, and further evolves the model.
[0023] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A foreign-related data intelligent identification and management system based on modulus-digital resonance, characterized in that: It includes a dual-stream sensing and vulnerability discovery module, an adaptive data forging workshop, a rhythm-synchronized online model evolution module, and a collaborative and mutually reinforcing closed-loop controller deployed in a computing device; the computing device includes a processor, memory, and a data communication interface; The dual-stream perception and weakness detection module is used to identify real-time foreign-related data streams using the current foreign-related data identification model, output classification results, record the correction operations of the classification results during manual review to obtain deviation correction samples, and determine target samples based on cognitive uncertainty and the deviation correction samples through the built-in uncertainty and deviation estimator, and send the target samples to the adaptive data forging workshop. The adaptive data forging workshop is used to perform data augmentation and annotation on the target sample, generate resonance data targeting the weaknesses of the current foreign-related data identification model, and send the resonance data to the rhythm-synchronized online model evolution module. The rhythm-synchronized online evolution module is used to perform incremental learning or small-sample rapid adaptation on the current foreign-related data identification model using the resonance data to obtain an updated foreign-related data identification model, and then transmits the updated foreign-related data identification model back to the dual-stream perception and weakness detection module. The collaborative and mutually reinforcing closed-loop controller is used to monitor the performance measurement data of the updated foreign-related data identification model, and dynamically adjust the operating parameters of the dual-flow perception and weakness discovery module, the adaptive data forging workshop, and the rhythm-synchronized online model evolution module according to the performance measurement data, so that the data governance process and the model iteration process form a closed loop of resonance.
2. The intelligent identification and management system for foreign-related data based on modulus-digital resonance as described in claim 1, characterized in that: The dual-stream sensing and vulnerability discovery module includes a data stream processing unit, an uncertainty and bias estimator, and a bias signal recording unit; the data stream processing unit, the uncertainty and bias estimator, and the bias signal recording unit are all deployed in the processor and memory of the computing device, and interact with each other through the data communication interface; The data stream processing unit is used to receive the real-time foreign-related data stream entering the system in real time, input the real-time foreign-related data stream into the current foreign-related data identification model for identification, and obtain the classification result. The uncertainty and bias estimator is used to estimate the cognitive uncertainty of the classification result, obtain the cognitive uncertainty, and identify data whose cognitive uncertainty is higher than a preset threshold as the first target sample; The deviation signal recording unit is used to record the correction operation of the classification result during manual review, use the data carrying the correction operation as the deviation correction sample, and determine the deviation correction sample as the second target sample; The data stream processing unit is further configured to send the first target sample and the second target sample as the target samples to the adaptive data forging plant.
3. The intelligent identification and management system for foreign-related data based on modulus-digital resonance as described in claim 2, characterized in that: The uncertainty and bias estimator is a Bayesian neural network or an ensemble learning model; the cognitive uncertainty is prediction variance, prediction entropy, or ensemble divergence; the preset threshold is dynamically adjusted by the collaborative and mutually reinforcing closed-loop controller based on the performance measurement data.
4. The intelligent identification and management system for foreign-related data based on modulus-digital resonance as described in claim 2, characterized in that: The adaptive data forging workshop includes a seed receiving unit, a data augmentation and mutation unit, and a quality purification and labeling unit; the seed receiving unit, the data augmentation and mutation unit, and the quality purification and labeling unit are all deployed in the processor and memory of the computing device, and interact with each other through the data communication interface; The seed receiving unit is used to receive the target sample sent by the dual-stream sensing and vulnerability discovery module, and to use the target sample as a seed sample. The data augmentation and mutation unit is used to perform at least one of the following transformations on the seed sample: synonym replacement, encoding format mutation, and entity obfuscation, which conform to the logic of foreign-related data, to obtain an augmented sample. The quality purification and labeling unit is used to automatically clean and accurately label the enhanced samples by combining the rule engine and the manual review interface to obtain the resonance data, and send the resonance data to the rhythm-synchronized online model evolution module.
5. The intelligent identification and management system for foreign-related data based on modulus-digital resonance according to claim 4, characterized in that: The data augmentation and mutation unit has a built-in foreign-related data logic rule library; the foreign-related data logic rule library is stored in the memory of the computing device; the data augmentation and mutation unit selects a transformation strategy from the foreign-related data logic rule library according to the data type corresponding to the seed sample, as well as the first weakness type characterized by cognitive uncertainty exceeding a preset threshold and the second weakness type characterized by the deviation correction sample, and performs corresponding data augmentation and mutation on the seed sample so that the augmented sample maintains the semantic and compliance characteristics of foreign-related data; The first vulnerability type and the second vulnerability type together constitute the vulnerability types of the current foreign-related data identification model.
6. The intelligent identification and management system for foreign-related data based on modulus-digital resonance as described in claim 1, characterized in that: The rhythm-synchronized online evolution module includes a resonance data receiving unit, a continuous learning pipeline, and a model parameter update unit; the resonance data receiving unit, the continuous learning pipeline, and the model parameter update unit are all deployed in the processor and memory of the computing device, and interact with each other through the data communication interface; The resonance data receiving unit is used to receive the resonance data produced by the adaptive data forging plant; The continuous learning pipeline is used to trigger incremental learning or rapid small-sample adaptation of the current foreign-related data identification model after the resonance data receiving unit receives the resonance data. The model parameter update unit is used to update the parameters of the current foreign-related data identification model using the resonance data under the constraint of forgetting old knowledge, so as to obtain the updated foreign-related data identification model, and then send the updated foreign-related data identification model back to the dual-stream perception and weakness detection module.
7. The intelligent identification and management system for foreign-related data based on modulus-digital resonance as described in claim 6, characterized in that: The continuous learning pipeline includes a memory replay buffer; the memory replay buffer is deployed in the memory of the computing device and is used to store historical representative samples; when the incremental learning or small sample fast adaptation is triggered, the continuous learning pipeline samples historical representative samples from the memory replay buffer, mixes the historical representative samples with the resonant data, and updates the parameters of the current foreign-related data identification model; or, when the incremental learning or small sample fast adaptation is triggered, the continuous learning pipeline updates the parameters of the current foreign-related data identification model using elastic weight consolidation or knowledge distillation.
8. The intelligent identification and management system for foreign-related data based on modulus-digital resonance as described in claim 1, characterized in that: The collaborative and mutually reinforcing closed-loop controller includes a performance measurement unit and a policy scheduling unit; both the performance measurement unit and the policy scheduling unit are deployed in the processor and memory of the computing device and interact with each other through the data communication interface; The performance measurement unit is used to monitor in real time the changes in the rare false alarm rate and the high-frequency false alarm rate of the updated foreign-related data identification model, and to obtain the performance measurement data. The strategy scheduling unit is used to generate strategy scheduling instructions based on the performance measurement data, and send the strategy scheduling instructions to the dual-stream perception and weakness detection module, the adaptive data forging workshop, and the rhythm-synchronized online model evolution module, respectively, to dynamically adjust the sensitivity of the dual-stream perception and weakness detection module, the forging intensity of the adaptive data forging workshop, and the evolution frequency of the rhythm-synchronized online model evolution module.
9. The intelligent identification and management system for foreign-related data based on modulus-digital resonance as described in claim 8, characterized in that: Under the constraints of system computing resources, storage resources, and manual review interface resources, the strategy scheduling unit aims to maximize the ratio of model performance improvement to resource consumption. It dynamically optimizes the sensitivity of the dual-stream perception and weakness detection module, the forging intensity of the adaptive data forging workshop, and the evolution frequency of the rhythm-synchronized online model evolution module, so that the data governance process and the model iteration process are synchronized in time.
10. The intelligent identification and management system for foreign-related data based on modulus-digital resonance according to claim 1, characterized in that: It also includes a foreign-related data access interface and a recognition result management execution interface; both the foreign-related data access interface and the recognition result management execution interface are deployed in the computing device and interact with each other through the data communication interface. The foreign-related data access interface is used to connect the cross-border business system and the international communication interface to collect the real-time foreign-related data stream; The identification result management execution interface is used to send the classification result output by the dual-stream perception and vulnerability discovery module to an external management execution module, so that the management execution module can perform at least one management action among blocking, alarming, or archiving.
