An international communication data flow dynamic management method and system based on an agent intelligent body

Abstract

The application provides an international communication data flow dynamic management method and system based on an Agent intelligent agent, relates to the technical field of data governance and cross-border compliance, and belongs to the technical field of data governance and cross-border compliance. The original data pool with semantic index is constructed through a distributed crawler, the governance instruction is generated based on the multi-modal perception and joint decision model, the operation of data filtering and desensitization is performed to form the governance result, the channel is divided according to the compliance ontology model, and the strategy is continuously and adaptively updated through the multi-target Bayesian optimization, so that the dynamic governance and optimization of cross-platform and multi-lingual data flow are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data governance and cross-border compliance, in particular to an international communication data flow dynamic management method and system based on Agent intelligent agents. BACKGROUND

[0002] Under the pattern of highly interconnected global digital media, international communication data flow has grown exponentially, with media platforms, social networks and content distribution networks running simultaneously in various countries. Massive cross-language, cross-platform and cross-time-zone data flows are simultaneously flowing into centralized monitoring and filtering systems. The industry usually relies on pre-set keyword matching, black and white lists and static data quality checks to screen, such as the invention patent with publication number CN105808604A, which proposes to improve data compliance through dirty data identification and correction, representing the typical approach of centralized rules and manual governance.

[0003] With the deepening of artificial intelligence technology, automatic semantic understanding, multi-agent collaborative decision-making and cross-domain execution framework have become a hot research direction. In recent years, some solutions have tried to abstract regulatory elements into a knowledge base or "compliance factor library" and use deep learning and encryption technology to achieve cross-border data protection and blocking. For example, the invention patent with publication number CN114760149B uses dynamic compliance factor matching and privacy identification to achieve automatic compliance control of cross-border data, providing technical support for rapid iteration of multiple regulations.

[0004] However, most existing systems are still an extension of monolithic models or rule engines: their compliance knowledge base updates and governance strategy execution chain are loose, and they cannot adapt to reorganization when global regulations change rapidly or major events occur. They also lack real-time evaluation of governance effectiveness and strategy closed-loop optimization, resulting in coexistence of strategy solidification, resource waste and compliance risks. SUMMARY

[0005] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide an international communication data flow dynamic management method and system based on Agent intelligent agents, which realizes the dynamic collection and adaptive governance of cross-platform and multi-language international communication data flow, significantly improving the accuracy, real-time and sustainable optimization ability of governance.

[0006] To achieve the above purpose, the present application provides the following scheme:

[0007] An international communication data flow dynamic management method based on Agent intelligent agents, comprising:

[0008] The distributed crawler nodes dynamically allocate crawling tasks according to an extensible hash algorithm to capture cross-platform and multi-lingual data streams, and attach source identification, timestamps, topic labels and sensitivity weights to each data stream to form a raw data pool with semantic indexing.

[0009] Autonomous agents are instantiated for each topic label in the raw data pool; the autonomous agents include a perception module, a decision module and an execution module; the perception module encodes the multi-modal content of the raw data pool using a hybrid Transformer-CNN network; the decision module outputs governance instructions through a multi-armed bandit-contextual bandit-policy gradient joint network combined with an international compliance rule base; each autonomous agent exchanges state summaries and issues synchronization instructions through a message bus, and maintains global policy consistency through a consensus algorithm;

[0010] The execution module performs processing operations including filtering, desensitization, rate limiting, reordering, multi-lingual synchronization on the raw data streams according to the governance instructions to obtain governance results;

[0011] The governance results are written to a governance result queue, and the data streams in the governance result queue are divided into fast channels, review channels and cold storage channels according to the priority of the compliance ontology model;

[0012] After each governance cycle ends, governance results are extracted from the governance result queue, and a comprehensive score is calculated based on timeliness, compliance and propagation effect;

[0013] Based on the comprehensive score, a multi-objective Bayesian optimization driven model-independent meta-learning algorithm is used to update the parameters of the decision module in small steps, and the updated parameters are written back to each autonomous agent to achieve adaptive migration and optimization of the governance strategy.

[0014] Preferably, the distributed crawler nodes dynamically allocate crawling tasks according to an extensible hash algorithm to capture cross-platform and multi-lingual data streams, and attach source identification, timestamps, topic labels and sensitivity weights to each data stream to form a raw data pool with semantic indexing, including:

[0015] A predetermined data source is obtained; the data source includes news websites, social platforms, video platforms, content distribution networks, open data interfaces and other information publishing channels, each data source corresponding to an access address, data format and access protocol;

[0016] Based on the access domain name, IP address and interface call information of the data source, a unique source identification is generated, and the platform type, language type and time zone information corresponding to the data source are recorded as metadata basis;

[0017] According to the computing power, network bandwidth and historical load of each crawler node, the task domain is dynamically divided by using a scalable hash algorithm, and the data source is mapped and distributed to different distributed crawler nodes according to platform type, language type and time zone information;

[0018] Each distributed crawler node real-time crawls a target data stream based on the task domain rule, and adds the source identifier and collection timestamp information to each collected target data stream;

[0019] For the collected target data stream, the theme label is extracted through natural language processing and theme modeling, and the sensitivity weight is calculated in combination with the content sensitivity model;

[0020] The source identifier, collection timestamp information, theme label and sensitivity weight are encapsulated as structured metadata with the target data stream, and are uniformly written into the original data pool.

[0021] Preferably, the perception module encodes the multi-modal content of the original data pool by using a hybrid Transformer-CNN network, including:

[0022] For the data stream associated with each theme label in the original data pool, text data, audio data, video data and image data in the data stream are extracted to form a multi-modal data set;

[0023] For the text data, a pre-trained Transformer encoding model is used for context modeling to output a text feature vector containing semantic relationships;

[0024] For the audio data and the video data, corresponding audio feature sequences and video frame image sequences are first extracted, and then convolutional neural networks are respectively inputted to extract time sequence and spatial local features, and output audio feature vectors and video feature vectors;

[0025] For the image data, a convolutional neural network is directly inputted to extract spatial level features, and output an image feature vector;

[0026] The text feature vector, audio feature vector, video feature vector and image feature vector are spliced to construct a unified multi-modal semantic feature matrix;

[0027] The multi-modal semantic feature matrix is inputted into a fusion layer, and a fusion Transformer encoding layer is used to further model the cross-modal association relationship to output a multi-modal unified semantic vector.

[0028] Preferably, the decision module outputs a governance instruction by a multi-armed bandit-contextual bandit-strategy gradient joint network in combination with an international compliance rule library, including:

[0029] receive the multi-modal unified semantic vector s output by the perception module, and extract a compliance constraint embedding vector c corresponding to the current propagation scenario from an international compliance rule library;

[0030] concatenate the multi-modal unified semantic vector s and the compliance constraint embedding vector c to form a context vector as the state input of the multi-armed contextual bandit;

[0031] For each governance action a in the action set A, calculate the exploration-exploitation score using the multi-armed contextual bandit and select a number of candidate actions;

[0032] Send the candidate actions together with the context vector into the policy gradient network, and calculate the compliance-weighted policy gradient according to the following formula and update the policy parameters:

[0033]

[0034] where θt is the policy network parameter vector; η is the learning rate; rt is the real-time propagation return; λ is the compliance penalty coefficient; l t is the compliance violation degree; a t is the selected governance action; s t is the perception state vector; c t is the compliance embedding vector; π θ (·) is a parameterized policy probability distribution; is the gradient operator

[0035] Re-evaluate the candidate actions according to the updated policy distribution, select the optimal action as the final governance instruction output to the execution module, and write the state-action-return-penalty quadruple of this round into the experience replay buffer.

[0036] Preferably, each autonomous agent exchanges state summaries and issues synchronization instructions through a message bus, and maintains global policy consistency through a consistency algorithm, including:

[0037] Each autonomous agent generates a state summary after completing local governance decision-making; the state summary at least contains the current local policy vector, the average return value and the timestamp, and is written into the message bus through the publish and subscribe method;

[0038] Each autonomous agent subscribes to external state summaries related to its own topic label from the message bus, and constructs a neighborhood state set;

[0039] For the neighborhood state set, a weighted consistency algorithm is used to calculate the global consistent policy, and the specific update rule is:

[0040]

[0041] where, is the local policy vector of agent i in the kth round; a is the neighborhood fusion coefficient; N i is the neighborhood set visible to agent i on the message bus; w ij is the trust weight from agent j to agent i, satisfying

[0042] write the resulting consistent policy vector into the local policy cache and generate a synchronization instruction broadcast to the execution module, causing each autonomous agent to execute the post-consistent governance policy in the next decision cycle.

[0043] Preferably, the execution module is used to implement processing operations including filtering, desensitization, rate limiting, reordering, and multi-language synchronization on the original data stream according to the governance instruction, to obtain governance results, including:

[0044] Parse the governance instruction to extract operation type, execution order, target field, and language tag, and build an operation list;

[0045] Based on the keywords, regular expressions, and black / white list rules in the operation list, perform content matching on the text, image, and audio content in the original data stream, delete or shield content related to illegal topics, to obtain a filtered data stream D filter ;

[0046] Perform desensitization processing on the fields in the filtered data stream D filter that involve identity information, geographic location information, and device identifier information, implement hierarchical marking and replacement by setting a sensitivity level, to obtain a desensitized result data stream D desens ;

[0047] Based on the bandwidth threshold and time window specified in the governance instruction, perform traffic shaping on the desensitized result data stream D desens , dynamically adjust the data packet sending rate, and generate a rate-limited data stream D throttle ;

[0048] According to the priority information and time stamp in the governance instruction, reorder the rate-limited data stream D throttle to obtain a sorted data stream D sorted , where high-priority data packets are sent first and low-priority data packets are queued for later transmission;

[0049] When the operation list contains a multi-language synchronization flag, based on the text content in the sorted data stream D sorted , call a neural machine translation model and a terminology library to generate a target language version aligned with the original language content, and reuse the original media format packaging to obtain a multi-language synchronization data stream D multi ;

[0050] The filtered data stream D filter , desensitization result data stream D desens , speed-limited data stream D throttle , sorted data stream D sorted , and the latest available data in the multi-language synchronous data stream D multi The final data stream is combined with operation records, processing labels and compliance status, and is encapsulated as a governance result.

[0051] Preferably, the governance result is written into a governance result queue, and the data streams in the governance result queue are divided into a fast lane, a review lane and a cold storage lane according to the compliance ontology model, including:

[0052] When the governance result is generated, the compliance risk score, time urgency and content heat prediction value are attached to it, and written into the governance result queue;

[0053] For each governance result in the governance result queue, the compliance ontology model is called to check the number of triggered clauses, and the priority score is calculated accordingly

[0054] S=tanh(R+T-E-C)

[0055] Where R is the compliance risk score; T is the time urgency; E is the content heat prediction value; C is the number of compliance clauses triggered; S is the priority score normalized by the hyperbolic tangent function;

[0056] According to the priority score S and the preset threshold set {s1, s2}, the channel is divided:

[0057] When s1≤S, it is allocated to the fast lane; when s2≤S<s1, it is allocated to the review lane; when S<s2, it is allocated to the cold storage lane;

[0058] The channel label is updated with the governance result in the queue.

[0059] Preferably, after each governance cycle, governance results are extracted from the governance result queue, and a comprehensive score is calculated according to the timeliness, compliance and propagation effect, including:

[0060] Read the generation timestamp of the governance result, and the difference Δt from the current time, and calculate the timeliness decay factor H, which is calculated as follows:

[0061] H=e -Δt / τ

[0062] Where τ is the time decay constant, Δt is the time difference between the current time and the generation time of the governance result, and H represents the timeliness decay factor;

[0063] Based on the number C of violation clauses triggered under the compliance ontology model based on the governance result, a compliance penalty factor P is calculated in the following manner:

[0064]

[0065] The click volume V, forwarding volume F and average stay duration D of the data flow corresponding to the governance result in the target audience are obtained, and a propagation effect index E is calculated in the following manner:

[0066]

[0067] Based on the above-mentioned indexes, a comprehensive score S is calculated in the following manner:

[0068] S = σ(ln(1+E)H-P)

[0069] Wherein, S is the comprehensive score, E is the propagation effect index, H is the timeliness decay factor, P is the compliance penalty factor, and σ is the logistic function.

[0070] Preferably, based on the comprehensive score, a multi-objective Bayesian optimization driven model-independent meta-learning algorithm is used to update the parameters of the decision module in small steps, and the updated parameters are written back to each autonomous intelligent agent to realize adaptive migration and optimization of the governance strategy, including:

[0071] After each governance cycle ends, the comprehensive score S of each autonomous intelligent agent is collected i and its current decision parameters A historical sample set D is constructed Wherein, S i is the comprehensive score of the i-th autonomous intelligent agent, is the decision parameter vector of the i-th autonomous intelligent agent at time t.

[0072] For the historical sample set D, a Gaussian process proxy model is trained for the minimization target min(1-S) and based on the Gaussian process proxy model A multi-objective expected hypervolume improvement function EHVI(θ) is defined.

[0073] The optimal increment is obtained by multi-objective Bayesian optimization, and the formula is:

[0074] Δθ * = argmax Δθ EHVI(θ (t) +Δθ)

[0075] Wherein, Δθ * is the optimal parameter increment obtained by maximizing EHVI(θ).

[0076] The model-independent meta-learning is used to update the rule execution small step migration, and the formula is:

[0077] θ (t+1) = θ (t) + ηΔθ *

[0078] Wherein, η is the step size coefficient adaptively calculated by the agent model average prediction variance; θ (t+1) is the decision parameter vector after small step update;

[0079] The updated parameter θ (t+1) is written back to the decision module of the corresponding autonomous intelligent agent to complete the adaptive optimization of the strategy.

[0080] An international communication data flow dynamic management system based on Agent intelligent agent, comprising:

[0081] A data acquisition and labeling unit is used to dynamically allocate crawling tasks by distributed crawler nodes according to an extensible hash algorithm to crawl cross-platform, multi-lingual data streams, and to attach source identification, timestamps, topic labels and sensitivity weights to each data stream to form a raw data pool with semantic index;

[0082] An intelligent agent construction and collaboration unit is used to instantiate autonomous intelligent agents for each of the topic labels in the raw data pool; the autonomous intelligent agents include a perception module, a decision module and an execution module; the perception module uses a hybrid Transformer-CNN network to encode the multi-modal content of the raw data pool; the decision module outputs governance instructions through a multi-armed bandit-contextual bandit-policy gradient joint network combined with an international compliance rule base; each autonomous intelligent agent exchanges state summaries and issues synchronization instructions through a message bus, and maintains global policy consistency through a consistency algorithm;

[0083] A data governance execution unit is used to implement processing operations including filtering, desensitization, rate limiting, reordering, multi-lingual synchronization on the raw data stream according to the governance instructions using the execution module, to obtain governance results;

[0084] A priority channel division unit is used to write the governance results into a governance result queue, and divide the data streams in the governance result queue into a fast channel, a review channel and a cold storage channel according to the priority of the compliance ontology model;

[0085] A comprehensive evaluation unit is used to extract governance results from the governance result queue after each governance cycle, and calculate a comprehensive score based on timeliness, compliance and propagation effect;

[0086] The policy optimization updating unit is configured to update parameters of the decision module in small steps by using a multi-objective Bayesian optimization driven model-agnostic meta-learning algorithm based on the comprehensive score, and write the updated parameters back to each autonomous intelligent agent to realize adaptive migration and optimization of the governance strategy.

[0087] According to the specific embodiments of the present application, the following technical effects are disclosed:

[0088] (1) The present application can dynamically and efficiently cover multilingual, cross-platform international communication data stream collection by introducing distributed crawlers and scalable hash algorithms, overcoming the defects of incomplete data acquisition and insufficient real-time performance under existing centralized and static rules, ensuring the comprehensiveness and timeliness of the original data pool.

[0089] (2) The present application constructs an autonomous agent system with perception, decision-making and execution capabilities, dynamically generates governance instructions based on multi-modal semantic encoding and reinforcement learning, replaces the existing governance mode relying on a single rule base or static model, has adaptive governance capabilities for complex communication situations, and significantly improves the accuracy and flexibility of cross-context and cross-regional data stream management.

[0090] (3) The present application constructs a closed-loop optimization link by dividing the governance result priority channel and the governance cycle comprehensive evaluation mechanism, combining multi-objective Bayesian optimization and model-agnostic meta-learning, overcoming the problem of disconnection between governance effect and strategy update in existing solutions, realizing dynamic migration and continuous optimization of governance strategy, and improving the long-term effectiveness and stability of governance.

[0091] (4) Each autonomous intelligent agent exchanges state summaries through a message bus and maintains global consistency, avoiding the risk of isolated monolithic models and fragmented strategies, ensuring overall collaborative effect under multi-theme and multi-task concurrent governance, and improving the robustness and adaptability of the system under dynamic changes in global compliance environment. BRIEF DESCRIPTION OF DRAWINGS

[0092] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0093] Figure 1 The method flowchart provided for the embodiments of the present application;

[0094] Figure 2 The system structure schematic diagram provided for the embodiments of the present application. DETAILED DESCRIPTION

[0095] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0096] The purpose of the present application is to provide an Agent intelligent agent-based international communication data flow dynamic management method and system, which realizes dynamic collection and adaptive management of cross-platform, multi-language international communication data flow, and significantly improves the accuracy, real-time and sustainable optimization ability of management.

[0097] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0098] Figure 1 The method flowchart provided by the embodiments of the present application is shown in Figure 1 The present application provides an Agent intelligent agent-based international communication data flow dynamic management method, which comprises:

[0099] Step 100: The distributed crawler node dynamically allocates crawling tasks according to the extensible hash algorithm to crawl cross-platform, multi-language data flow, and adds source identification, timestamp, topic label and sensitivity weight to each data flow to form a raw data pool with semantic index;

[0100] Step 200: Instantiating autonomous intelligent agents for each topic label in the raw data pool; the autonomous intelligent agent includes a perception module, a decision module and an execution module; the perception module encodes the multi-modal content of the raw data pool using a hybrid Transformer-CNN network; the decision module outputs management instructions through a multi-armed contextual bandit-strategy gradient joint network combined with an international compliance rule base; each autonomous intelligent agent exchanges state summaries and publishes synchronization instructions through a message bus, and keeps the global strategy consistent through a consistency algorithm;

[0101] Step 300: Using the execution module to implement processing operations including filtering, desensitization, speed limiting, reordering, multi-language synchronization on the raw data flow according to the management instructions, to obtain the management result;

[0102] Step 400: Write the management result into the management result queue, and divide the data flow in the management result queue into a fast channel, a review channel and a cold storage channel according to the compliance ontology model by priority;

[0103] Step 500: After the end of each governance cycle, the governance results are extracted from the governance result queue, and the comprehensive score is calculated according to the timeliness, compliance and propagation effect;

[0104] Step 600: Based on the comprehensive score, a multi-objective Bayesian optimization driven model-independent meta-learning algorithm is used to update the parameters of the decision module in small steps, and the updated parameters are written back to each autonomous intelligent agent to realize adaptive migration and optimization of the governance strategy.

[0105] In the embodiment, preferably, the distributed crawler node dynamically allocates crawling tasks according to the scalable hash algorithm to capture cross-platform and multi-lingual data streams, and appends source identification, timestamp, topic label and sensitivity weight to each data stream to form a process of forming a raw data pool with semantic index. Specifically, the following steps are included: first, the embodiment presets a data source set, which includes news websites, social platforms, video platforms, content distribution networks, open data interfaces and other information publishing channels, and configures a unique access address, data format and access protocol for each data source. Based on the access domain name, IP address or interface call information of the data source, the embodiment generates a source identification for uniquely identifying the data source, and records the specific platform type (such as news, social, video, etc.), language type (such as Chinese, English, Japanese, etc.) and time zone information of the corresponding data source as metadata basis for subsequent data management and processing.

[0106] Further, according to the real-time computing capacity, network bandwidth and historical load of each distributed crawler node, the task domain is dynamically divided by using the scalable hash algorithm, and the data source is mapped and allocated to different distributed crawler nodes according to the platform type, language type and time zone information and other dimensions. Each crawler node captures the corresponding target data stream in real time according to the allocation result, and automatically appends the aforementioned source identification and collection timestamp information to each collected data stream. Subsequently, the embodiment performs language analysis on the collected target data stream by using a natural language processing method, identifies its propagation theme by combining a theme modeling algorithm, extracts the corresponding theme label, and calls a content sensitivity model to calculate a sensitivity weight. The source identification, collection timestamp, theme label and sensitivity weight are packaged together with the raw data stream as structured metadata, and are uniformly stored in the raw data pool as basic data for subsequent autonomous agent scheduling and governance.

[0107] Specifically, in the present embodiment, preferably, the perception module encodes the multi-modal content of the original data pool using a hybrid Transformer-CNN network, specifically including the following steps: first, for the data stream in the original data pool associated with each of the topic labels, the present embodiment extracts text data, audio data, video data and image data in the data stream respectively to form a multi-modal data set. For the text data, a Transformer encoder based on a pre-trained BERT (Bidirectional Encoder Representations from Transformers) model is used for context modeling, the BERT model includes 12 encoder layers, each layer uses 12 multi-head self-attention mechanisms, the hidden layer dimension is set to 768, the maximum input sequence length is set to 512, and the output includes a text feature vector containing semantic relationships. For the audio data and the video data, the audio is subjected to short-time Fourier transform to extract a mel-spectrogram feature sequence, and the video is extracted to extract a key frame sequence of 5 frames per second, which are uniformly input into a five-layer convolutional neural network (CNN), which includes convolution kernels of 64, 128, 256, 512, and 512 channels in each layer, the convolution kernel size is 3x3, the activation function uses ReLU, the maximum pooling kernel size is 2x2, and the corresponding audio feature vector and video feature vector are output respectively.

[0108] Further, for the image data, a convolutional neural network with the same structure as the above-mentioned CNN is directly inputted to extract spatial level features and output an image feature vector. Subsequently, the present embodiment concatenates the text feature vector, the audio feature vector, the video feature vector and the image feature vector to construct a unified multi-modal semantic feature matrix, and inputs it into a fusion layer. The fusion layer is based on a Transformer encoder architecture, and uses a 6-layer stacked encoder structure, each layer includes 8 multi-head self-attention units, the hidden dimension is 512, and the residual connection and layer normalization use standard configurations. The fusion Transformer encoder takes the concatenated multi-modal feature matrix as input and outputs a fused multi-modal unified semantic vector, which is used by the decision module of the subsequent autonomous agent to generate governance instructions, ensuring that the present embodiment can accurately capture the semantic association and propagation characteristics of different propagation topics in multi-modal data.

[0109] Preferably, the decision module outputs governance instructions through a multi-armed bandit-contextual bandit-policy gradient joint network combined with an international compliance rule base, including:

[0110] receiving the multi-modal unified semantic vector s output by the perception module, and extracting a compliance constraint embedding vector c corresponding to the current propagation scene from the international compliance rule base;

[0111] The multi-modal unified semantic vector s is spliced with the compliance constraint embedding vector c to form a context vector as a state input of a multi-armed contextual bandit;

[0112] For each governance action a in the action set A, an exploration-exploitation score is calculated using a multi-armed contextual bandit and a number of candidate actions are selected;

[0113] The candidate actions are fed into a policy gradient network together with the context vector, and a compliance-weighted policy gradient is calculated according to the following formula and the policy parameters are updated:

[0114]

[0115] Where θt is the policy network parameter vector; η is the learning rate; rt is the real-time propagation reward; λ is the compliance penalty coefficient; l t is the compliance violation degree; a t is the selected governance action; s t is the perception state vector; c t is the compliance embedding vector; π θ (·) is a parameterized policy probability distribution; is a gradient operator;

[0116] The candidate actions are re-evaluated according to the updated policy distribution, and the optimal action is selected as the final governance instruction output to the execution module, and the state-action-reward-penalty quadruple of this round is written into the experience replay buffer.

[0117] In this embodiment, preferably, the decision module is combined with a multi-armed contextual bandit and a policy gradient joint network, and combined with an international compliance rule library to output corresponding governance instructions. Specifically, the decision module first receives the multi-modal unified semantic vector output by the perception module, which represents the content features, propagation tendency and public opinion attributes of the target data stream. At the same time, the decision module automatically retrieves and extracts the embedding representation of compliance constraints from the pre-set international compliance rule library according to the country, region and corresponding propagation scenario to which the current data stream belongs. The compliance embedding includes but is not limited to data privacy clauses, propagation sensitivity levels, content safety restrictions, etc., and is converted into vector information that can be recognized by the model through rule analysis and semantic vectorization technology. Subsequently, the multi-modal unified semantic vector and the compliance constraint embedding vector are spliced to form a context vector, which is used as a state input of a multi-armed contextual bandit.

[0118] Further, the multi-arm context gambling machine calculates the exploration-exploitation scores of each governance action under the current propagation context in combination with the context vector for a preset set of governance actions (including filtering, desensitization, rate limiting, reordering, multi-language synchronization, etc.), and selects a number of candidate actions with higher scores. Subsequently, the candidate actions and the context vector are jointly input into a policy gradient network, and the governance strategy parameters are iteratively updated by dynamically adjusting the learning step size and the reward function in combination with historical governance records and the latest feedback information, and the benefits and risk balance of each candidate action are re-evaluated. Finally, the decision module selects a governance action with the optimal comprehensive benefit from the candidate actions as the final governance instruction and outputs it to the execution module for execution, while storing the state vector corresponding to the current round, the selected action, the actual governance benefit, and the corresponding violation penalty information in the experience replay buffer for subsequent continuous optimization of the governance strategy and acceleration of model convergence. The above parameter data are calculated and updated in real time by the system during operation, without the need for manual presetting or manual adjustment, thereby ensuring the automation and self-adaptation of system governance.

[0119] In the present embodiment, preferably, each autonomous agent exchanges state summaries and issues synchronization instructions through a message bus to achieve consistent cooperation of the global governance strategy. Specifically, each autonomous agent automatically generates a state summary including the current local policy vector, the cumulative average governance benefit value in the current period, and the corresponding timestamp after completing its local data flow governance decision. The state summary is written to a unified message bus through a message publishing-subscribing mechanism for real-time acquisition by other autonomous agents. The state summary is used to represent the running state and benefit performance of the current agent governance strategy, ensuring transparent information flow within the system and providing a data basis for subsequent global consistency calculation.

[0120] Further, each autonomous agent actively subscribes to the state summaries published by external agents associated with the theme tag it is responsible for through the message bus, and automatically aggregates to form a neighborhood state set. The neighborhood state set is used to describe the local governance strategy state and benefit performance of other related agents in the current system. The present embodiment dynamically identifies the neighborhood set according to preset rules, including but not limited to theme tag overlap degree, historical cooperation frequency, governance effect correlation, etc. At the same time, the present embodiment generates a trust weight for each neighborhood agent according to historical interaction data or preset rules, representing the influence proportion of its local strategy on the update of the current agent's global consistent strategy. The weight is preset in the system initialization stage and updated in real time according to the strategy convergence effect and governance benefit in the running stage.

[0121] After obtaining the neighborhood state set, each autonomous agent calculates an updated global consensus strategy using a weighted consensus algorithm, and caches the calculated consensus strategy vector to the local strategy library for execution in the next governance cycle. At the same time, each autonomous agent generates a synchronization instruction based on the updated consensus strategy vector and broadcasts it to the execution module through the message bus, so that all autonomous agents can execute the data flow governance task according to the consensus governance strategy in the subsequent cycle. The above-mentioned consistency calculation and synchronization instruction updating are automatically completed by the system, which ensures the collaborative consistency of multiple autonomous agents in a distributed environment and the global stability of the governance effect.

[0122] In the present embodiment, preferably, the execution module implements processing operations including filtering, desensitization, rate limiting, reordering, multi-language synchronization on the original data stream according to the governance instruction, and the specific implementation manner is as follows: the execution module first parses the governance instruction, extracts the corresponding operation type, execution order, target field and language label, and constructs an operation list as the basis for execution. Subsequently, according to the keywords, regular expressions and black and white list rules in the operation list, the text, image and audio content in the original data stream are identified and matched, and the content related to the illegal theme is deleted or shielded, forming a filtered data stream. On this basis, for the sensitive fields such as identity information, geographic positioning information and device identifier information contained in the filtered data stream, hierarchical marking and replacement operations are performed according to the preset sensitive level rules, forming a desensitization result data stream. Further, according to the preset bandwidth threshold and time window requirement in the governance instruction, the desensitization result data stream is subjected to traffic shaping operation, and the data packet sending rate is adjusted to obtain a rate-limited data stream. According to the priority information and time stamp in the governance instruction, the execution module reorders the rate-limited data stream, preferentially sends high-priority data packets, and delays sending low-priority data packets, to obtain an ordered data stream. If the operation list contains multi-language synchronization marks, the neural machine translation model and the terminology library are called based on the text content of the ordered data stream to generate a target language version aligned with the original language content, and the original media format is reused for packaging to obtain a multi-language synchronization data stream. Finally, the latest available data in each stage is taken as the final data stream, and the corresponding operation record, processing label and compliance status information are combined to form a governance result, which is stored in the governance result queue for subsequent channel division and evaluation module calling.

[0123] In the embodiment, preferably, when the governance result is generated, the embodiment automatically attaches a compliance risk score, a time urgency, and a content heat prediction value to the governance result, and writes them uniformly into a governance result queue as a basis for subsequent priority division. The compliance risk score is calculated according to the historical governance record of the data stream, the severity and frequency of the triggered violation rule; the time urgency is set according to the time urgency requirement and content update cycle of the field to which the data stream belongs; and the content heat prediction value is estimated in combination with real-time propagation trends, click rates, forwarding volumes and other data on the platform. All governance results are uniformly packaged as structured items and written into the governance result queue in real time for subsequent scheduling and calling.

[0124] Further, for each governance result in the governance result queue, the embodiment calls a compliance ontology model to perform content verification, counts the number of compliance clauses triggered, and calculates a priority score based on the compliance risk score, the time urgency, the content heat prediction value, and the number of triggered clauses. The score is mapped to a fixed interval by a normalization function and used for subsequent channel division. According to a set of preset score thresholds, the embodiment assigns governance results with a priority score higher than a first threshold to a fast channel for fast and low-latency processing, assigns governance results with a score between the first threshold and a second threshold to a review channel for manual review and in-depth governance, and assigns governance results with a score lower than the second threshold to a cold storage channel for long-term archiving and subsequent tracing. The embodiment updates and stores the channel label together with the governance result in the index structure of the governance result queue, ensuring the transparency and traceability of the data stream throughout its life cycle.

[0125] In the embodiment, after each governance period ends, the embodiment extracts all governance results from the governance result queue and calculates a comprehensive score according to time urgency, compliance, and propagation effect. First, the embodiment reads the generation timestamp corresponding to the governance result and compares it with the current time to obtain a time difference, which represents the time decay of the governance result. According to a preset time decay constant, which is set according to the sensitivity of different data types to time urgency, the embodiment calculates a time urgency decay factor corresponding to the time difference to reflect the time decay trend of the governance result. The timestamp information is directly read from the governance result queue, the current time is provided by the system clock, and the time decay constant is a fixed value set by the embodiment based on historical data experience and does not need to be adjusted manually.

[0126] Further, the embodiment calls the established compliance ontology model for rule checking for each governance result corresponding data stream, counts the number of triggered violation clauses in the governance process, and calculates the corresponding compliance penalty factor according to the number of triggers to reflect the compliance risk level of the data stream. For the evaluation of the propagation effect, the embodiment obtains the click volume, forwarding volume and audience stay time data of the data stream through the third-party data interface connected with the target audience platform. The foregoing data is used to measure the propagation effect and generate a propagation effect index accordingly. Finally, the embodiment uniformly normalizes the timeliness decay factor, the compliance penalty factor and the propagation effect index, calculates a comprehensive score, and the comprehensive score is used as the basis for subsequent governance strategy optimization and data channel division.

[0127] In the embodiment, preferably, after the end of each governance cycle, the embodiment collects the comprehensive scores corresponding to each autonomous agent in the current cycle and the governance decision parameters used by each autonomous agent in the cycle, and uniformly constructs a historical sample set. The comprehensive score is calculated by the timeliness, compliance and propagation effect, and is directly obtained from the evaluation module. The governance decision parameters of each autonomous agent are derived from the decision module currently executed by the autonomous agent, and the parameters are used to represent the state of the governance strategy network in the cycle. The embodiment forms a corresponding relationship between the comprehensive score and the governance parameters, continuously accumulates by period, forms a historical sample data set, and is used to support the subsequent strategy optimization process.

[0128] Further, the embodiment constructs a Gaussian process proxy model based on the historical sample set, the proxy model takes minimizing the difference between the governance effect and the strategy as the optimization target, evaluates the influence of different parameter configurations on the governance effect, and defines a multi-objective expected improvement function according to the influence, which is used to balance the balance between the governance effect and the compliance risk. Under the guidance of the proxy model, the embodiment automatically solves the optimal parameter adjustment increment by the Bayesian optimization method combined with the prediction performance of each historical sample in the model. The step adjustment coefficient is adaptively set according to the prediction variance of the proxy model on the sample, which guarantees the stability and effectiveness of the small step update. Finally, the obtained update parameter is written back to the decision module of each corresponding autonomous agent to replace the original parameter, so as to realize the adaptive migration and continuous optimization of the governance strategy in the next cycle.

[0129] Further, the comprehensive score of the embodiment is obtained by real-time calculation of the timeliness, compliance and propagation effect of the governance result. The decision parameters are derived from the strategy network output used by each autonomous agent in the current period, and the historical sample set is automatically accumulated according to the governance period. The training data of the Gaussian process proxy model is directly provided by the aforementioned historical sample set, and the hyperparameters in the model are automatically optimized by cross-validation or maximum likelihood estimation according to the fitting effect of the sample set. The multi-objective expected improvement function balances the governance effect and compliance risk, and the prediction error and distribution of the proxy model are automatically determined according to the historical data fitting. The step adjustment coefficient is adaptively set according to the uncertainty of the current parameter prediction by the proxy model, i.e. the variance, without pre-setting a fixed value.

[0130] In the embodiment, preferably, after the governance result comprehensive score calculation and decision module parameter update are completed, the embodiment further synchronously generates a traceable report containing the data stream governance result, rule trigger item, autonomous agent state and parameter update record. The report summarizes the final processing result of each data stream in the governance result queue, records the applicable governance instruction, the corresponding compliance rule clause number and the trigger number, the state summary of the autonomous agent in the current period, including the strategy parameters, the performance and the state synchronization information, and the historical sample data and the final update result in the governance strategy parameter update process in the current period. The above data is uniformly packaged as a structured report file, associated with a timestamp and a unique identifier of the data stream, stored in an audit archive database, used for external compliance audit, effect tracing and subsequent iteration optimization to provide data basis, and ensure the transparency, verifiability and reproducibility of the whole governance process.

[0131] Corresponding to the above method, as shown in Figure 2 The embodiment also provides an international communication data stream dynamic management system based on an Agent intelligent agent, which comprises:

[0132] A data acquisition and labeling unit is configured to distribute crawling tasks by distributed crawler nodes according to an extensible hash algorithm to crawl cross-platform and multi-lingual data streams, and to attach source identification, a timestamp, a topic label and a sensitivity weight to each data stream to form an original data pool with semantic indexing.

[0133] An agent construction and cooperation unit is configured to instantiate autonomous agents for each topic label in the original data pool. The autonomous agents comprise a perception module, a decision module and an execution module. The perception module encodes the multi-modal content of the original data pool by using a hybrid Transformer-CNN network. The decision module outputs governance instructions by a multi-armed contextual bandit-strategy gradient joint network and in combination with an international compliance rule library. Each autonomous agent exchanges a state summary and issues a synchronization instruction through a message bus, and a consistency algorithm is used to keep the global strategy consistent.

[0134] a data governance execution unit configured to implement, by using the execution module, processing operations including filtering, desensitization, rate limiting, reordering, and multi-language synchronization on the original data stream according to the governance instructions, to obtain governance results;

[0135] a priority channel division unit configured to write the governance results into a governance result queue, and divide data streams in the governance result queue into a fast channel, a review channel, and a cold storage channel according to a compliance ontology model and priority;

[0136] a comprehensive evaluation unit configured to extract governance results from the governance result queue after each governance cycle ends, and calculate a comprehensive score according to timeliness, compliance, and propagation effect;

[0137] a strategy optimization update unit configured to update parameters of the decision module in small steps based on the comprehensive score, by using a multi-objective Bayesian optimization driven model-agnostic meta-learning algorithm, and write the updated parameters back to each autonomous intelligent agent, to realize adaptive migration and optimization of the governance strategy.

[0138] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the system disclosed in the embodiments, the description is relatively simple because it corresponds to the method disclosed in the embodiments. The relevant parts can be referred to the description of the method.

[0139] The principles and implementation manners of the present application are described by using specific examples in the specification. The above description of the embodiments is only used to help understand the method of the present application and its core idea. For those skilled in the art, the specific implementation manners and application scope can be changed according to the idea of the present application. In conclusion, the content of the specification should not be understood as a limitation of the present application.

Claims

1. An Agent-based intelligent agent international communication data flow dynamic management method, characterized in that, The method comprises the following steps: distributing crawling tasks by distributed crawler nodes according to an extensible hash algorithm to crawl cross-platform and multi-language data streams, and attaching source identification, timestamps, topic labels, and sensitivity weights to each data stream to form a raw data pool with semantic indexing; instantiating autonomous agents for each topic label in the raw data pool; the autonomous agents include a perception module, a decision module, and an execution module; the perception module encodes the multi-modal content of the raw data pool using a hybrid Transformer-CNN network; the decision module outputs governance instructions by combining a multi-armed bandit-contextual bandit-policy gradient joint network with an international compliance rule base; each autonomous agent exchanges state summaries and issues synchronization instructions through a message bus, and maintains global policy consistency through a consistency algorithm; using the execution module to implement processing operations including filtering, desensitization, rate limiting, reordering, and multi-language synchronization on the raw data pool according to the governance instructions to obtain governance results; writing the governance results into a governance result queue, and dividing the data streams in the governance result queue into a fast lane, a review lane, and a cold storage lane according to the priority of a compliance ontology model; after each governance cycle, extracting governance results from the governance result queue and calculating a comprehensive score based on timeliness, compliance, and propagation effect; based on the comprehensive score, using a multi-objective Bayesian optimization-driven model-agnostic meta-learning algorithm to update the parameters of the decision module in small steps, and writing the updated parameters back to each autonomous agent to achieve adaptive migration and optimization of governance strategies; the decision module outputs governance instructions by combining a multi-armed bandit-contextual bandit-policy gradient joint network with an international compliance rule base, including: receiving the multi-modal unified semantic vector output by the perception module and extracting a compliance constraint embedding vector corresponding to the current propagation scenario from an international compliance rule base ; unifying the multi-modal semantic vectors embedding vectors for the compliance constraints concatenating to form a context vector as a state input to a multi-armed bandit For action sets Every governance action Explore using multi-armed contextual gambling machine calculations - utilize scores and select several candidate actions; sending the candidate actions together with the context vector into a policy gradient network, calculating a compliance-weighted policy gradient according to the following formula, and updating the policy parameters: ; wherein, is a policy network parameter vector; is a learning rate; is a real-time propagation reward; is a compliance penalty coefficient; is a compliance violation degree; is a selected governance action; is a perception state vector; is a compliance embedding vector; is a parameterized policy probability distribution; is a gradient operator; re-evaluating the candidate actions according to the updated policy distribution, selecting the optimal action as the final governance instruction output to the execution module, and writing the state-action-reward-penalty quadruple of this round into an experience replay buffer; based on the comprehensive score, using a multi-objective Bayesian optimization-driven model-agnostic meta-learning algorithm to update the parameters of the decision module in small steps, and writing the updated parameters back to each autonomous agent to achieve adaptive migration and optimization of governance strategies, including: At the end of each governance cycle, the comprehensive score corresponding to each autonomous agent is collected and its current decision parameter to construct a historical sample set ; wherein, is the comprehensive score of the th autonomous agent, is the decision parameter vector of the th autonomous agent at time ; for the history sample set for minimizing an objective training a gaussian process surrogate model and based on the gaussian process surrogate model defining a multi-objective expected hypervolume improvement function ; finding the optimal increment by multi-objective Bayesian optimization, with the formula being: ； wherein, to maximize the resulting optimal parameter increment; updating the rule execution in small steps using model-agnostic meta-learning, with the formula being: ； wherein, is a step length coefficient computed by the proxy model mean prediction variance adaptation; is the decision parameter vector after the small step update; updating the parameters write back to the decision module corresponding to the autonomous intelligent agent to complete the policy adaptive optimization.

2. The Agent-based intelligent entity international communication data flow dynamic management method according to claim 1, characterized in that, distributing crawling tasks by distributed crawler nodes according to an extensible hash algorithm to crawl cross-platform and multi-language data streams, and attaching source identification, timestamps, topic labels, and sensitivity weights to each data stream to form a raw data pool with semantic indexing, including: acquiring a preset data source; the data source includes news websites, social platforms, video platforms, content distribution networks, open data interfaces, and other information publishing channels, each data source corresponding to an access address, data format, and access protocol; Based on the access domain name, IP address and interface call information of the data source, a unique source identifier is generated, and the platform type, language type and time zone information corresponding to the data source are recorded as metadata basis; According to the computing power, network bandwidth and historical load of each crawler node, the task domain is dynamically divided by using a scalable hash algorithm, and the data sources are mapped and distributed to different distributed crawler nodes according to the platform type, language type and time zone information; Each distributed crawler node real-time crawls the target data stream based on the task domain rule, and adds the source identifier and collection timestamp information to each collected target data stream; For the collected target data stream, the theme label is extracted through natural language processing and theme modeling, and the sensitivity weight is calculated combined with the content sensitivity model; The source identifier, collection timestamp information, theme label and sensitivity weight are encapsulated with the target data stream as structured metadata, and are uniformly written into the original data pool.

3. The agent-based intelligent entity international communication data flow dynamic management method according to claim 1, characterized in that, The perception module encodes the multi-modal content of the original data pool by using a hybrid Transformer-CNN network, including: For the data streams associated with each theme label in the original data pool, text data, audio data, video data and image data in the data stream are extracted respectively to form a multi-modal data set; For the text data, a pre-trained Transformer encoding model is used for context modeling to output a text feature vector containing semantic relationships; For the audio data and the video data, the corresponding audio feature sequence and video frame image sequence are first extracted, and then input into a convolutional neural network to extract temporal and spatial local features, respectively, to output an audio feature vector and a video feature vector; For the image data, a convolutional neural network is directly inputted to extract spatial level features to output an image feature vector; The text feature vector, audio feature vector, video feature vector and image feature vector are spliced to construct a unified multi-modal semantic feature matrix; The multi-modal semantic feature matrix is inputted into a fusion layer to further model the cross-modal association relationship by using a fusion Transformer encoding layer to output a multi-modal unified semantic vector.

4. The agent-based intelligent entity international communication data flow dynamic management method according to claim 1, characterized in that, Each autonomous agent exchanges state summaries and issues synchronization instructions through a message bus, and maintains global policy consistency by using a consistency algorithm, including: Each autonomous agent generates a state summary after completing local governance decision; the state summary at least contains the current local policy vector, average revenue value and timestamp, and is written into the message bus through publishing and subscribing; Each autonomous agent subscribes to external state summaries related to its theme label from the message bus to construct a neighborhood state set; For the neighborhood state set, a weighted consistency algorithm is used to calculate a global consistent policy, and the specific update rule is: ； wherein, is the wheel time agent local policy vector of the agent; is the neighborhood fusion coefficient; is the neighborhood set visible to the agent on the message bus; is the trust weight from the agent to the agent satisfying ; The obtained consistent policy vector is written into the local policy cache, and a synchronization instruction is generated and broadcast to the execution module, so that each autonomous agent executes the consistent governance strategy in the next decision cycle.

5. The agent-based intelligent entity international communication data flow dynamic management method according to claim 1, characterized in that, The execution module is used to implement processing operations including filtering, desensitization, speed limiting, reordering, and multi-language synchronization on the original data pool according to the governance instruction, to obtain a governance result, including: The governance instruction is parsed to extract operation types, execution sequences, target fields, and language tags, and an operation list is constructed; Based on the keywords, regular expressions and black and white list rules in the operation list, content matching is performed on the text, image and audio content in the original data pool, content related to violation topics is deleted or shielded, and a filtered data stream is obtained ; The filtered data stream is desensitized The fields related to identity information, geographical positioning information and device identification information in the filtered data stream are desensitized, hierarchical marking and replacement are implemented by setting a sensitivity level, and a desensitized result data stream is obtained ; Based on the bandwidth threshold and time window specified in the governance instruction, the desensitization result data stream Implement traffic shaping, dynamically adjust the data packet sending rate, and generate a limited speed data stream ; According to the priority information and the time stamp in the governance instruction, the limited-speed data stream is reordered to obtain a sorted data stream wherein high-priority data packets are transmitted preferentially and low-priority data packets are queued for delay​ When the operation list contains multi-lingual synchronization markers, invoke a neural machine translation model based on the sorted data stream 's text content, and a terminology library to generate a target language version aligned with the original language content, and re-use the original media format packaging to obtain a multi-lingual synchronization data stream ; The filtered data stream , desensitized result data stream , speed-limited data stream , sorted data stream , and the latest available data in the multi-language synchronous data stream as the final data stream, combined with operation records, processing labels and compliance status, summarized and packaged as governance results.

6. The agent-based intelligent entity international communication data flow dynamic management method according to claim 1, characterized in that, The governance result is written into a governance result queue, and data streams in the governance result queue are divided into a fast channel, a review channel, and a cold storage channel according to a compliance ontology model and priority, including: When the governance result is generated, a compliance risk score, time urgency, and content heat prediction value are attached to the governance result, and written into the governance result queue; For each governance result in the governance result queue, the compliance ontology model is called to check the number of trigger clauses, and the priority score is calculated accordingly ； wherein, is a compliance risk score; is a time urgency; is a content heat prediction value; is a number of compliance clause triggers; is the priority score normalized by a hyperbolic tangent function; According to the priority score with a set of preset thresholds Channel division is performed: When assigned to the fast track; when assigned to the review track; when assigned to the cold storage track; The channel label is updated in the queue together with the governance result.

7. The agent-based intelligent entity international communication data flow dynamic management method according to claim 1, characterized in that, After each governance cycle ends, governance results are extracted from the governance result queue, and a comprehensive score is calculated according to time effectiveness, compliance, and propagation effect, including: reading a generation time stamp of the governance result, and a difference value with a current time , calculating a timeliness decay factor , and the calculation manner is ； wherein, is a time decay constant, is a time difference between the current time and the governance result generation time, denotes an age decay factor; a number of violated clauses triggered under a compliance ontology model based on the governance results a compliance penalty factor calculated as ； acquire a click quantity of a data stream corresponding to the governance result in a target audience , a forwarding quantity and an average stay duration , calculate a propagation effect index , and the calculation manner is: ； The comprehensive score is calculated according to the following formula based on the above indexes : ； wherein, is a comprehensive score, is a propagation effectiveness index, is a timeliness decay factor, is a compliance penalty factor, is a logistic function.

8. An Agent-based intelligent agent international communication data flow dynamic management system, characterized in that, including: A data collection and labeling unit is configured to dynamically allocate crawling tasks according to an extensible hash algorithm by a distributed crawler node to capture cross-platform, multi-language data streams, and attach source identification, timestamps, topic tags, and sensitivity weights to each data stream to form an original data pool with semantic indexing; An agent construction and collaboration unit is configured to instantiate autonomous agents for each topic tag in the original data pool; the autonomous agents include a perception module, a decision module, and an execution module; the perception module encodes multi-modal content of the original data pool using a hybrid Transformer-CNN network; the decision module outputs a governance instruction through a multi-armed contextual bandit-strategy gradient joint network combined with an international compliance rule library; each autonomous agent exchanges state summaries and issues synchronization instructions through a message bus, and maintains global policy consistency through a consistency algorithm; A data governance execution unit is configured to use the execution module to implement processing operations including filtering, desensitization, speed limiting, reordering, and multi-language synchronization on the original data pool according to the governance instruction, to obtain a governance result; A priority channel division unit is configured to write the governance result into a governance result queue, and divide data streams in the governance result queue into a fast channel, a review channel, and a cold storage channel according to a compliance ontology model and priority; A comprehensive evaluation unit is configured to extract governance results from the governance result queue after each governance cycle ends, and calculate a comprehensive score according to time effectiveness, compliance, and propagation effect; A strategy optimization update unit is configured to update the parameters of the decision module in small steps based on the comprehensive score using a multi-objective Bayesian optimization-driven model-agnostic meta-learning algorithm, and write the updated parameters back to each autonomous agent to achieve adaptive migration and optimization of governance strategies; The decision module outputs a governance instruction through a multi-armed contextual bandit-strategy gradient joint network combined with an international compliance rule library, including: receiving the multi-modal unified semantic vector output by the perception module and extracting a compliance constraint embedding vector corresponding to the current propagation scenario from an international compliance rule base ; unifying the multi-modal semantic vectors embedding the compliance constraints vectors concatenating to form a context vector as a state input to a multi-armed bandit For each governance action in the set of actions , compute an explore-exploit score using a multi-armed bandit and select a number of candidate actions . The candidate action is input into a policy gradient network together with the context vector, a compliance-weighted policy gradient is calculated according to the following formula, and the policy parameters are updated: ； wherein, is a policy network parameter vector; is a learning rate; is a real-time propagation reward; is a compliance penalty coefficient; is a compliance violation degree; is a selected governance action; is a perception state vector; is a compliance embedding vector; is a parameterized policy probability distribution; is a gradient operator; The candidate action is re-evaluated according to the updated policy distribution, the optimal action is selected as the final governance instruction, and the state-action-reward-punishment quadruple of this round is written into the experience replay buffer; Based on the comprehensive score, a multi-objective Bayesian optimization driven model-agnostic meta-learning algorithm is used to update the parameters of the decision module in small steps, and the updated parameters are written back to each autonomous intelligent agent to realize adaptive migration and optimization of the governance strategy, including: At the end of each governance cycle, the comprehensive score corresponding to each autonomous agent is collected and its current decision parameter , a historical sample set is constructed ; wherein, is the comprehensive score of the th autonomous agent, is the decision parameter vector of the th autonomous agent at time ; for the history sample set for minimizing objectives training a gaussian process surrogate model and based on the gaussian process surrogate model defining a multi-objective expected hypervolume improvement function ; The optimal increment is obtained by multi-objective Bayesian optimization, and the formula is: ； wherein, to maximize the resulting optimal parameter increment; The model-agnostic meta-learning update rule is used to perform small-step migration, and the formula is: ； wherein, is a step length coefficient computed by the proxy model mean prediction variance adaptation; is the decision parameter vector after the small step update; updating the parameters write back to the decision module corresponding to the autonomous intelligent agent to complete the policy adaptive optimization.

Images