A supervision method based on industry chain digitalization collaboration

Abstract

The present application relates to the technical field of digital intelligent supervision, and more particularly to a supervision method based on digital collaboration of an industrial chain, comprising efficient and reliable collection and fusion of heterogeneous data of the industrial chain through blockchain and federated learning technology, protection of data privacy and construction of a unified data base; dynamic learning of entity relationships of the industrial chain through a graph neural network, accurate identification of a key path and a vulnerability node, revelation of an implicit risk structure, and improvement of risk positioning capability; real-time calculation of risk entropy values and generation of early warnings through fusion of policy, behavior and topological data, realization of multi-dimensional risk quantification and dynamic monitoring, and enhancement of early warning accuracy; automatic triggering of supervision instructions based on smart contracts and risk matching, ensuring timeliness and accuracy of supervision measures, and reducing human intervention errors; and dynamic adjustment of a graph and risk model parameters through a reinforcement learning mechanism, realization of continuous optimization of supervision strategies and minimization of business interference.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital intelligent supervision, and in particular to a supervision method based on digital collaboration of an industrial chain. BACKGROUND

[0002] Under the background of current globalization and rapid development of digitization, efficient collaboration and safe supervision of the industrial chain have become key elements to ensure stable operation of the economy and promote industrial upgrading. However, there are still some challenges in current industrial chain supervision, including the following aspects: In traditional industrial chain supervision, different information systems and management software are usually used by various participating subjects (such as suppliers, manufacturers, distributors, etc.), resulting in non-uniform data formats, incompatible interfaces, data islands, and difficulties for regulatory agencies to obtain comprehensive and accurate data information; traditional supervision methods rely on post-examination and regular reports for risk identification, lack real-time and forward-looking, and for potential logistics disruptions and other risks in the industrial chain, traditional methods often fail to provide timely warning and effective response, resulting in lagging supervision and increasing instability and risk costs of the industrial chain; traditional supervision methods often rely on static rules and experience for risk identification, making it difficult to accurately identify dynamically changing industrial chain risks; in traditional supervision methods, the execution and feedback mechanism of supervision instructions is not perfect, making it difficult to evaluate and optimize the supervision effect; at the same time, there is a lack of effective tracking and feedback mechanism, making it difficult to monitor the execution of supervision instructions in real time, resulting in low transparency of supervision and difficulty in ensuring effective implementation of supervision measures. Therefore, the present application proposes a supervision method based on digital collaboration of an industrial chain. SUMMARY

[0003] The purpose of the present application is to solve the problems in the background art and propose a supervision method based on digital collaboration of an industrial chain.

[0004] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0005] A supervision method based on digital collaboration of an industrial chain, comprising:

[0006] S1, a cross-domain data acquisition and fusion network is constructed, based on blockchain anchoring and federated learning technology, heterogeneous business data of participating subjects in the industrial chain is standardized collected and credibly notarized, and an industrial chain collaboration data unit is generated;

[0007] S2, based on the generated industrial chain collaboration data unit, a dynamic industrial chain knowledge graph is constructed; wherein the dynamic industrial chain knowledge graph takes participating subjects as entity nodes, takes logistics, capital flow and information flow as relationship edges, and dynamically learns and updates the hidden correlation strength between nodes based on graph neural network, so as to label the key path and the vulnerability node in the graph;

[0008] S3, implanting a multi-modal risk perception model based on a dynamic industry chain knowledge graph; by constructing a multi-modal feature fusion network, the multi-modal input data composed of policy modal data, behavior modal data and topology modal data are fused and analyzed to realize dynamic evaluation of industry chain risk and generation of risk early warning signals;

[0009] S4, based on the risk early warning signal, the intelligent contract engine deployed in the blockchain is used to automatically execute the differentiated supervision instructions matched with the risk characteristics, wherein the differentiated supervision instructions include data verification request, fund flow temporary permission or logistics state enhanced tracking;

[0010] S5, the feedback data generated after the execution of the supervision instruction is returned to the dynamic industry chain knowledge graph and the multi-modal risk perception model, and the graph structure and model parameters are dynamically adjusted through the reinforcement learning mechanism to realize the adaptive iteration of the supervision strategy.

[0011] Further, a cross-domain data collection and fusion network is constructed, and based on the blockchain anchoring and federated learning technology, the heterogeneous business data of the participating subjects in the industry chain is standardized collected and credibly notarized, and the process of generating the industry chain collaborative data unit includes:

[0012] S11, through the data agent gateway deployed in the local of the participating subject, the original business data from its internal business system is collected, wherein the internal business system includes enterprise resource planning system, warehouse management system and logistics tracking system;

[0013] S12, the original business data is desensitized and standardized to form a standardized data package conforming to the preset data mode;

[0014] S13, each participating subject uses a federated learning model to extract features from the standardized data package locally, generates a local feature vector, and uploads the local feature vector and its hash value to the cross-domain fusion node;

[0015] S14, the cross-domain fusion node uses a model parameter aggregation algorithm to aggregate and update the federated learning model parameters of each participating subject, and generates a global feature view based on the aggregated model; at the same time, the cross-domain fusion node records the hash value of each local feature vector and the aggregation proof of the global feature view to the blockchain network, completing the credible notarization;

[0016] S15, a unique industry chain collaborative data unit identifier is assigned to each batch of data that has successfully completed the credible notarization, which is associated with the corresponding data source subject, timestamp and blockchain notarization information.

[0017] Further, based on the generated industry chain collaborative data unit, the process of constructing a dynamic industry chain knowledge graph includes:

[0018] S21, taking the participating subject as an entity node, the attributes of the entity node including enterprise type, registered capital, historical performance record and global feature vector;

[0019] S22, taking logistics, fund flow and information flow as the relationship edges between entity nodes, the attributes of the relationship edges including flow, frequency and stability index;

[0020] S23, based on the graph neural network, dynamically learning the dynamic industry chain knowledge graph, updating the embedding representation of the entity node by aggregating the attributes of the adjacent nodes and the features of the relationship edges, and calculating the implicit correlation strength between the entity nodes;

[0021] S24, according to the updated dynamic industry chain knowledge graph, using community discovery algorithm to identify industry chain clusters, and through key path analysis algorithm and node centrality calculation, marking out the key path and vulnerability node set which are crucial to the stable operation of the industry chain, wherein the vulnerability node set includes node identification and dynamically updated vulnerability score.

[0022] Further, based on the dynamic industry chain knowledge graph, a multi-modal risk perception model is implanted; by constructing a multi-modal feature fusion network, the multi-modal input data composed of policy modal data, behavior modal data and topology modal data are fused and analyzed, realizing the process of dynamic evaluation and risk early warning signal generation of the industry chain risk, including:

[0023] S31, collecting and constructing multi-modal input data, including policy modal data, behavior modal data and topology modal data:

[0024] S32, real-time collection and analysis of macro policy and regulation texts related to the industry chain, extracting policy keywords, control tendency and influence range vector through natural language processing technology to form a policy influence vector;

[0025] S33, from the dynamic industry chain knowledge graph, preferentially extracting the micro transaction behavior sequence of the vulnerability node and its first degree associated entity node to constitute the behavior time series data;

[0026] S34, real-time monitoring of the supply chain topology structure change of the dynamic industry chain knowledge graph: monitoring the adjacency relationship change of the vulnerability node and the mutation of the relationship edge weight on the key path, extracting and forming the topology evolution feature;

[0027] S35, constructing a multi-modal feature fusion network, which includes:

[0028] S351, modal exclusive coding layer: using time series convolution network to process behavior time series data, using graph convolution network to process topology evolution feature, using full connection network to process policy influence vector, and coding each modal data to a unified dimensional feature space;

[0029] S352, a cross-modal attention fusion layer: taking the encoded features of each modality as input, the cross-modal attention fusion layer calculates the correlation weight between the modalities, i.e., the cross-modal attention fusion layer weight; wherein, taking the encoded behavior timing feature as a query vector, and taking the encoded policy impact feature and the topological evolution feature as a key-value pair, the weighted fusion is performed to generate a comprehensive risk feature representation of the vulnerability node level;

[0030] S353, a risk entropy output layer: inputting the comprehensive risk feature representation of the vulnerability node level into a multi-layer perceptron, respectively outputting the global risk entropy value of the industrial chain and the local risk entropy value of each node in the vulnerability node set;

[0031] S36, when the global risk entropy value or the local risk entropy value exceeds the preset risk threshold, a risk warning signal is generated; wherein, the risk warning signal contains a risk feature vector, which at least includes the risk entropy value triggering this warning, the change trend of the risk entropy value, and the associated entity node information.

[0032] Further, based on the risk warning signal, the process of automatically executing the differentiated supervision instructions matched with the risk features through the smart contract engine deployed on the blockchain includes:

[0033] S41, realizing the differentiation of supervision instructions through pre-defined supervision rule logic, wherein the supervision rule logic maps the risk feature vector of the risk warning signal to the executable supervision instruction code;

[0034] S42, compiling the supervision rule logic into a smart contract deployed on the blockchain; wherein, the smart contract is embedded with an instruction parser, which converts the mapped supervision instruction code into a standardized instruction message recognizable by the target execution system;

[0035] S43, when the smart contract listens to the newly added risk warning signal on the chain and matches the pre-defined supervision rule, it automatically triggers the execution; wherein, the smart contract calls the oracle service of the blockchain, and routes the standardized instruction message to the corresponding target execution system via an encrypted channel, wherein the target execution system includes a data proxy gateway for executing data verification requests, a fund clearing platform for executing fund flow temporary permission, and a logistics tracking system for executing logistics state enhanced tracking;

[0036] S44, the smart contract simultaneously listens to the instruction execution receipt returned by the oracle service, and records the content of the risk warning signal triggered for execution, the standardized instruction message sent, and the execution result status obtained from the oracle as a supervision operation log on the blockchain.

[0037] Further, the feedback data generated after the execution of the regulatory instructions is fed back to the dynamic industry chain knowledge graph and the multi-modal risk perception model. Through the reinforcement learning mechanism, the graph structure and model parameters are dynamically adjusted to realize the adaptive iteration of the regulatory strategy. The process includes:

[0038] S51, collect the newly generated industry chain coordination data units of the target regulatory object and its associated entity nodes in a complete monitoring period after the issuance of the regulatory instructions, to form a feedback data set;

[0039] S52, input the feedback data set into two optimization channels simultaneously:

[0040] The first optimization channel is used to optimize the dynamic industry chain knowledge graph: based on the feedback data set, update the attributes and relationship edge weights of the related entity nodes in the graph, and perform community discovery, key path analysis and node centrality calculation to update the vulnerability node set and the key path;

[0041] The second optimization channel is used to optimize the multi-modal risk perception model: the feedback data set is used as a new training sample with a time label to perform incremental training on the multi-modal risk perception model;

[0042] S53, use a reinforcement learning algorithm based on strategy search to construct a joint reward function with the goal of minimizing the overall risk level after regulation and business interference;

[0043] S54, the reinforcement learning algorithm dynamically outputs optimization parameters through the strategy network, wherein the optimization parameters include: a first parameter set for adjusting the contribution weight of the graph neural network in the first optimization channel, and a second parameter set for adjusting the weight of the cross-modal attention fusion layer in the second optimization channel;

[0044] S55, by iteratively updating the optimization parameters, the dynamic industry chain knowledge graph correlation weight and the multi-modal risk perception model evaluation parameter are driven to be calibrated synchronously, thereby completing the adaptive optimization of the regulatory strategy.

[0045] Compared with the prior art, the beneficial effects of the present application are: by anchoring and federated learning technology through the blockchain, a cross-domain data collection and fusion network is constructed, efficient and reliable collection and standardized storage of heterogeneous data in the industry chain are realized, data privacy is protected, and a unified data representation is formed to lay a high-quality data foundation for subsequent analysis; by taking the participating subjects as nodes and logistics, capital flow and information flow as edges, a dynamic industry chain knowledge graph is constructed, the correlation strength between nodes is dynamically learned and updated based on the graph neural network, and the key path and vulnerability nodes in the graph are labeled, so that the dependence relationship of the industry chain can be reflected in real time, the implicit risk structure can be revealed, and the risk positioning and conduction analysis capability can be improved; by fusing policy modal, behavior modal and topological modal data, a multi-modal feature fusion network is constructed, risk entropy is calculated in real time and early warning signals are generated, multi-dimensional risk quantification is realized, early warning accuracy is enhanced, and potential risks can be intervened in advance to help regulatory agencies; based on the intelligent contract engine, regulatory instructions matching the risk features are automatically triggered to ensure timely and accurate regulatory measures, reduce human intervention errors and improve regulatory efficiency; through the reinforcement learning mechanism, the regulatory feedback data is fed back to the knowledge graph and risk model, the graph structure and model parameters are dynamically adjusted, the regulatory strategy is continuously optimized, and the long-term effectiveness of regulation is improved. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A flowchart of a regulatory method based on industry chain digital collaboration is proposed. DETAILED DESCRIPTION

[0047] 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 labor fall within the scope of protection of the present application.

[0048] REFERENCE Figure 1 A regulatory method based on industry chain digital collaboration, comprising:

[0049] S1, a cross-domain data collection and fusion network is constructed, and based on the blockchain anchoring and federated learning technology, the heterogeneous business data of the participating subjects in the industry chain is standardized collected and reliably stored to generate an industry chain collaborative data unit;

[0050] S2, based on the generated industry chain collaborative data unit, a dynamic industry chain knowledge graph is constructed; wherein the dynamic industry chain knowledge graph takes the participating subjects as entity nodes, takes logistics, capital flow and information flow as relationship edges, and dynamically learns and updates the implicit correlation strength between nodes based on the graph neural network, so as to label the key path and the vulnerability nodes in the graph;

[0051] S3, implanting a multi-modal risk perception model based on a dynamic industry chain knowledge graph; by constructing a multi-modal feature fusion network, multi-modal input data composed of policy modal data, behavior modal data and topology modal data are fused and analyzed to realize dynamic assessment of industry chain risks and generation of risk early warning signals;

[0052] S4, based on the risk early warning signal, through the smart contract engine deployed in the blockchain, automatically execute the differentiated supervision instructions matched with the risk characteristics, wherein the differentiated supervision instructions include data verification request, fund flow temporary permission or logistics state enhanced tracking;

[0053] S5, the feedback data generated after the execution of the supervision instruction is returned to the dynamic industry chain knowledge graph and the multi-modal risk perception model, and the graph structure and model parameters are dynamically adjusted through the reinforcement learning mechanism to realize the adaptive iteration of the supervision strategy.

[0054] It needs to be further explained that in the specific implementation process, a cross-domain data collection and fusion network is constructed, and based on the blockchain anchoring and federated learning technology, the heterogeneous business data of the participating subjects in the industry chain is standardized collected and credibly notarized, and the process of generating the industry chain collaborative data unit includes:

[0055] S11, through the data agent gateway deployed in the local of the participating subject, the original business data from its internal business system is collected, wherein the internal business system includes enterprise resource planning system, warehouse management system and logistics tracking system;

[0056] S12, the original business data is desensitized and standardized to form a standardized data package conforming to the preset data mode;

[0057] S13, each participating subject uses a federated learning model to extract features from the standardized data package locally, generates a local feature vector, and uploads the local feature vector and its hash value to the cross-domain fusion node;

[0058] S14, the cross-domain fusion node uses a model parameter aggregation algorithm to aggregate and update the federated learning model parameters of each participating subject, and generates a global feature view based on the aggregated model; at the same time, the cross-domain fusion node records the hash value of each local feature vector and the aggregation proof of the global feature view to the blockchain network, completing the credible notarization;

[0059] In the data collection and fusion stage, the federated learning model is a pre-trained deep autoencoder, the input layer dimension of which matches the dimension of the standardized data packet, and the output layer dimension is the compressed feature dimension; for example, the number of input layer neurons of the deep autoencoder is [X1], the number of neurons in the hidden layer is [X2], [X3] and [X4] respectively, and the number of output layer neurons is [X5]; the deep autoencoder adopts ReLU function as the activation function, and is pre-trained based on the public industry chain data set by using batch gradient descent method, and the training termination condition is set as the loss function value being less than 0.01 or the iteration number reaching 1000 times;

[0060] In the parameter aggregation link, after the cross-domain fusion node receives all the local feature vectors, the federated average algorithm is preferably used to aggregate and update the model parameters of the deep autoencoder; in the parameter aggregation, the weight of each participating subject's local model parameters is distributed according to the proportion of its data volume in the total data volume; in the parameter aggregation process, for abnormal values or noise data, the median filtering method is used to eliminate the influence;

[0061] In the blockchain storage link, the SHA-256 hash function is preferably used to calculate the hash value of each participating subject's uploaded local feature vector, and a Merkle tree is constructed; the consortium chain is selected as the blockchain network, and the proof-of-stake algorithm is used to package the root hash of the Merkle tree, the digest of the global feature view and the aggregation batch information of the federated average into one storage transaction, and broadcast to the blockchain network for consensus and persistent storage;

[0062] S15, assign a unique industry chain collaborative data unit identifier to each data batch that successfully completes the trusted storage, the identifier is associated with the corresponding data source subject, timestamp and blockchain storage information; wherein the data batch refers to a group of associated data that is assigned a unique identifier after being standardized, feature extraction and trusted storage in the data collection and fusion process;

[0063] It can be understood that by combining the federated learning and blockchain storage technology, efficient and trusted fusion and unified representation of cross-domain heterogeneous data are realized under the premise of protecting the data privacy of each participating subject, which lays a data foundation for subsequent construction of high-quality knowledge graph.

[0064] It needs to be further explained that in the specific implementation process, based on the generated industry chain collaborative data unit, the process of constructing a dynamic industry chain knowledge graph includes:

[0065] S21, taking the participating subject as an entity node, the attributes of the entity node including enterprise type, registered capital, historical performance record and global feature vector; wherein the global feature vector is derived from the global feature view generated after aggregation in the S14 step through federated learning, ensuring the consistency and comparability of node features in the global range;

[0066] S22, taking logistics, fund flow and information flow as the relationship edges between entity nodes, the attributes of the relationship edges including flow, frequency and stability index; it can be understood that the quantification of the attributes of the relationship edges is crucial when constructing the dynamic industry chain knowledge graph; wherein the flow refers to the amount of goods, the quantity of goods or the amount of information data flowing along the relationship edge in a certain monitoring period, which is a direct reflection of the relationship strength; the frequency refers to the number of interaction events (such as transactions, logistics delivery, information transmission) occurring on the relationship edge in a certain monitoring period, which reflects the activity of business transactions; the stability index is a comprehensive measure for evaluating the fluctuation of flow and frequency of the relationship edge in the historical period, for example, it can be obtained by calculating the coefficient of variation or stability index of the historical flow or frequency sequence, the higher the value, the more stable the business relationship represented by the edge;

[0067] S23, based on the graph neural network, the dynamic industry chain knowledge graph is dynamically learned, the embedding representation of the entity node is updated by aggregating the attributes of adjacent nodes and the features of the relationship edges, and the implicit correlation strength between the entity nodes is calculated;

[0068] S24, according to the updated dynamic industry chain knowledge graph, the industry chain cluster is identified by using the community discovery algorithm, and the key path and the vulnerability node set which are crucial to the stable operation of the industry chain are marked out by the key path analysis algorithm and the node centrality calculation, wherein the vulnerability node set includes the node identifier and the dynamically updated vulnerability score;

[0069] Wherein, when constructing the dynamic industry chain knowledge graph, the graph neural network preferably adopts a multi-head attention network architecture; for example, 8 attention heads are set, each attention head adopts a query-key-value attention mechanism for calculation; the cross-entropy loss function is used for training the graph neural network, the Adam optimizer is selected as the optimizer, the learning rate is set to 0.001, and the batch size is set to 32; the contribution weight of adjacent nodes and relationship edges to the embedding representation of the center node is calculated through the attention mechanism, and the new embedding representation of the center node is generated in a weighted aggregation manner; the implicit correlation strength between nodes is measured by calculating the cosine similarity between the updated node embedding vectors;

[0070] In labeling the critical path and the vulnerability node, the community discovery preferably adopts the Louvain algorithm, which plans the order of community merging by calculating the increment of modularity and gradually merges the communities to improve the modularity value by using the greedy algorithm; wherein, in order to improve its calculation efficiency, the parallel computing method is preferably used to accelerate it; the critical path analysis is performed by using the critical path method based on dynamic programming, and the longest path from the starting point to the ending point is calculated by setting a specific state transition equation, so as to label the node sequence on the critical path;

[0071] In the node centrality calculation, the betweenness centrality is included, and the Dijkstra algorithm is preferably used to calculate the shortest path of all node pairs; wherein, the dynamically updated vulnerability score is obtained by comprehensively calculating the betweenness centrality of the node, the frequency of the node appearing on the critical path, and the stability index of the associated edge of the node, specifically, in the monitoring period , the calculation formula of the node vulnerability score can be expressed as:

[0072] ,

[0073] In the formula, is the normalized betweenness centrality, is the frequency of the node appearing in all critical paths identified in the monitoring period , is the average stability index (normalized to the range of 0-1) of all relationship edges of the node in the monitoring period , is a preset weight coefficient, and satisfies ;

[0074] According to the historical data statistics, the betweenness centrality threshold is set, for example, the nodes with vulnerability scores higher than the median of all node scores are formally labeled and included in the vulnerability node set; this set is recalculated with each dynamic update of the dynamic industry chain knowledge graph, so as to realize the dynamic update of the vulnerability score;

[0075] It can be understood that through the dynamic analysis of the combination of the graph neural network and various graph algorithms, the hidden and dynamically changing dependency relationships and risk structures in the industry chain can be accurately captured, and the intelligent and refined identification of the critical path and the vulnerability node is realized.

[0076] It needs to be further explained that, in the specific implementation process, based on the dynamic industry chain knowledge graph, a multi-modal risk perception model is implanted; by constructing a multi-modal feature fusion network, the multi-modal input data composed of policy modal data, behavior modal data and topology modal data are fused and analyzed, and the process of dynamic evaluation of industry chain risk and generation of risk early warning signal includes:

[0077] S31, collect and construct multi-modal input data, including policy modal data, behavior modal data and topology modal data:

[0078] S32, real-time collection and analysis of macro policy and regulation texts related to industry chain, extraction of policy keywords, control tendency and influence range vector through natural language processing technology, formation of policy influence vector;

[0079] S33, from the dynamic industry chain knowledge graph, preferentially extract the micro transaction behavior sequence of the vulnerability node and its first associated entity node, and constitute the behavior time series data;

[0080] S34, real-time monitoring of the change of supply chain topology structure of dynamic industry chain knowledge graph: monitoring the change of adjacency relationship of vulnerability node and the mutation of relationship edge weight on critical path, extracting and forming topology evolution characteristics;

[0081] S35, construct a multi-modal feature fusion network, which includes:

[0082] S351, modal exclusive coding layer: respectively use time series convolution network to process behavior time series data, use graph convolution network to process topology evolution characteristics, and use full connection network to process policy influence vector, encode each modal data to a unified dimensional feature space;

[0083] S352, cross-modal attention fusion layer: take the encoded modal features as input, calculate the correlation weight between modal through cross attention mechanism, that is, the cross-modal attention fusion layer weight; among them, take the encoded behavior time series feature as query vector, take the encoded policy influence feature and topology evolution feature as key-value pair, perform weighted fusion, and generate comprehensive risk feature representation of vulnerability node level;

[0084] S353, risk entropy output layer: take the comprehensive risk feature representation of vulnerability node level Input into a multi-layer perception machine, respectively output the global risk entropy value of industry chain And the local risk entropy value of each node in the vulnerability node set ; wherein the calculation function of this layer is defined as:

[0085] ,

[0086] wherein, are weights and bias parameters of the multi-layer perceptron, is a Sigmoid activation function used to constrain the output in the range of (0, 1) representing the risk probability;

[0087] S36, when the global risk entropy value or the local risk entropy value exceeds the preset risk threshold , a risk warning signal is generated; wherein the risk warning signal contains a risk feature vector, which at least includes the risk entropy value (global or local) triggering this warning, the change trend of the risk entropy value and the associated entity node information; Specifically, the risk entropy value triggering this warning is directly taken from the real-time calculation result of the risk entropy value output layer; For global risk, it is the output global risk entropy value, for local risk, it is the local risk entropy value output for the specific vulnerability node; The change trend of the risk entropy value is obtained by recording the risk entropy value sequence in a specific time window (for example, the last 6 monitoring periods) and calculating its slope or first-order difference, which is used to represent whether the risk is rising, falling or stable; The associated entity node information: when the risk warning signal is triggered by the local risk entropy value, the information is the calculated target vulnerability node identifier, when triggered by the global risk entropy value, the information can be associated with the affected industry chain cluster identifier or core enterprise node identifier, these identifier information is directly obtained from the dynamic industry chain knowledge graph.

[0088] It needs to be further explained that, in the specific implementation process, based on the risk warning signal, through the smart contract engine deployed in the blockchain, the process of automatically executing the differentiated supervision instructions matched with the risk features includes:

[0089] S41, realize the differentiation of supervision instructions through pre-defined supervision rule logic, wherein the supervision rule logic maps the risk feature vector of the risk warning signal to the executable supervision instruction code; for example, map the warning signal with high entropy value and associated core enterprise to fund flow temporary license, and map the warning signal with steep change trend of entropy value to logistics state enhanced tracking;

[0090] S42, compile the supervision rule logic into a smart contract deployed on the blockchain; wherein the smart contract is embedded with an instruction parser, which converts the mapped supervision instruction code into a standardized instruction message recognizable by the target execution system;

[0091] S43, when the smart contract listens to the newly added risk warning signal on the chain and matches the pre-defined supervision rules, it automatically triggers the execution; wherein the smart contract calls the oracle service of the blockchain, and routes the standardized instruction message to the corresponding target execution system through the encryption channel, wherein the target execution system includes a data proxy gateway for executing data verification request, a fund clearing platform for executing fund flow temporary permission, and a logistics tracking system for executing logistics state enhanced tracking; it can be understood that the newly added on-chain is a specific blockchain event or data record that can be listened to by the smart contract, which is essentially a trusted existence proof of the risk warning signal on the blockchain; due to the tamper-proof nature of the blockchain, once this new action occurs, it provides a deterministic trigger basis for subsequent automatic execution; specifically, after the risk warning signal is generated, it will be stored on the chain in one or more ways, thereby becoming a newly added object that the smart contract can listen to: 1) after the multi-modal risk perception model (or its associated contract) calculates the risk warning signal, it will call a function of the smart contract, which does not change the on-chain state, but records an event containing the risk feature vector as a transaction log, and the smart contract can trigger execution by listening to the specific event; 2) write the risk feature vector or its hash value into a public on-chain mapping or array, this write operation itself is a state update, and the smart contract can listen to the change of the storage variable; 3) treat the risk feature vector itself as a data that needs to be stored, generate a unique identifier (such as hash) for it, and pack the identifier and the risk feature vector into a storage transaction, and record it on the blockchain, and the smart contract triggers by listening to the generation of such storage transaction;

[0092] S44, the smart contract listens to the instruction execution receipt returned by the oracle service, and records the risk warning signal content triggered by the execution, the standardized instruction message sent, and the execution result state obtained from the oracle as an unalterable supervision operation log on the blockchain.

[0093] It needs to be further explained that in the specific implementation process, the feedback data generated after the execution of the supervision instruction is returned to the dynamic industry chain knowledge graph and the multi-modal risk perception model, and the graph structure and model parameters are dynamically adjusted through the reinforcement learning mechanism to realize the adaptive iteration of the supervision strategy, which includes:

[0094] S51, collect the newly generated industry chain collaborative data units of the target supervision object and its associated entity nodes in a complete monitoring period after the supervision instruction is issued, to form a feedback data set;

[0095] S52, synchronously input the feedback data set to two optimization channels:

[0096] The first optimization channel is used for optimizing the dynamic industry chain knowledge graph: based on the feedback dataset, the attributes and relationship edge weights of related entity nodes in the graph are updated, and community discovery, key path analysis and node centrality calculation are performed to update the vulnerability node set and key path;

[0097] The second optimization channel is used for optimizing the multi-modal risk perception model: the feedback dataset is used as a new training sample with a time label to perform incremental training on the multi-modal risk perception model;

[0098] S53, a reinforcement learning algorithm based on policy search is adopted to construct a joint reward function with the goal of minimizing the overall risk level after regulation and business interference; Specifically, based on the policy search reinforcement learning algorithm, an actor-critic framework is implemented: the actor network acts as a policy network, responsible for outputting optimized policy parameters, i.e., the first parameter set and the second parameter set, according to the current environment state (i.e., the state of the latest dynamic industry chain knowledge graph and multi-modal risk perception model); the critic network acts as a value network, responsible for evaluating the long-term value of the policy taken by the actor network, i.e., estimating the cumulative reward (i.e., the expected value of the joint reward function) that can be obtained in the future under the current policy; the joint reward function is quantified as:

[0099] ,

[0100] wherein, represents the immediate reward obtained in the th iteration, is the reinforcement learning iteration step index, is the monitoring period time point corresponding to the th iteration, is the risk change evaluation window, represents the total number of regulatory instructions triggered within the th iteration evaluation window, is the action vector of the th regulatory instruction, is the regulatory instruction summation index, represents the estimated business interference cost of the th regulatory instruction, is the weight coefficient of the risk reduction reward, is the weight coefficient of the regulatory cost penalty; is the relative reduction rate of the global risk entropy value; is the total regulatory cost;

[0101] The goal of the critic network is to maximize the expected value of the future cumulative reward , wherein represents the future cumulative reward starting from the th reinforcement learning iteration, Discount factor ( (used to weigh the importance of current rewards against future rewards) For the offset of future reinforcement learning iteration steps, Indicates the future The reward obtained during each reinforcement learning iteration; the critic network is updated using temporal difference error, and the actor network is updated using the policy gradient method based on the advantage function given by the critic network; through this actor-critic interactive learning, the optimal optimization parameters can be found step by step, thereby achieving accurate, efficient and minimally invasive adaptive supervision;

[0102] S54. The reinforcement learning algorithm dynamically outputs optimization parameters through the policy network, wherein the optimization parameters include: a first set of parameters used to adjust the contribution weights of the graph neural network in the first optimization channel, and a second set of parameters used to adjust the weights of the cross-modal attention fusion layer in the second optimization channel;

[0103] S55. By iteratively updating and optimizing parameters, the synchronous calibration of the dynamic industrial chain knowledge graph association weights and the multimodal risk perception model evaluation parameters is driven, thereby achieving adaptive optimization of regulatory strategies. It can be understood that the dynamic industrial chain knowledge graph association weights are specifically the weight parameters involved in the update logic used to generate node embedding representations in the graph neural network, and the multimodal risk perception model evaluation parameters are specifically the weights and bias parameters involved in the calculation logic of the risk entropy value output layer. By optimizing these underlying parameters, the coordinated adjustment of the knowledge graph structure and risk perception capabilities is achieved.

[0104] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0105] It should be understood that determining B based on A does not mean determining B solely based on A; it also means determining B based on A and / or other information.

[0106] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0107] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A supervision method based on industry chain digitalization collaboration, characterized in that: S1, a cross-domain data collection and fusion network is constructed, based on blockchain anchoring and federated learning technology, heterogeneous business data of participating subjects in the industry chain is standardized collected and credibly stored, and an industry chain collaboration data unit is generated; S2, based on the generated industry chain collaboration data unit, a dynamic industry chain knowledge graph is constructed, and the implicit correlation strength between nodes is dynamically learned and updated based on graph neural network, so as to label the key path and the vulnerability node in the graph; S3, based on the dynamic industry chain knowledge graph, a multi-modal risk perception model is implanted; by constructing a multi-modal feature fusion network, the multi-modal input data composed of policy modal data, behavior modal data and topology modal data are fused and analyzed, and the dynamic assessment and risk warning signal generation of the industry chain risk are realized; S4, based on the risk warning signal, through the smart contract engine deployed in the blockchain, the differentiated supervision instructions matched with the risk characteristics are automatically executed; S5, the feedback data generated after the execution of the supervision instruction is returned to the dynamic industry chain knowledge graph and the multi-modal risk perception model, and the graph structure and the model parameters are dynamically adjusted through the reinforcement learning mechanism, so as to realize the adaptive iteration of the supervision strategy.

2. The method of claim 1, wherein, In S1, the process of constructing a cross-domain data collection and fusion network, based on blockchain anchoring and federated learning technology, heterogeneous business data of participating subjects in the industry chain is standardized collected and credibly stored, and an industry chain collaboration data unit is generated, including: S11, through the data agent gateway deployed in the participating subject, the original business data from its internal business system is collected, wherein the internal business system includes enterprise resource planning system, warehouse management system and logistics tracking system; S12, the original business data is desensitized and standardized to form a standardized data package conforming to the preset data mode; S13, each participating subject locally uses a federated learning model to extract features from the standardized data package, generates a local feature vector, and uploads the local feature vector and its hash value to a cross-domain fusion node; S14, the cross-domain fusion node uses a model parameter aggregation algorithm to update the federated learning model parameters of each participating subject, and generates a global feature view based on the aggregated model; at the same time, the cross-domain fusion node records the hash value of each local feature vector and the aggregation proof of the global feature view to the blockchain network, completing the credible storage; S15, a unique industry chain collaboration data unit identifier is assigned to each data batch that has successfully completed the credible storage, which is associated with the corresponding data source subject, timestamp and blockchain storage information.

3. The method of claim 2, wherein, In S2, the process of constructing a dynamic industry chain knowledge graph based on the generated industry chain collaboration data unit includes: S21, the participating subjects are taken as entity nodes, and the attributes of the entity nodes include enterprise type, registered capital, historical performance record and global feature vector; S22, logistics, capital flow and information flow are taken as the relationship edges between entity nodes, and the attributes of the relationship edges include flow, frequency and stability index; S23, based on the graph neural network, the dynamic industry chain knowledge graph is learned dynamically, the embedding representation of the entity node is updated by aggregating the attributes and relationship edge features of adjacent nodes, and the implicit correlation strength between the entity nodes is calculated; S24, according to the updated dynamic industry chain knowledge graph, the community discovery algorithm is used to identify the industry chain cluster, and the key path analysis algorithm and node centrality calculation are used to mark the key path and the vulnerability node set which are crucial to the stable operation of the industry chain, wherein the vulnerability node set includes node identification and dynamically updated vulnerability score.

4. The method according to claim 3, wherein, In S3, based on the dynamic industry chain knowledge graph, a multi-modal risk perception model is implanted; by constructing a multi-modal feature fusion network, the multi-modal input data composed of policy modal data, behavior modal data and topology modal data are fused and analyzed, and the process of dynamic evaluation and risk warning signal generation of industry chain risk is realized, including: S31, collect and construct multi-modal input data, including policy modal data, behavior modal data and topology modal data: S32, real-time collection and analysis of macro policy and regulation texts related to the industry chain, extraction of policy keywords, control tendency and influence range vector through natural language processing technology to form a policy influence vector; S33, from the dynamic industry chain knowledge graph, preferentially extract the micro transaction behavior sequence of the vulnerability node and its first degree associated entity node to constitute the behavior time series data; S34, real-time monitoring of the supply chain topology structure change of the dynamic industry chain knowledge graph: monitoring the adjacency relationship change of the vulnerability node and the mutation of the relationship edge weight on the key path, extracting and forming the topology evolution feature; S35, construct a multi-modal feature fusion network, which includes: S351, modal exclusive coding layer: use time series convolution network to process behavior time series data, use graph convolution network to process topology evolution feature, and use full connection network to process policy influence vector, and encode each modal data to a unified dimensional feature space; S352, cross-modal attention fusion layer: take the encoded modal features as input, calculate the correlation weight between modal through cross attention mechanism, that is, the cross-modal attention fusion layer weight; wherein, take the encoded behavior time series feature as query vector, take the encoded policy influence feature and topology evolution feature as key-value pair, perform weighted fusion, and generate comprehensive risk feature representation of vulnerability node level; S353, risk entropy output layer: input the comprehensive risk feature representation of vulnerability node level into a multi-layer perception machine, and output the global risk entropy value of the industry chain and the local risk entropy value of each node in the vulnerability node set, respectively; S36, when the global risk entropy value or the local risk entropy value exceeds the preset risk threshold, a risk warning signal is generated; wherein, the risk warning signal contains a risk feature vector, which at least includes the risk entropy value triggering this warning, the change trend of the risk entropy value and the associated entity node information.

5. The method according to claim 4, wherein, In S4, based on the risk warning signal, through the smart contract engine deployed in the blockchain, the process of automatically executing the differentiated supervision instructions matched with the risk features includes: S41, differentiating the supervision instructions by a pre-defined supervision rule logic, wherein the supervision rule logic maps the risk feature vector of the risk early warning signal to a supervision instruction code that can be executed; S42, compiling the supervision rule logic into a smart contract deployed on the blockchain; wherein the smart contract is embedded with an instruction parser to convert the supervision instruction code obtained by the mapping into a standardized instruction message recognizable by the target execution system; S43, when the smart contract listens to a newly added risk early warning signal on the chain that matches the pre-defined supervision rule, it automatically triggers execution; S44, the smart contract listens to the instruction execution return from the oracle service at the same time, and records the content of the risk early warning signal triggered for execution, the standardized instruction message sent, and the execution result status obtained from the oracle as a supervision operation log on the blockchain.

6. The method according to claim 5, wherein, In S43, the smart contract calls the oracle service of the blockchain, and routes the standardized instruction message to the corresponding target execution system via an encrypted channel, wherein the target execution system includes a data proxy gateway for executing data verification requests, a fund clearing platform for executing fund flow temporary permission, and a logistics tracking system for executing logistics state enhanced tracking.

7. The method according to claim 5, wherein, In S5, the feedback data generated after the execution of the supervision instruction is returned to the dynamic industry chain knowledge graph and the multi-modal risk perception model, and the graph structure and model parameters are dynamically adjusted through the reinforcement learning mechanism to realize the adaptive iteration of the supervision strategy, including: S51, collecting the newly generated industry chain collaborative data units of the target supervision object and its associated entity nodes within a complete monitoring period after the issuance of the supervision instruction, to form a feedback data set; S52, synchronously inputting the feedback data set into two optimization channels: the first optimization channel is used to optimize the dynamic industry chain knowledge graph; the second optimization channel is used to optimize the multi-modal risk perception model; S53, using a reinforcement learning algorithm based on policy search to construct a joint reward function with the goal of minimizing the overall risk level after supervision and business interference; S54, the reinforcement learning algorithm dynamically outputs optimization parameters through a policy network; S55, by iteratively updating the optimization parameters, the dynamic industry chain knowledge graph correlation weight and the multi-modal risk perception model evaluation parameter are simultaneously calibrated, thereby completing the adaptive optimization of the supervision strategy.

8. The method according to claim 7, wherein, In S52, the process of the first optimization channel for optimizing the dynamic industry chain knowledge graph includes: updating the attributes and relationship edge weights of the related entity nodes in the graph based on the feedback data set, and performing community discovery, key path analysis and node centrality calculation to update the vulnerability node set and the key path.

9. The method according to claim 7, wherein, In S52, the process of the second optimization channel for optimizing the multi-modal risk perception model includes: taking the feedback data set as a new training sample with a time label, and performing incremental training on the multi-modal risk perception model.

10. The method of claim 7, wherein the method is based on digitalization of the industrial chain. In S54, the optimization parameters include: a first parameter set for adjusting the contribution weight of the graph neural network in the first optimization channel, and a second parameter set for adjusting the weight of the cross-modal attention fusion layer in the second optimization channel.

Images