A multi-source heterogeneous data trusted fusion method in a blockchain environment

Abstract

The application discloses a kind of multi-source heterogeneous data trusted fusion methods under blockchain environment, it is related to blockchain data fusion technical field, comprising the following steps: S1, by blockchain consortium chain node, the identity authentication of multi-source heterogeneous data source is carried out, the integrity check is carried out to multi-source heterogeneous data passing through identity authentication, hierarchical filtering is carried out according to the check result, all operations and results are chained and stored as evidence;S2, the multi-source heterogeneous data that is not intercepted after hierarchical filtering is preprocessed, and standardized feature data is obtained.The multi-source heterogeneous data trusted fusion method under blockchain environment proposed in the application realizes data full-link trusted control and reverse traceability through the deep combination of blockchain technology and multi-source heterogeneous data fusion, builds a full-process trusted system from data source identity authentication, transmission integrity check to fusion process black-box cracking, effectively suppresses the interference of low-trust abnormal data on the result.

Description

Technical Field

[0001] This invention relates to the field of blockchain data fusion technology, specifically to a trusted fusion method for multi-source heterogeneous data in a blockchain environment. Background Technology

[0002] As business operations in areas such as financial risk control and big data analytics continue to deepen, the decision-making accuracy and business adaptability of risk control models heavily rely on multi-source, cross-platform data support. The required data covers various formats, including structured data within business systems, semi-structured data from third-party platforms, and unstructured data such as text and images. Currently, multimodal data fusion technology has achieved feature extraction and unified processing of heterogeneous data in different formats, laying the foundation for collaborative applications of cross-format data. Blockchain technology, with its immutable and traceable characteristics, is also being gradually applied to data trust management. The combination of these two technologies has become an important development direction in the field of multi-source heterogeneous data fusion.

[0003] Existing technical solutions still have many aspects that need optimization in practical applications. Multimodal data fusion solutions mostly focus on the weight allocation of feature semantic associations, without fully adapting and optimizing the fusion weights based on data trust levels, making it difficult to avoid interference from low-trust data in the fusion results. Existing blockchain and data fusion solutions mostly only implement basic identity verification and result notarization at the data input and output stages, failing to deeply integrate blockchain trust capabilities into the core processes of data preprocessing and feature fusion. This makes it impossible to achieve trusted control and reverse traceability of the entire data chain. Furthermore, there is still significant room for improvement in the adaptation of heterogeneous data for differentiated verification, processing efficiency, and adaptability to business scenarios. To address these issues, we propose a trusted fusion method for multi-source heterogeneous data in a blockchain environment. Summary of the Invention

[0004] To address the aforementioned technical issues, this paper provides a trusted fusion method for multi-source heterogeneous data in a blockchain environment. This technical solution solves the problems of existing multimodal fusion schemes that do not optimize fusion weights based on data trust levels, making them susceptible to interference from low-trust data; the shallow integration of blockchain and data fusion, which only achieves basic evidence storage at the beginning and end stages and lacks full-link trusted control and traceability capabilities; and the fact that there is still room for improvement in heterogeneous data verification and adaptation, processing efficiency, and scenario adaptability.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A trusted fusion method for multi-source heterogeneous data in a blockchain environment includes the following steps: S1. Through blockchain consortium chain nodes, perform identity authentication on multi-source heterogeneous data sources, perform integrity verification on multi-source heterogeneous data that has passed identity authentication, perform hierarchical filtering based on the verification results, and store all operations and results on the blockchain for evidence. S2. Preprocess the multi-source heterogeneous data that was not intercepted after hierarchical filtering to obtain standardized feature data. All preprocessing rules and operation records are hashed and uploaded to the chain, and the corresponding on-chain trust scores are associated with them. S3. Through a feature-level fusion model based on an attention mechanism, the preprocessed feature data is fused, the trust score on the blockchain is integrated into the attention weight calculation, a unified fusion feature vector is generated, and the core parameters of the fusion process are stored on the blockchain. S4. Store the fused feature vector on the blockchain, standardize the output of the fused feature vector to the risk control model, and provide full-link traceability and verification capabilities for the fused data.

[0006] Preferably, in step S1, the authentication of multi-source heterogeneous data sources specifically involves: The management node of the blockchain consortium chain conducts qualification review of the data sources applying for access. After the review is approved, a unique distributed digital identity (DID) identifier is generated on the chain for each data source, and an asymmetric public-private key pair is generated accordingly. The public key is bound to the DID identifier and written into the DID registry on the blockchain to complete the on-chain registration, while the private key is kept by the data source itself. When a data source sends a batch of data, it uses its own private key to digitally sign the unique batch number of that batch of data. After the consortium blockchain verification node obtains the data, it completes the signature verification through the public key corresponding to the DID registry on the chain. If the signature verification is successful, the identity of the data source is confirmed to be legitimate.

[0007] Preferably, in step S1, the integrity verification of the data that has passed identity authentication specifically involves: For a single data entry, a unique hash value is generated using the SHA-256 algorithm and transmitted synchronously with the data. The verification node recalculates the hash value of the received data and compares it with the transmitted hash value to complete the integrity verification of the single data entry. For heterogeneous data in the same batch, a Merkle tree is constructed according to the data form. The generated Merkle root hash is synchronously uploaded to the blockchain for evidence storage. After the batch data transmission is completed, the verification node reconstructs the Merkle tree and compares it with the on-chain Merkle root hash to complete the batch data integrity verification.

[0008] Preferably, in step S1, the hierarchical filtering based on the verification results specifically involves: Based on the sequence of the verification process, a progressive three-level exception interception mechanism is set up: First, perform data source identity authentication. If the identity verification fails, it is judged as a level 1 anomaly, and all data in that batch is directly intercepted, preventing it from entering the subsequent process. Simultaneously, an on-chain anomaly alarm record is generated and a data source qualification review is triggered. If the identity verification passes, the data integrity verification is performed. If the integrity verification fails, it is judged as a level 2 exception, and the corresponding abnormal data is returned to the data source for retransmission. If the number of retransmissions of a single batch of data from the same data source exceeds the limit, the data source whitelist temporary freezing mechanism is triggered. If the integrity check passes, a data format and quality check is performed. If there are format abnormalities, missing values ​​exceeding the limit, or value range abnormalities, it is judged as a level three abnormality. Abnormal data is marked with an abnormal label on the chain and the abnormality type is recorded synchronously, and it is prohibited from directly entering the feature fusion process.

[0009] Preferably, in step S2, the preprocessing of the multi-source heterogeneous data that was not intercepted after hierarchical filtering specifically includes: For structured data, the following steps are performed sequentially: unified metadata mapping, data cleaning, and feature normalization, to generate a standardized structured feature matrix. For semi-structured data, the schema adaptive parsing engine completes the nested structure recognition and flattening process, maps the parsed core business fields to a unified metadata standard, and generates a standardized semi-structured feature matrix. For unstructured data, high-dimensional features are extracted using a pre-trained model, and then dimensionality reduction is performed to generate feature vectors that match the dimensions of structured data. All preprocessing rules and operation records are synchronized on the blockchain and bound to the on-chain trust score of the corresponding data.

[0010] Preferably, in step S3, before fusing the preprocessed feature data, the unified representation and dimension alignment of the features of the multi-source heterogeneous data are first completed. The preprocessed structured, semi-structured, and unstructured data are respectively generated into initial feature vectors with unified dimensions through the Embedding layer. Then, all initial feature vectors are mapped to the same feature space through the linear transformation layer to generate feature matrices with consistent dimensions.

[0011] Preferably, in step S3, incorporating the trust score on the blockchain into the attention weight calculation specifically involves: The data source DID trust score, data integrity verification score, and historical data quality score are obtained from the blockchain and normalized by the sigmoid function to generate a trustworthy prior weight. A multi-head self-attention mechanism is used to calculate the dimension-aligned feature matrix, capture the semantic association and business value between features, and generate semantic association weights. The trusted prior weights and semantic association weights are weighted and fused to generate the final fused attention weights. The feature matrix is ​​then weighted using the fused attention weights.

[0012] Preferably, in step S3, generating a unified fusion feature vector specifically involves: The dimensionally aligned feature matrix is ​​weighted and summed using fused attention weights to obtain a weighted optimized fused feature matrix. The fused feature matrix is ​​input into the fully connected layer and then transformed nonlinearly to generate a unified fused feature vector of fixed dimensions. The weighting coefficients, model parameters, and core hash values ​​from the calculation process are simultaneously recorded on the blockchain for evidence storage.

[0013] Preferably, in step S4, the on-chain storage of the fused feature vector specifically involves: It adopts a distributed storage architecture that combines on-chain and off-chain collaboration. On-chain storage includes the SHA-256 hash value of the fused feature vector, the hash of all parameters of the fusion process, the corresponding data source DID list, the data batch number, the timestamp and the signature of the operation node. All on-chain storage content is written into the block after consensus among all nodes of the consortium blockchain. The original fusion feature vector, full-process detailed data, and original data backup are stored off-chain. Off-chain data is uniquely anchored by on-chain hash values.

[0014] Preferably, in step S4, the standardized output of the fused feature vector to the risk control model, while providing end-to-end traceability and verification capabilities for the fused data, specifically includes: Two types of standardized interfaces are designed to interface with the risk control model. The standardized interfaces include: a trusted data synchronization interface and a full-link trusted verification interface. The trusted data synchronization interface uses a standardized communication protocol to push the fused feature vector to the risk control model, and simultaneously attaches the corresponding on-chain evidence hash value, block height and trusted verification certificate. The end-to-end trusted verification interface allows the risk control model to query the end-to-end trusted information of the corresponding data by data batch number or hash value. Based on the on-chain DID identifier and hash anchoring mechanism, the entire chain of reverse tracing from the fusion result to the original data source is performed, and a tracing report is generated.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: The proposed method for trusted fusion of multi-source heterogeneous data in a blockchain environment achieves trusted control and reverse traceability of the entire data chain through the deep integration of blockchain technology and multi-source heterogeneous data fusion. It constructs a trusted system covering the entire process from data source identity authentication and transmission integrity verification to black-box cracking of the fusion process. It ensures the authenticity of the data source through distributed digital identity identification and asymmetric encryption, and verifies the data's immutability by combining the SHA-256 algorithm and Merkle tree technology. It integrates on-chain trust scores into the attention weight calculation of the multimodal fusion algorithm, giving high-trust data a greater weight in feature fusion and effectively suppressing the interference of low-trust abnormal data on the results. The on-chain and off-chain collaborative storage architecture anchors on-chain and off-chain data through hash values, ensuring data immutability and solving the efficiency bottleneck of full on-chain storage. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0018] Reference Figure 1 As shown, a trusted fusion method for multi-source heterogeneous data in a blockchain environment includes the following steps: S1. Through blockchain consortium chain nodes, perform identity authentication on multi-source heterogeneous data sources, perform integrity verification on multi-source heterogeneous data that has passed identity authentication, perform hierarchical filtering based on the verification results, and store all operations and results on the blockchain for evidence. In step S1, the authentication of multi-source heterogeneous data sources specifically involves: The management node of the blockchain consortium chain conducts qualification review of the data sources applying for access. After the review is approved, a unique distributed digital identity (DID) identifier is generated on the chain for each data source, and an asymmetric public-private key pair is generated accordingly. The public key is bound to the DID identifier and written into the DID registry on the blockchain to complete the on-chain registration, while the private key is kept by the data source itself. When a data source sends a batch of data, it uses its own private key to digitally sign the unique batch number of that batch of data. After the consortium blockchain verification node obtains the data, it completes the signature verification through the public key corresponding to the DID registry on the chain. If the signature verification is successful, the identity of the data source is confirmed to be legitimate.

[0019] The entire process of issuing, verifying, renewing, and revoking digital certificates requires consensus approval from consortium blockchain nodes before being written to the blockchain. All operation records are permanently stored and traceable. Approved data sources are simultaneously added to the on-chain access whitelist. Only data sources on the whitelist can initiate data transmission requests; all data requests from nodes not on the whitelist are directly blocked.

[0020] In step S1, the integrity verification of the data that has passed identity authentication specifically involves: For a single data entry, a unique hash value is generated using the SHA-256 algorithm and transmitted synchronously with the data. The verification node recalculates the hash value of the received data and compares it with the transmitted hash value to complete the integrity verification of the single data entry. For heterogeneous data in the same batch, a Merkle tree is constructed according to the data form. The generated Merkle root hash is synchronously uploaded to the blockchain for evidence storage. After the batch data transmission is completed, the verification node reconstructs the Merkle tree and compares it with the on-chain Merkle root hash to complete the batch data integrity verification.

[0021] The hash value calculation for a single data entry covers all fields of the data, including metadata and all business data, preventing undetectable partial tampering. The construction of the customized Merkle tree follows rules to adapt to heterogeneous data formats. Structured data uses individual data records as leaf nodes, semi-structured XML or JSON data uses parsed core field units as leaf nodes, and unstructured data uses feature blocks as leaf nodes. All leaf nodes are sorted in ascending order by batch number, and hash values ​​are calculated for each pair. The Merkle root hash is generated iteratively upwards layer by layer, ensuring that any tampering with any data within a batch can be quickly located through Merkle tree verification.

[0022] In step S1, the hierarchical filtering based on the verification results specifically involves: Based on the sequence of the verification process, a progressive three-level exception interception mechanism is set up: First, perform data source identity authentication. If the identity verification fails, it is judged as a level 1 anomaly, and all data in that batch is directly intercepted, preventing it from entering the subsequent process. Simultaneously, an on-chain anomaly alarm record is generated and a data source qualification review is triggered. If the identity verification passes, the data integrity verification is performed. If the integrity verification fails, it is judged as a level 2 exception, and the corresponding abnormal data is returned to the data source for retransmission. If the number of retransmissions of a single batch of data from the same data source exceeds the limit, the data source whitelist temporary freezing mechanism is triggered. If the integrity check passes, a data format and quality check is performed. If there are format abnormalities, missing values ​​exceeding the limit, or value range abnormalities, it is judged as a level three abnormality. Abnormal data is marked with an abnormal label on the chain and the abnormality type is recorded synchronously, and it is prohibited from directly entering the feature fusion process.

[0023] If a single batch of data from the same data source fails to pass retransmission more than three times, a temporary freeze mechanism for the data source whitelist is triggered. During the freeze period, all data transmission requests from that data source are blocked until qualification review and compliance rectification are completed. Abnormal data marked as Level 3 abnormal can only enter the targeted cleaning process in the preprocessing stage and cannot directly enter the subsequent feature fusion step. All changes to the abnormality handling rules must be approved by the consensus of all nodes on the consortium blockchain before taking effect on the chain, preventing unauthorized tampering of the handling rules. All abnormality handling processes and results are synchronously written to the blockchain for notarization, ensuring that the entire abnormality handling process is traceable and auditable.

[0024] S2. Preprocess the multi-source heterogeneous data that was not intercepted after hierarchical filtering to obtain standardized feature data. All preprocessing rules and operation records are hashed and uploaded to the chain, and the corresponding on-chain trust scores are associated with them. In step S2, the preprocessing of the multi-source heterogeneous data that was not intercepted after hierarchical filtering specifically involves: For structured data, the following steps are performed sequentially: unified metadata mapping, data cleaning, and feature normalization, to generate a standardized structured feature matrix. For structured data, a unified metadata standard was developed based on financial risk control scenarios. This standard clarifies field naming conventions, data type constraints, and encoding format requirements, completing the mapping of synonymous field names across data sources, unified data type conversion, and standardization of UTF-8 encoding format. This addresses the issues of heterogeneous field definitions and inconsistent encoding across different business systems. Data cleaning is performed based on on-chain anomaly tags. For continuous features, missing values ​​are filled using the median of the same field from the same data source; for discrete features, missing values ​​are filled using the mode of the same field from the same data source; outliers are corrected using the 3σ principle; and full deduplication of duplicate data is performed simultaneously. For cleaned continuous numerical features, min-max normalization is applied, linearly mapping feature values ​​to the interval between 0 and 1. The calculation formula is as follows:

[0025] In the formula, These are the normalized eigenvalues. These are the original eigenvalues. This is the minimum value of the feature field in the corresponding batch of data. This represents the maximum value of the feature field in the corresponding batch of data. Discrete categorical features are processed using one-hot encoding, ultimately generating a standardized structured feature matrix with uniform dimensions. All filling, correction, and normalization rules and operation details generate corresponding hash values.

[0026] For semi-structured data, the schema adaptive parsing engine completes the nested structure recognition and flattening process, maps the parsed core business fields to a unified metadata standard, and generates a standardized semi-structured feature matrix. For semi-structured data, the schema adaptive parsing engine can automatically identify the nesting level, core business fields, and auxiliary attribute values ​​of XML or JSON format data. Based on preset risk control field priority rules, it completes the flattening decomposition of the nested structure, solving the problems of inconsistent nesting levels and non-standard field definitions in semi-structured data. The parsed core business fields are mapped to a unified metadata standard for risk control scenarios, and non-core auxiliary attributes are tagged and retained. The extracted field information, parsing rules, and mapping logic all generate corresponding hash values, which are synchronously stored on the blockchain, ultimately generating a standardized semi-structured feature matrix.

[0027] For unstructured data, high-dimensional features are extracted using a pre-trained model, and then dimensionality reduction is performed to generate feature vectors that match the dimensions of structured data. All preprocessing rules and operation records are synchronized on the blockchain and bound to the on-chain trust score of the corresponding data.

[0028] For unstructured data, text-based data employs a BERT pre-trained model fine-tuned for financial risk control scenarios to extract keyword features, semantic features, and sentiment and entity features related to default risk. Image-based data utilizes a CNN convolutional neural network paired with an OCR text recognition module to extract visual contour features, text recognition features, document authenticity verification features, and tampering trace recognition features. The extracted high-dimensional sparse features are linearly transformed and dimensionality reduced through fully connected layers to generate dense feature vectors that perfectly match the feature dimensions of the structured data. The model version used for feature extraction, training hyperparameters, and inference core parameters all generate corresponding hash values, which are simultaneously stored on the blockchain and deeply bound to the on-chain trust score of the corresponding data source, providing a reliable prior basis for subsequent fusion processes.

[0029] S3. Through a feature-level fusion model based on an attention mechanism, the preprocessed feature data is fused, the trust score on the blockchain is integrated into the attention weight calculation, a unified fusion feature vector is generated, and the core parameters of the fusion process are stored on the blockchain. In step S3, before fusing the preprocessed feature data, the unified representation and dimension alignment of the features of the multi-source heterogeneous data are first completed. The preprocessed structured, semi-structured, and unstructured data are respectively generated into initial feature vectors with unified dimensions through the Embedding layer. Then, all initial feature vectors are mapped to the same feature space through the linear transformation layer to generate feature matrices with consistent dimensions.

[0030] The feature outputs of the preprocessed structured, semi-structured, and unstructured data are encoded and mapped through independent embedding layers to generate initial feature vectors with uniform dimensions, thus achieving a unified feature representation for the three types of heterogeneous data. Then, a shared linear transformation layer maps these initial feature vectors from different sources and semantic spaces to the same high-dimensional feature space, generating feature matrices with completely consistent row and column dimensions. This eliminates the differences in feature dimensions and semantic space between heterogeneous data, laying a unified data foundation for subsequent feature fusion calculations.

[0031] In S3, incorporating the trust score on the blockchain into the attention weight calculation specifically involves: The data source DID trust score, data integrity verification score, and historical data quality score are obtained from the blockchain and normalized by the sigmoid function to generate a trustworthy prior weight. The trusted prior weights are generated by retrieving the data source DID trust score, data integrity verification score, and historical data quality score from the blockchain, and then normalizing them using the sigmoid function. The calculation formula is as follows:

[0032] In the formula, This represents the reliable prior weight for the corresponding feature, with a value ranging from 0 to 1. The on-chain trust score corresponding to the data source DID ranges from 0 to 100 and is calculated based on the data source's qualification level and historical compliance records. The integrity score for the corresponding data ranges from 0 to 100 and is calculated by combining the hash verification result and the data integrity score. The historical data quality score for the corresponding data source ranges from 0 to 100 and is calculated by combining the historical data anomaly rate and business adaptability. , , The weights are trainable coefficients, corresponding to the proportions of influence of identity credibility, data integrity, and historical data quality on the credible prior weights, respectively, satisfying the constraints. add add It equals 1, and , , The values ​​range from 0 to 1 and can be flexibly adjusted according to the credibility requirements of the risk control scenario. The sigmoid function is calculated using the following formula:

[0033] In the formula, the sigmoid function is used to normalize the weighted sum to the range of 0 to 1, ensuring that features from highly reliable data sources receive higher base weights.

[0034] A multi-head self-attention mechanism is used to calculate the dimension-aligned feature matrix, capture the semantic association and business value between features, and generate semantic association weights. A multi-head self-attention mechanism is employed to perform multi-dimensional attention calculations on the dimension-aligned feature matrix. The feature matrix is ​​mapped to the query space Q, key space K, and value space V, respectively. Scaling dot product attention calculations capture the semantic relationships and risk control business value between different features. The formula for single-head attention is:

[0035] In the formula, √(3 / 2) represents the dimension of the key vector. This is a scaling factor used to prevent the gradient of the softmax function from vanishing due to excessively large inner product values. Multi-head attention concatenates and linearly transforms the outputs of multiple single-head attention processes to generate the final semantic association weights. The value ranges from 0 to 1, enabling the focus on high-value features that are highly relevant to the credit risk control default identification target, while suppressing the weights of irrelevant and noisy features.

[0036] The trusted prior weights and semantic association weights are weighted and fused to generate the final fused attention weights. The feature matrix is ​​then weighted using the fused attention weights.

[0037] The trusted prior weights and semantic association weights are weighted and fused to generate the final fused attention weights. The calculation formula is as follows:

[0038] In the formula, This is the final fusion attention weight, with a value ranging from 0 to 1. The reliability balancing coefficient ranges from 0 to 1 and can be flexibly adjusted according to the compliance requirements of the risk control scenario. In scenarios with strong compliance and strong risk control, the value of λ can be increased to prioritize data reliability. In regular business scenarios, the weights of the two can be balanced to take into account both data reliability and business value. The feature matrix after dimension alignment is weighted by the finally generated fused attention weight.

[0039] In step S3, generating a unified fusion feature vector specifically involves: The dimensionally aligned feature matrix is ​​weighted and summed using fused attention weights to obtain a weighted optimized fused feature matrix. The fused feature matrix is ​​input into the fully connected layer and then transformed nonlinearly to generate a unified fused feature vector of fixed dimensions. The weighting coefficients, model parameters, and core hash values ​​from the calculation process are simultaneously recorded on the blockchain for evidence storage.

[0040] By employing fused attention weights, element-wise weighted summation is performed on the dimension-aligned multi-source heterogeneous feature matrix to obtain a weighted optimized fused feature matrix. This weighted fused feature matrix is ​​then input into a fully connected layer, where a ReLU nonlinear activation function is used to perform a nonlinear transformation, generating a unified fused feature vector of fixed dimensions. This completes end-to-end feature-level fusion of multi-source heterogeneous data. All weight coefficients, model hyperparameters, and intermediate result hash values ​​from the fusion process are synchronously uploaded to the blockchain for notarization, ensuring that the entire fusion process is traceable, reproducible, and verifiable, completely resolving the issues of black-box and unauditable traditional fusion processes.

[0041] S4. Store the fused feature vector on the blockchain, standardize the output of the fused feature vector to the risk control model, and provide full-link traceability and verification capabilities for the fused data.

[0042] In step S4, the on-chain storage of the fused feature vector specifically involves: It adopts a distributed storage architecture that combines on-chain and off-chain collaboration. On-chain storage includes the SHA-256 hash value of the fused feature vector, the hash of all parameters of the fusion process, the corresponding data source DID list, the data batch number, the timestamp and the signature of the operation node. All on-chain storage content is written into the block after consensus among all nodes of the consortium blockchain. The original fusion feature vector, full-process detailed data, and original data backup are stored off-chain. Off-chain data is uniquely anchored by on-chain hash values.

[0043] All on-chain stored content must undergo full-node consensus verification via the PBFT consensus mechanism of the consortium blockchain before being formally written into the block body for on-chain notarization, ensuring that the notarized content is tamper-proof and tamper-proof. The on-chain stored content also simultaneously includes complete link anchoring information such as preprocessing rule hashes and block height, achieving closed-loop hash anchoring of data throughout the entire process. Off-chain storage uses IPFS distributed storage or a financial-grade distributed database to achieve distributed redundant storage of the original data, ensuring high availability and disaster recovery capabilities. All data in off-chain storage is anchored using a unique SHA-256 hash value on the blockchain. Any tampering with the original off-chain data will result in a mismatch between the recalculated hash value and the on-chain notarized hash value, achieving full identification and verification of data tampering.

[0044] In step S4, the standardized output of the fused feature vector is sent to the risk control model, while providing full-link traceability and verification capabilities for the fused data. Two types of standardized interfaces are designed to interface with the risk control model. The standardized interfaces include: a trusted data synchronization interface and a full-link trusted verification interface. The trusted data synchronization interface uses a standardized communication protocol to push the fused feature vector to the risk control model, and simultaneously attaches the corresponding on-chain evidence hash value, block height and trusted verification certificate. The trusted data synchronization interface adopts standardized communication protocols such as RESTful or gRPC, supporting high concurrency and low latency data push, and is compatible with the access requirements of mainstream credit risk control models and anti-fraud models. During the interface call, the on-chain notarized hash value, block height, trusted verification certificate, and end-to-end trusted report summary corresponding to the batch of data are pushed synchronously, providing high-quality data input with end-to-end trusted endorsement for the risk control model. All interface call records, transmitted data hashes, and call node information are stored on the blockchain in real time to ensure the traceability of data transmission.

[0045] The end-to-end trusted verification interface allows the risk control model to query the end-to-end trusted information of the corresponding data by data batch number or hash value. The end-to-end trusted verification interface allows risk control models and auditing institutions to query the end-to-end trusted information of corresponding data in real time through data batch number, data hash value or fusion feature hash value. This includes data source identity authentication information, integrity verification results, preprocessing rules and operation records, fusion model parameters and calculation process, and full-process evidence storage records, meeting the compliance audit requirements and regulatory verification needs of financial risk control decisions.

[0046] Based on the on-chain DID identifier and hash anchoring mechanism, the entire chain of reverse tracing from the fusion result to the original data source is performed, and a tracing report is generated.

[0047] Based on on-chain DID identification and a full-process hash anchoring mechanism, end-to-end reverse traceability is achieved from the fusion result to the original data source. Inputting the hash value of the fused data to be traced, the data batch number, or the fusion feature vector identifier, the system retrieves the corresponding block storage record on the blockchain, obtaining the data source DID list, full-process hash anchoring record, and consensus node signature information for that batch of data. Based on the on-chain hash anchoring record, the system traces back the model parameters and calculation details of the fusion process, the rules and operation records of the preprocessing stage, the integrity verification records of the data transmission stage, and the digital signature information of the original data. Then, based on the data source DID identifier, the system traces the data source's qualification review information, identity registration information, historical compliance records, and credit rating, ultimately generating a complete end-to-end traceability report. This achieves end-to-end traceability, verifiability, and auditability from the fusion result to the fusion process, preprocessing stage, original data, and data source identity, fully meeting the compliance and regulatory requirements of financial risk control scenarios.

[0048] This invention relies on a consortium blockchain to assign a unique distributed digital identity to multi-source heterogeneous data sources. It completes the verification of the legality of the data source identity through asymmetric encryption, and combines SHA-256 single data hash verification and Merkle tree batch data integrity verification to set up a three-level progressive abnormal data interception mechanism. It filters low-trust abnormal data from three dimensions: identity, integrity and quality, to ensure the authenticity and reliability of the integrated data source.

[0049] Differentiated preprocessing schemes are developed for three types of heterogeneous data: structured, semi-structured, and unstructured. Embedding encoding and linear transformation are used to achieve unified alignment between feature dimensions and semantic space, eliminating format barriers and feature differences between heterogeneous data, and laying a unified data foundation for feature fusion of multi-source data.

[0050] The on-chain stored data source credibility score, data integrity score, and historical quality score are transformed into credible prior weights, which are then weighted and fused with the semantic association weights generated by the multi-head self-attention mechanism. This allows high-credibility data to occupy a higher weight in the feature fusion process, effectively suppressing the interference of low-credibility data on the fusion result and improving the reliability of the fused features.

[0051] It adopts a distributed storage architecture that combines on-chain and off-chain collaboration. On-chain storage includes core information such as the fusion result hash, process parameters, and data source identifiers, while off-chain storage includes the original data and detailed content. Based on hash anchoring and distributed digital identity, it enables full-link reverse traceability from the fusion result to the original data source, meeting the compliance audit and regulatory verification needs of scenarios such as financial risk control.

[0052] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A trusted fusion method for multi-source heterogeneous data in a blockchain environment, characterized in that, Includes the following steps: S1. Through blockchain consortium chain nodes, perform identity authentication on multi-source heterogeneous data sources, perform integrity verification on multi-source heterogeneous data that has passed identity authentication, perform hierarchical filtering based on the verification results, and store all operations and results on the blockchain for evidence. S2. Preprocess the multi-source heterogeneous data that was not intercepted after hierarchical filtering to obtain standardized feature data. All preprocessing rules and operation records are hashed and uploaded to the chain, and the corresponding on-chain trust scores are associated with them. S3. Through a feature-level fusion model based on an attention mechanism, the preprocessed feature data is fused, the trust score on the blockchain is integrated into the attention weight calculation, a unified fusion feature vector is generated, and the core parameters of the fusion process are stored on the blockchain. S4. Store the fused feature vector on the blockchain, standardize the output of the fused feature vector to the risk control model, and provide full-link traceability and verification capabilities for the fused data.

2. The method for trusted fusion of multi-source heterogeneous data in a blockchain environment according to claim 1, characterized in that, In step S1, the authentication of multi-source heterogeneous data sources specifically involves: The management node of the blockchain consortium chain conducts qualification review of the data sources applying for access. After the review is approved, a unique distributed digital identity (DID) identifier is generated on the chain for each data source, and an asymmetric public-private key pair is generated accordingly. The public key is bound to the DID identifier and written into the DID registry on the blockchain to complete the on-chain registration, while the private key is kept by the data source itself. When a data source sends a batch of data, it uses its own private key to digitally sign the unique batch number of that batch of data. After the consortium blockchain verification node obtains the data, it completes the signature verification through the public key corresponding to the DID registry on the chain. If the signature verification is successful, the identity of the data source is confirmed to be legitimate.

3. The trusted fusion method for multi-source heterogeneous data in a blockchain environment according to claim 2, characterized in that, In step S1, the integrity verification of the data that has passed identity authentication specifically involves: For a single data entry, a unique hash value is generated using the SHA-256 algorithm and transmitted synchronously with the data. The verification node recalculates the hash value of the received data and compares it with the transmitted hash value to complete the integrity verification of the single data entry. For heterogeneous data in the same batch, a Merkle tree is constructed according to the data form. The generated Merkle root hash is synchronously uploaded to the blockchain for evidence storage. After the batch data transmission is completed, the verification node reconstructs the Merkle tree and compares it with the on-chain Merkle root hash to complete the batch data integrity verification.

4. The trusted fusion method for multi-source heterogeneous data in a blockchain environment according to claim 3, characterized in that, In step S1, the hierarchical filtering based on the verification results specifically involves: Based on the sequence of the verification process, a progressive three-level exception interception mechanism is set up: First, perform data source identity authentication. If the identity verification fails, it is judged as a level 1 anomaly, and all data in that batch is directly intercepted, preventing it from entering the subsequent process. Simultaneously, an on-chain anomaly alarm record is generated and a data source qualification review is triggered. If the identity verification passes, the data integrity verification is performed. If the integrity verification fails, it is judged as a level 2 exception, and the corresponding abnormal data is returned to the data source for retransmission. If the number of retransmissions of a single batch of data from the same data source exceeds the limit, the data source whitelist temporary freezing mechanism is triggered. If the integrity check passes, a data format and quality check is performed. If there are format abnormalities, missing values ​​exceeding the limit, or value range abnormalities, it is judged as a level three abnormality. Abnormal data is marked with an abnormal label on the chain and the abnormality type is recorded synchronously, and it is prohibited from directly entering the feature fusion process.

5. The trusted fusion method for multi-source heterogeneous data in a blockchain environment according to claim 1, characterized in that, In step S2, the preprocessing of multi-source heterogeneous data that was not intercepted after hierarchical filtering specifically involves: For structured data, the following steps are performed sequentially: unified metadata mapping, data cleaning, and feature normalization, to generate a standardized structured feature matrix. For semi-structured data, the schema adaptive parsing engine completes the nested structure recognition and flattening process, maps the parsed core business fields to a unified metadata standard, and generates a standardized semi-structured feature matrix. For unstructured data, high-dimensional features are extracted using a pre-trained model, and then dimensionality reduction is performed to generate feature vectors that match the dimensions of structured data. All preprocessing rules and operation records are synchronized on the blockchain and bound to the on-chain trust score of the corresponding data.

6. The trusted fusion method for multi-source heterogeneous data in a blockchain environment according to claim 1, characterized in that, In step S3, before fusing the preprocessed feature data, the unified representation and dimension alignment of the features of the multi-source heterogeneous data are first completed. The preprocessed structured, semi-structured, and unstructured data are respectively generated into initial feature vectors with unified dimensions through the Embedding layer. Then, all initial feature vectors are mapped to the same feature space through the linear transformation layer to generate feature matrices with consistent dimensions.

7. The trusted fusion method for multi-source heterogeneous data in a blockchain environment according to claim 6, characterized in that, In S3, incorporating the trust score on the blockchain into the attention weight calculation specifically involves: The data source DID trust score, data integrity verification score, and historical data quality score are obtained from the blockchain and normalized by the sigmoid function to generate a trustworthy prior weight. A multi-head self-attention mechanism is used to calculate the dimension-aligned feature matrix, capture the semantic association and business value between features, and generate semantic association weights. The trusted prior weights and semantic association weights are weighted and fused to generate the final fused attention weights. The feature matrix is ​​then weighted using the fused attention weights.

8. The trusted fusion method for multi-source heterogeneous data in a blockchain environment according to claim 7, characterized in that, In step S3, generating a unified fusion feature vector specifically involves: The dimensionally aligned feature matrix is ​​weighted and summed using fused attention weights to obtain a weighted optimized fused feature matrix. The fused feature matrix is ​​input into the fully connected layer and a fixed-dimensional unified fused feature vector is generated through nonlinear transformation. The weighting coefficients, model parameters, and core hash values ​​from the calculation process are simultaneously recorded on the blockchain for evidence storage.

9. A trusted fusion method for multi-source heterogeneous data in a blockchain environment according to claim 1, characterized in that, In step S4, the on-chain storage of the fused feature vector specifically involves: It adopts a distributed storage architecture that combines on-chain and off-chain collaboration. On-chain storage includes the SHA-256 hash value of the fused feature vector, the hash of all parameters of the fusion process, the corresponding data source DID list, the data batch number, the timestamp and the signature of the operation node. All on-chain storage content is written into the block after consensus among all nodes of the consortium blockchain. The original fusion feature vector, full-process detailed data, and original data backup are stored off-chain. Off-chain data is uniquely anchored by on-chain hash values.

10. A trusted fusion method for multi-source heterogeneous data in a blockchain environment according to claim 9, characterized in that, In step S4, the standardized output of the fused feature vector is sent to the risk control model, while providing full-link traceability and verification capabilities for the fused data. Two types of standardized interfaces are designed to interface with the risk control model. The standardized interfaces include: a trusted data synchronization interface and a full-link trusted verification interface. The trusted data synchronization interface uses a standardized communication protocol to push the fused feature vector to the risk control model, and simultaneously attaches the corresponding on-chain evidence hash value, block height and trusted verification certificate. The end-to-end trusted verification interface allows the risk control model to query the end-to-end trusted information of the corresponding data by data batch number or hash value. Based on the on-chain DID identifier and hash anchoring mechanism, the entire chain of reverse tracing from the fusion result to the original data source is performed, and a tracing report is generated.

Images