Multi-source data integration AI knowledge base construction method and system

Abstract

The invention provides an AI knowledge base construction method and system for multi-source data integration, and belongs to the technical field of knowledge base construction. The method comprises the following steps: adaptively pulling unstructured data at an API-free site through a reversible crawler of a heterogeneous data access framework; the mode drift detector senses field change in real time and updates a protocol in a hot plug mode within 500ms, so that zero-stop synchronization is realized; carrying out weighted resolution on conflicts according to source credibility index attenuation, and generating version stamps with digital watermarks and consanguinity URI (Uniform Resource Identifier) for records; uniformly mapping to a shared semantic space through cross-modal comparative learning, and constructing an incremental hypergraph knowledge graph which can be evolved by probability weight; and extracting an optimal sub-graph by using generative adversarial reinforcement learning, writing the optimal sub-graph into a time sequence knowledge warehouse by using a zero knowledge evidence chain and carrying out block chain solidification, and outputting a verifiable certificate. By the adoption of the AI knowledge base construction method and system for multi-source data integration, end-to-end automatic, high-credibility and traceable large-scale knowledge base construction is achieved.

Description

Technical Field

[0001] This invention relates to the field of knowledge base construction technology, and in particular to a method and system for constructing an AI knowledge base by integrating multi-source data. Background Technology

[0002] With the rapid development of Web 3.0 and the big data industry, enterprises, governments, and research institutions are eager to integrate multi-source, heterogeneous data scattered across the internet, the Internet of Things, and internal business systems into high-quality, reasonable, and traceable AI knowledge bases to support high-value applications such as intelligent decision-making, risk warning, and personalized services. However, existing technologies generally suffer from the following insurmountable shortcomings during implementation:

[0003] (1) Data source access is limited: Traditional solutions rely on public APIs or writing crawler scripts once. Once the target site is redesigned, a new anti-crawling strategy is implemented, or the IP is blocked, the data flow will be interrupted. Manual maintenance of scripts is costly and cannot form an automated link that can be replayed and reused.

[0004] (2) Protocol drift causes downtime: When the upstream page structure changes due to the addition or deletion of fields or changes in type, the existing system needs to be shut down to release a new version of the parsing rules, resulting in a data vacuum; even if a simple regular expression or XPath template is used, it is impossible to complete the adaptive switching in minutes or even seconds.

[0005] (3) Decay in the credibility of conflict resolution: Most existing conflict resolution algorithms use static weights or simple majority voting, ignoring the accuracy drift of data sources over time. Once the quality of a data source declines, erroneous information is still cited for a long time, leading to a continuous decrease in the overall confidence of the knowledge base.

[0006] (4) Multimodal semantic fragmentation: Multimodal data such as text, images, and time-series signals are usually encoded independently, lacking a unified semantic space mapping, making it difficult to achieve cross-modal retrieval and reasoning; traditional methods do not introduce a time decay mechanism for old samples, further amplifying the risk of concept drift.

[0007] (5) Static solidification of knowledge graphs: Most existing graphs are built in batches with fixed super-edge weights, which cannot be split, merged or weighted in milliseconds as business evolves; incremental updates often require full graph recalculation, resulting in huge computation and storage overhead.

[0008] (6) The dilemma of subgraph extraction efficiency and privacy: In real-time query scenarios, traditional subgraph extraction uses heuristic rules or offline calculations, which are slow and difficult to balance privacy; when federated or multi-party collaboration, there is a lack of dynamic optimization mechanism that balances differential privacy and information entropy gain.

[0009] (7) The contradiction between the credibility of the evidence chain and the cost: Existing traceability schemes either rely on centralized databases, which pose a risk of single-point tampering; or they directly put all the data on the chain, resulting in extremely high storage and gas costs, making it difficult to popularize in large-scale knowledge bases.

[0010] The aforementioned shortcomings are intertwined, making the construction of a large-scale, highly reliable, low-cost, and traceable AI knowledge base still a pain point for the industry. Summary of the Invention

[0011] The purpose of this invention is to provide a method and system for constructing an AI knowledge base by integrating multi-source data, so as to realize the construction of a large-scale knowledge base that is automated, highly reliable, and traceable from end to end.

[0012] To achieve the above objectives, this invention provides a method for constructing an AI knowledge base by integrating multi-source data, comprising the following steps:

[0013] Step S1: Through the reversible crawler orchestrator in the heterogeneous data access framework, automatically generate a replayable crawler script on the target site without a public API, pull unstructured documents, and output the original unstructured data set.

[0014] Step S2: Using the original unstructured data set as input, the pattern drift detector monitors the addition or deletion of fields and triggers the protocol adaptive plugin to perform hot-plug updates, outputting a continuously synchronized zero-downtime data stream;

[0015] Step S3: Taking the continuously synchronized zero-downtime data stream as input, perform conflict resolution based on the decay of the source credibility index, generate a version stamp with digital watermark and lineage URI for each record, and output a set of resolved records with version stamp;

[0016] Step S4: Taking text, image, and time-series signal modalities from the resolved record set with version stamps as input, the unified contrastive learning loss function is used to map them to the shared semantic space, and the output is a multi-dimensional continuous vector with time decay awareness.

[0017] Step S5: Using multidimensional continuous vectors as the node set and hyperedges with evolving probability weights that drift over time as connections, construct an incremental hypergraph knowledge graph that supports hyperedge splitting and merging, and output the hypergraph knowledge graph.

[0018] Step S6: Using the hypergraph knowledge graph as the environment, train the optimal subgraph extraction strategy using the generative adversarial reinforcement learning framework. The generator simulates the query distribution and outputs candidate subgraphs, while the discriminator gives a reward based on information entropy gain and delay penalty.

[0019] Step S7: Write the optimal subgraph instance extracted by the optimal subgraph extraction strategy into the time-series knowledge warehouse, generate a zero-knowledge evidence chain that conforms to the W3CVerifiable Credentials data model, and solidify its immutability through the blockchain timestamp service to output verifiable credentials.

[0020] Preferably, step S1 specifically includes:

[0021] Using the target site's page as input, the script is recorded and played back in a sandbox environment, and the updated CSS / XPath selectors are automatically generated when the target site is redesigned.

[0022] Taking CSS / XPath selectors and the target site's anti-crawling verification mechanism as input, it adopts a dual-channel parallel approach of headless browser and traffic replay, maintains crawling through behavioral fingerprint perturbation, and outputs an uninterrupted raw data pull stream.

[0023] The script instructions from the fetching process are used as input, and the generated script is stored in the form of reversible bytecode.

[0024] Using historical anti-scraping rule change data as input, the model is trained through a reinforcement learning-based anti-scraping strategy prediction module to generate behavioral perturbation schemes to cope with new anti-scraping mechanisms in advance.

[0025] Taking the target site's IP reputation score and the proxy node status as input, the system dynamically switches proxy nodes through intelligent scheduling of a distributed proxy pool, and finally outputs a set of raw unstructured data.

[0026] Preferably, step S2 specifically includes:

[0027] Using the field features of the original unstructured dataset as input, a bivariate statistic of field occurrence frequency and field type entropy is calculated using a sliding window approach. A dynamic threshold model is constructed based on the historical drift features of the data source, and a drift event is triggered when the joint deviation exceeds the corresponding threshold.

[0028] Using drift events and original data protocol features as input, the protocol adaptive plugin generates and registers a new protocol template within 500ms after detecting a drift event by online matching and Bayesian updating based on a protocol template library.

[0029] Using the output data of the old and new protocol templates as input, after verifying the consistency difference between the two outputs in parallel through the shadow channel to be ≤2%, switch to production traffic;

[0030] Taking the warning signal that the field statistics are close to the threshold as input, lightweight preprocessing is triggered in advance, and finally the continuous synchronous zero-downtime data stream is output.

[0031] Preferably, in step S3, conflict resolution based on the decay of the source credibility index, taking a continuously synchronized zero-downtime data stream as input, performs the following processing and outputs a set of resolved records with version stamps, specifically including:

[0032] Using the historical records of each data source in the data stream as input, an accurate prior based on beta-Bernoulli dynamic updates is established for each data source, where the prior parameters decay with the real-time feedback error in an exponentially weighted moving average manner.

[0033] The attenuated accuracy is taken as input and mapped to source credibility weights, which are used for weighted voting to resolve field-level conflicts. The source credibility weights are calculated as follows:

[0034] ω source (t)=ω source (t-1)×λ Δt ;

[0035] Where, ω source (t) represents the source credibility weight at time t, ω source (t-1) represents the source credibility weight at time t-1, Δt represents the time interval, and λ represents the decay coefficient;

[0036] Using the conflict resolution results and data source reliability as inputs, ω source (t) is written as the version stamp of the corresponding record as metadata, and a version stamp with digital watermark and lineage URI is generated for each record.

[0037] Using the data on changes in the accuracy of the data source as input, when the accuracy drops by more than 10%, the attenuation coefficient is temporarily adjusted to 0.8;

[0038] By combining the domain prior weights of the data source, the final output is a set of resolved records with version stamps.

[0039] Preferably, in step S3, the generation and embedding of the digital watermark takes the field data of the output de-watermarked record set as input, performs the following processing, and outputs records with digital watermarks, specifically including:

[0040] Using the text content of the record field as input, a hidden watermarking technique based on zero-width Unicode characters is employed to embed a 128-bit hash digest into the blank space of the field value.

[0041] Using watermarked field data as input, during the data leakage tracing phase, zero-width character sequences are extracted and decoded using regular expressions to achieve lossless tracing of the leakage source;

[0042] Using the field sensitivity rating as input, the watermark embedding strength is dynamically adjusted to maintain a decoding accuracy of ≥99.9%, and the final output is a record with a digital watermark.

[0043] Preferably, step S4 specifically includes:

[0044] Using three modal data as input, cross-modal anchor-positive sample-negative sample triples are constructed, and text, images, and time-series signals associated with the same business entity are marked as positive sample triples;

[0045] Perform cross-modal hard negative sample mining within the same training batch;

[0046] Using sample features and time decay parameters as input, a loss function L is used to make the gradient contribution of old samples decrease exponentially with time. The loss function L is as follows:

[0047] L = L txt-img +L txt-ts +L img-ts +αL temporal-decay ;

[0048] Among them, L txt-img L represents the contrastive learning loss between text and image modalities. txt-ts L represents the contrastive learning loss between text and temporal signal modes. img-ts L represents the contrastive learning loss between the image and the temporal signal modes, α represents the time decay regularization coefficient, and L represents the time decay regularization coefficient. temporal-decay This represents the time decay loss term;

[0049] After training, the cosine distance between synonymous concepts of different modalities in the shared semantic space is ≤0.05, and the cosine distance between heteronymous concepts is ≥0.35.

[0050] Emerging modalities such as audio and 3D models are accessed through a standardized feature extraction protocol of the modal adaptation interface layer. The weight of modalities with high noise is reduced to 0.8 times the original weight, while the weight of stable modalities is maintained.

[0051] Preferably, step S5 specifically includes:

[0052] The dynamic changes of the superedge weights follow a stochastic process driven by drift rate and volatility, where the drift rate is set by the domain prior and the volatility is updated in real time through online maximum likelihood estimation;

[0053] A merge operation is triggered when the weight of the superedge drops to the first threshold of 0.05, and a split operation is triggered when the weight of the superedge rises to the second threshold of 0.95 and the number of associated nodes exceeds 10.

[0054] A dynamic adjacency tensor storage structure accelerated by a graphics processor is used to achieve millisecond-level sparse updates based on a coordinate list format.

[0055] Record the triggering factors for weight changes in the hyperedge attribute;

[0056] Perform automatic aggregation on redundant superedges with a weight less than 0.02.

[0057] Preferably, in step S6, the adversarial reinforcement learning framework performs the following operations:

[0058] The generator progressively selects nodes or superedges to add to the candidate subgraph using a pointer network structure.

[0059] The discriminator reward function is defined as: the information entropy gain value minus 0.01 times the response delay value;

[0060] In a federated training scenario, inject parameters ε=1 and δ=10 that satisfy privacy parameters. -5 Differential privacy noise is introduced, and training is terminated before the privacy budget is exhausted.

[0061] Preferably, in step S7, the zero-knowledge evidence chain performs the following operations:

[0062] By embedding STARK zero-knowledge proofs in the main fields of the evidence chain, we can achieve inference path verification without exposing the original nodes;

[0063] A two-layer Rollup architecture is used to anchor the digest hash to the blockchain, while the complete chain of evidence is stored in the InterPlanetary File System.

[0064] The cost of a single anchor is kept below $0.001, and cross-chain bridging protocol supports mutual recognition between Ethereum and Polygon.

[0065] Evidence chain updates are implemented based on on-chain verifiable delay functions, with a 10-minute delay period set to defend against rollback attacks.

[0066] This invention also provides an AI knowledge base construction system that integrates multi-source data, comprising:

[0067] A heterogeneous data access framework with a built-in reversible crawler orchestrator, used to automatically generate replayable crawler scripts and output raw unstructured data sets on target sites without public APIs.

[0068] A pattern drift detector and protocol adaptive plugin are used to monitor field additions and deletions and perform hot-swap updates, outputting a continuously synchronized zero-downtime data stream;

[0069] The conflict resolution engine generates a version stamp with a digital watermark and lineage URI for each record based on the decay of the source credibility index, and outputs a set of resolved records with version stamps.

[0070] A multimodal contrastive learning encoder is used to map text, images, and time-series signals to a shared semantic space, outputting a multidimensional continuous vector with time decay awareness.

[0071] An incremental hypergraph knowledge graph builder uses multidimensional continuous vectors as nodes and probabilistically weighted, evolving hyperedges as connections. It supports hyperedge splitting and merging and outputs a hypergraph knowledge graph.

[0072] Generative adversarial reinforcement learning subgraph extractor is used to train the optimal subgraph extraction policy in the hypergraph knowledge graph and output candidate subgraphs.

[0073] The temporal knowledge repository and blockchain solidification module are used to write the optimal subgraph instance into the temporal knowledge repository, generate a zero-knowledge evidence chain that conforms to the W3C Verifiable Credentials data model, and solidify its immutability through blockchain timestamp service, outputting verifiable credentials.

[0074] Therefore, the present invention employs the above-described method and system for constructing an AI knowledge base based on multi-source data integration, and the beneficial technical effects are as follows:

[0075] (1) Full-domain automation and zero downtime: Reversible crawlers, adaptive protocols and distributed agents work together to completely get rid of dependence on public APIs and manual scripts. Any site redesign or anti-crawling upgrade can be completed in milliseconds to ensure uninterrupted data flow 24 / 7.

[0076] (2) Continuous evolution of credibility and self-healing of conflicts: Through the exponential decay credibility model, digital watermarking and lineage URI, dynamic weight voting and version tracing are performed on each record. Erroneous data is automatically downgraded as the source quality decreases, and the overall confidence of the knowledge base is maintained in the long term.

[0077] (3) Dynamic graph and low-cost trusted traceability: Incremental supergraph supports millisecond-level splitting, merging and weight evolution, and with generative adversarial reinforcement learning, it achieves privacy-preserving subgraph extraction; Zero-knowledge evidence chain anchors the blockchain with a two-layer rollup, taking into account immutability and extremely low on-chain cost, and meets the needs of large-scale trusted knowledge services in high compliance scenarios. Attached Figure Description

[0078] Figure 1 This is a flowchart illustrating a method for constructing an AI knowledge base through multi-source data integration, as described in this invention.

[0079] Figure 2 Here is a flowchart of step S1;

[0080] Figure 3 This is a flowchart of step S2. Detailed Implementation

[0081] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0082] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0083] Example 1

[0084] like Figures 1-3 A method for constructing an AI knowledge base by integrating multi-source data includes the following steps:

[0085] Step S1: Through the reversible crawler orchestrator in the heterogeneous data access framework, automatically generate a replayable crawler script on the target site without a public API, pull unstructured documents, and output the original unstructured data set.

[0086] 1. Script recording and playback mechanism.

[0087] The system performs event capture and state synchronization on the target page in an isolated sandbox, recording the user interaction sequence and the stable selectors of the corresponding DOM (Document Object Model) elements. When the page structure changes, it automatically recalculates and generates new CSS / XPath paths through a differential DOM comparison algorithm to ensure selector availability. At the same time, it stores the old and new versions in reversible bytecode form, supporting rollback and version tracking.

[0088] 2. Dual-channel parallel acquisition.

[0089] Headless Browser Channel: Based on a containerized headless browser instance, it simulates real user behavior characteristics by dynamically adjusting User-Agent, screen resolution, and mouse trajectory parameters, thereby reducing the risk of behavior consistency detection.

[0090] Traffic replay channel: Based on the complete TLS (Transport Layer Security) session and header sequence of historical successful requests, the request is directly replayed through the session reuse mechanism to reduce interaction latency; if an abnormal TLS fingerprint or response is detected, it automatically switches to the Headless channel for compensation collection.

[0091] 3. Reversible bytecode storage.

[0092] The script instruction sequence generated during the recording process is encoded in intermediate language, including operation type, element path, timestamp and page hash check value. It is stored in an object storage system using compression and encryption mechanisms, and supports version retrieval and difference comparison.

[0093] 4. Anti-crawling strategy prediction.

[0094] A training set is constructed based on historical anti-scraping rule change logs. A two-layer LSTM and attention mechanism model is used to model the request feature sequence, predict possible anti-scraping upgrade strategies in the near future, and generate corresponding behavior perturbation schemes and header mutation rules in advance.

[0095] Two-layer LSTM:

[0096] First LSTM layer: Set the hidden layer size to 256, the input dimension is determined according to the dimension of the requested feature sequence, and the output dimension is 256.

[0097] The second LSTM layer is also set to a hidden layer size of 256, an input dimension of 256 (receiving the output of the first LSTM layer), and an output dimension of 256.

[0098] Each LSTM layer uses dropout technology with a dropout rate of 0.2 to prevent overfitting.

[0099] Attention mechanism layer:

[0100] Scaled Dot-Product Attention is used to calculate the attention score between the query, key, and value.

[0101] The attention mechanism has an input dimension of 256 (receiving the output of the second LSTM layer) and an output dimension of 256.

[0102] Training parameters: The optimizer used is the Adam optimizer, with an initial learning rate of 0.001. During training, a learning rate decay strategy is adopted, with the learning rate multiplied by 0.9 every 10 epochs. The batch size is set to 64, and the number of iterations (epochs) is set to 50. The loss function used is the cross-entropy loss function, which is used to measure the difference between the predicted result and the actual result.

[0103] 5. Distributed proxy pool scheduling.

[0104] The health of proxy nodes is determined by a comprehensive score of latency, blocking rate, and IP reputation. The scheduler periodically reassesses node availability, and nodes below the threshold are automatically taken offline. At the same time, based on the geographical location and IP reputation of the target site, high-reputation nodes in the same region are prioritized to ensure the stability and concealment of data collection.

[0105] 6. Output data specifications.

[0106] The collected results retain the original format and complete metadata, including page URL, collection timestamp, page hash, and proxy node identifier; the SimHash algorithm is used to deduplicate the content, ensuring the uniqueness and integrity of the output data.

[0107] Step S2: Using the original unstructured data set as input, the pattern drift detector monitors the addition or deletion of fields and triggers the protocol adaptive plugin to perform hot-plug updates, outputting a continuously synchronized zero-downtime data stream.

[0108] 1. Field drift detection model.

[0109] The system continuously calculates the frequency and type entropy of each field in the original dataset using a sliding window approach, constructing a bivariate joint distribution. A dynamic threshold model is trained based on historical drift features; a drift event is triggered when the joint deviation exceeds the threshold. Drift events are labeled as "structural" or "semantic" for subsequent protocol matching strategy selection.

[0110] 2. Protocol Adaptive Plugin.

[0111] The plugin has a built-in protocol template library. Templates are described in the form of quadruples of field path, type constraint, nesting level, and value range. After detecting a drift event, the plugin matches the most similar template in the template library through a Bayesian update mechanism and automatically generates completion rules for missing fields. The new template runs in parallel in the shadow channel and performs consistency checks with the old template output; the hot-swap switch is completed when the difference rate is less than 2%, and the switchover process is imperceptible to production traffic.

[0112] 3. Shadow channel verification.

[0113] The shadow channel shares the input data stream with the main channel but outputs independently. The system performs field-level hash comparison and statistical distribution checks on the two outputs to ensure semantic consistency. If the difference rate exceeds a threshold, a rollback mechanism is triggered, the template is marked as "to be corrected," and an alarm event is generated.

[0114] 4. Early warning and pre-processing.

[0115] When a field statistic approaches a threshold but no drift event is triggered, the system initiates a lightweight preprocessing procedure: generating placeholders for potentially missing fields and performing type casting on fields with anomalous types to ensure the continuity of subsequent steps. The warning information is also written to the metadata log for manual auditing or model retraining.

[0116] 5. Zero downtime guarantee.

[0117] The protocol switching employs a dual-buffer architecture: the old protocol buffer continuously serves the current request, while the new protocol buffer atomically switches read and write pointers after warming up, with switching latency controlled within 500ms. The entire process does not require restarting the acquisition task, ensuring continuous data output.

[0118] Step S3: Using a continuously synchronized zero-downtime data stream as input, perform conflict resolution based on source credibility index decay, generate a version stamp with digital watermark and lineage URI for each record, and output a set of resolved records with version stamp.

[0119] 1. Source credibility index decay.

[0120] For each continuously synchronized data stream, the system establishes a beta-Bernoulli dynamic prior for each independent data source. The prior parameters are updated using an exponentially weighted moving average based on real-time feedback errors, forming a confidence weight that decreases over time. The weights are used for weighted voting in field-level conflict resolution. The weight update formula strictly follows the exponential decay law, and time intervals and decay coefficients are introduced to quantify the rate of confidence decline.

[0121] The source credibility weight is calculated as follows:

[0122] ω source (t)=ω source (t-1)×λ Δt ;

[0123] Where, ω source (t) represents the source credibility weight at time t, ω source (t-1) represents the source credibility weight at time t-1, Δt represents the time interval, and λ represents the decay coefficient;

[0124] 2. Conflict resolution mechanism.

[0125] In field-level conflict scenarios, the system extracts multi-value records of the same entity from different data sources and performs weighted majority voting based on the current credibility weight of each data source. If the weight differences are significant, the highest weight value is directly adopted. If the weights are close, a priori weight from the domain is introduced for secondary weighting, and finally a unique resolution result is output.

[0126] 3. Version stamp and lineage URI.

[0127] Each resolved record is appended with a version stamp, which includes: data source identifier, trust weight, generation timestamp, digital watermark hash, and lineage URI. The lineage URI uses a resolvable path structure, pointing to all upstream records and data sources that the record depends on, supporting end-to-end traceability.

[0128] 4. Digital watermark embedding.

[0129] To embed a hidden watermark in the text field: a zero-width Unicode character set is selected, and the 128-bit hash digest is mapped bit by bit to a zero-width sequence, which is then inserted into the blank spaces of the field. The watermark strength is dynamically adjusted according to the field's sensitivity level to ensure that the embedded text is visually indistinguishable and that the decoding accuracy is higher than 99.9%.

[0130] 5. Dynamic adjustment of attenuation coefficient.

[0131] When the real-time accuracy of any data source drops by more than 10%, the system automatically adjusts the attenuation coefficient of that source to 0.8 temporarily to accelerate its weight attenuation. If the accuracy is still below the threshold after three consecutive evaluations, a manual review is triggered, and its effective weight is reduced to 50% of its original value during the conflict resolution phase.

[0132] Step S4: Taking text, image, and time-series signal modalities from the resolved record set with version stamps as input, the data is mapped to the shared semantic space through a unified contrastive learning loss function, and the output is a multi-dimensional continuous vector with time decay awareness.

[0133] 1. Construction of cross-modal samples.

[0134] Taking a set of resolved records with version stamps as input, the system extracts three modalities of data: text, image, and time-series signal. Based on the business entity identifier, it constructs an anchor-positive sample-negative sample triplet:

[0135] Anchor points and positive samples represent the same entity in different modalities.

[0136] Negative samples are selected through a cross-modal hard negative sample mining strategy, which selects records that are semantically similar but have different entities, to ensure that the training gradient is effective.

[0137] 2. Unify the contrastive learning loss function.

[0138] The loss function consists of three parts: intermodal contrast losses for text-image, text-time series, and image-time series, coupled with a time decay regularization term. The regularization term exponentially reduces the gradient weights of older samples, making the model focus more on recent data and suppressing concept drift. After training, the cosine distance between synonymous concepts in the shared semantic space is no higher than 0.05, and the cosine distance between heteronymous concepts is no lower than 0.35.

[0139] The loss function L is as follows:

[0140] L = L txt-img +L txt-ts +L img-ts +αL temporal-decay ;

[0141] Among them, L txt-img L represents the contrastive learning loss between text and image modalities. txt-ts L represents the contrastive learning loss between text and temporal signal modes. img-ts L represents the contrastive learning loss between the image and the temporal signal modes, α represents the time decay regularization coefficient, and L represents the time decay regularization coefficient. temporal-decay This represents the time decay loss term.

[0142] in:

[0143]

[0144] L temporal-decay =λ(t)||F(t)-F(t-Δt)||2;

[0145] Where, N pos N represents the number of positive samples. neg S represents the negative sample logarithm, τ represents the temperature parameter, and S pos,i S represents the cosine similarity between the i-th pair of positive samples (text features, image features). neg,j S represents the cosine similarity of the j-th negative sample (text features, image features). pos,i1 S represents the cosine similarity of the i-th pair of positive samples (textual features, temporal features). neg,j1 S represents the cosine similarity of the j-th negative sample (textual features, temporal features). pos,i2 S represents the cosine similarity of the i-th pair of positive samples (image features, temporal features). neg,j2 Let F(t) represent the cosine similarity of the j-th negative sample (image features, temporal features), F(t) represent the feature vector at time t, and F(t-Δt) represent the feature vector at the previous time t-Δt.

[0146] 3. Modal adaptation and weight adjustment.

[0147] The system accesses new modalities (such as audio, 3D models, etc.) through a standardized feature extraction protocol. For modalities with high noise levels or unstable data distribution, the training weights are automatically reduced to 0.8 times their original values; for modalities with stable data quality, the original weights are kept unchanged to maintain overall representation consistency.

[0148] 4. Time decay sensing mechanism.

[0149] Each sample is appended with timestamp metadata before entering the training batch; the time decay regularization coefficient is dynamically calculated based on the interval between the sample generation time and the current training time to ensure that the contribution of old data to the total loss decays exponentially over time. The decay coefficient is determined by the domain-preset half-life and can be adjusted online.

[0150] 5. Shared semantic space verification.

[0151] A shared semantic space is a unified vector space that can be mapped to all data samples from different modalities (such as text, images, and time-series signals) after training through multimodal contrastive learning. In this space, samples of the same entity from different modalities have high similarity in their vector representations, thus supporting cross-modal retrieval and matching operations.

[0152] After training, the system performs cross-modal retrieval validation: using the representation vector of any modality as the query, a nearest neighbor search is performed in the shared space, requiring that the hit rate of samples of the same entity from different modalities in the returned results be no less than 95%, ensuring the effectiveness of multimodal alignment.

[0153] Step S5: Using multidimensional continuous vectors as the node set and hyperedges with evolving probability weights that drift over time as connections, construct an incremental hypergraph knowledge graph that supports hyperedge splitting and merging, and output the hypergraph knowledge graph.

[0154] 1. Modeling of hyperedge-weighted stochastic processes.

[0155] Using multidimensional continuous vectors as the node set, the system assigns an initial weight to each hyperedge, and its dynamic evolution follows a drift-fluctuation coupled stochastic process:

[0156] The drift rate is set by domain priors, reflecting the long-term expectations of business rules regarding the strength of association;

[0157] Volatility is updated in real time via online maximum likelihood estimation, reflecting the uncertainty of recent observations.

[0158] The weight update frequency is synchronized with the incremental refresh cycle of the knowledge graph to ensure statistical consistency.

[0159] 2. Triggering conditions for splitting and merging.

[0160] The super-edge state machine defines two levels of thresholds:

[0161] When the weight drops to the first threshold of 0.05 and persists for two observation periods, a merging operation is triggered, and its associated node is merged into the adjacent superedge with the highest weight.

[0162] When the weight rises to the second threshold of 0.95 and the number of associated nodes exceeds 10, a split operation is triggered. The nodes are divided using k-medoids based on the cosine distance between them, generating two independent hyperedges. The weights are inherited proportionally and then renormalized.

[0163] In this context, proportional weight inheritance refers to determining the initial weights of the two newly generated independent hyperedges based on the weight contribution of each node in the original hyperedge. Specifically, before the split, the weights of each node in the hyperedge are statistically analyzed, and the nodes are divided into two groups using k-medoids partitioning. Each group corresponds to one of the two newly generated hyperedges. For each group, the total weight of that group in the original hyperedge is calculated, and its proportion within the original hyperedge's total weight is determined. This proportion serves as the basis for inheriting the weights of the new hyperedges. For example, suppose the original hyperedge has a total weight value, the first group of nodes has a total weight value, and the second group has a total weight value equal to the original hyperedge's total weight value. Then, the initial weight of the first newly generated hyperedge will be determined according to the proportion of the first group's total weight within the original hyperedge's total weight, and similarly, the initial weight of the second newly generated hyperedge will be determined according to the proportion of the second group's total weight. Subsequently, the weights of the two newly generated hyperedges are renormalized so that the sum of their weights equals the total weight of the hyperedges before the split, thus ensuring that the weight system of the entire hypergraph knowledge graph remains relatively stable and reasonable after the split operation.

[0164] 3. Incremental storage structure.

[0165] The GPU-accelerated dynamic adjacency tensor maintains a sparse structure internally in coordinate list (COO) format, supporting millisecond-level insertion, deletion, and weight updates. Weight change events are written to the hyperedge attributes in append log form, retaining the triggering reason, timestamp, and operation type for easy backtracking and auditing.

[0166] 4. Redundant overedge recycling.

[0167] The system periodically scans hyperedges with a weight lower than 0.02 and fewer than 3 nodes, marking them as redundant. Redundant hyperedges are aggregated to generate a representative hyperedge, with the weights taken as a weighted average. The node lists are merged and deduplicated to ensure that the graph size is controlled.

[0168] 5. Consistency guarantee.

[0169] All weight change operations are implemented through optimistic concurrency control: the current weight version number is obtained before the update, and the version number is checked to ensure that it has not changed when writing. If there is a conflict, the operation is retried. This mechanism ensures that the graph state is consistent under concurrent updates and does not require a global lock.

[0170] Step S6: Using the hypergraph knowledge graph as the environment, train the optimal subgraph extraction strategy using a generative adversarial reinforcement learning framework. The generator simulates the query distribution and outputs candidate subgraphs, while the discriminator gives a reward based on information entropy gain and delay penalty.

[0171] 1. Generate an adversarial reinforcement learning framework.

[0172] Using an incremental hypergraph knowledge graph as the environment, the system constructs a two-layer architecture of generator-discriminator:

[0173] The generator employs a pointer network to progressively select nodes or hyperedges to construct candidate subgraphs. The pointer network makes decisions by calculating the probability of selecting each node or hyperedge in the current state. Specifically, for each possible node or hyperedge, the pointer network calculates a probability value based on factors such as its relevance to the currently constructed subgraph and its importance in the hypergraph knowledge graph. It then selects the next node or hyperedge to add to the candidate subgraph according to this probability value. For example, at each step of constructing the candidate subgraph, the generator first evaluates all unselected nodes and hyperedges, calculates their selection probabilities, and then samples according to this probability distribution, selecting one node or hyperedge to add to the subgraph. This process is repeated until a suitable candidate subgraph is constructed.

[0174] The discriminator calculates a reward based on a weighted sum of information entropy gain and response latency, guiding the generator's policy updates. Information entropy gain is determined by comparing the difference in information entropy between the candidate subgraph and the original hypergraph knowledge graph. Specifically, the information entropy of the candidate subgraph and the original hypergraph knowledge graph are calculated separately, and then the information entropy of the candidate subgraph is subtracted from the information entropy of the original hypergraph knowledge graph; the difference is the information entropy gain. Response latency refers to the time spent processing the candidate subgraph.

[0175] The training objective is to maximize the subgraph information entropy gain within a limited latency budget.

[0176] 2. Subgraph extraction strategy optimization.

[0177] At each step, the generator calculates the pointer distribution based on the current subgraph state vector and the environment context vector, and selects the next node or hyperedge through sampling; the discriminator evaluates the complete subgraph, and the reward function R is defined as...

[0178] R = Information entropy gain - 0.01 response delay;

[0179] The information entropy gain is measured by the entropy difference in the distribution of subgraph node attributes, and the response latency is recorded using a millisecond-level timer.

[0180] 3. Federal training privacy protection.

[0181] In a federated scenario, the system injects the gradients of the generator and discriminator that satisfy (ε = 1 and δ = 10). -5 Differential privacy noise; the privacy budget is tracked cumulatively, and training terminates when the budget is exhausted to prevent privacy leaks. The noise level is dynamically adjusted based on gradient sensitivity to ensure model availability.

[0182] 4. Strategy convergence and evaluation.

[0183] Training employs asynchronous advantage Actor-Critic updates, generating a fixed number of candidate subgraphs in each round. After evaluation by the discriminator, the subgraph with the highest reward is retained. Policy convergence is determined based on the reward increase being less than 1% for three consecutive rounds. At this point, the model parameters are frozen and the optimal subgraph extraction policy is output.

[0184] Asynchronous Advantage Actor-Critic Network Structure:

[0185] Actor Networks: These networks take the current state of the incremental hypergraph knowledge graph as input and output a probability distribution for selecting nodes or hyperedges to construct candidate subgraphs. For example, at each step of generating candidate subgraphs, the Actor Network calculates the probability of selecting each node or hyperedge based on the connection relationships and weights of nodes and hyperedges in the current hypergraph. Then, it samples according to this probability distribution to determine the next node or hyperedge to be added to the candidate subgraph.

[0186] The Critic network, similarly, takes the current state of the hypergraph knowledge graph as input and outputs a value estimate for that state. This value estimate is used to evaluate the merits of the candidate subgraph construction strategies generated by the Actor network, providing guidance for strategy updates.

[0187] Asynchronous update method: The system simultaneously launches multiple parallel training threads, each exploring and learning in an independent hypergraph knowledge graph environment instance. Each thread generates candidate subgraphs based on its own Actor network and submits them to the discriminator for evaluation, receiving corresponding rewards. Then, based on its experience in this process, each thread calculates the gradients of the Actor and Critic networks and asynchronously passes these gradients to the global model parameters for updating. This method can fully utilize the computational resources of multi-threading, accelerating training speed, while the exploration of different environments by different threads also increases sample diversity and improves the model's generalization ability.

[0188] Advantage function calculation: The advantage function measures the strength of the current policy relative to the average policy. In this system, the formula is: Advantage value = Discriminator reward + Discount factor × Next state value estimate - Current state value estimate. The discount factor is used to weigh the importance of the current reward and future rewards and can be set according to actual needs. By calculating the advantage value, the merits of the candidate subgraph construction policies generated by the Actor network can be evaluated more accurately, thus guiding policy updates more effectively.

[0189] Training process:

[0190] Initialize the global Actor and Critic network parameters.

[0191] Start multiple parallel training threads.

[0192] Each thread generates a fixed number of candidate subgraphs in its own hypergraph knowledge graph environment based on the current Actor network, and submits these candidate subgraphs to the discriminator for evaluation, receiving a reward for each candidate subgraph.

[0193] Each thread retains the candidate subgraph with the highest reward from the discriminator's evaluation results, calculates the gradients of the Actor and Critic networks based on this, and then asynchronously passes the gradients to the global model for parameter updates.

[0194] Repeat the above process until the policy converges. The policy convergence criterion is that the reward increase is less than 1% over three consecutive training rounds.

[0195] When the strategy converges, the model parameters are frozen, and the optimal subgraph extraction strategy is output. This strategy can be used to efficiently extract subgraphs that meet the requirements in incremental hypergraph knowledge graphs.

[0196] 5. Subgraph instantiation and caching.

[0197] The subgraph instances extracted by the optimal strategy are written into the time-series knowledge warehouse after lossless serialization, and are also cached in a high-speed memory pool in key-value form. The cache key is composed of the query vector hash and the strategy version number, ensuring that the same query can be directly reused in the future, reducing latency.

[0198] Step S7: Write the optimal subgraph instance extracted by the optimal subgraph extraction strategy into the time-series knowledge warehouse, generate a zero-knowledge evidence chain that conforms to the W3CVerifiable Credentials data model, and solidify its immutability through the blockchain timestamp service to output verifiable credentials.

[0199] 1. Construction of zero-knowledge evidence chains.

[0200] The optimal subgraph instance is used as a blueprint to generate the main body of the evidence chain based on the W3C Verifiable Credentials data model. The main body fields are embedded with STARK zero-knowledge proofs, ensuring that the proof verification process does not expose the original node content or reasoning path. The proof system parameters are publicly verifiable, meeting the requirements of transparency and auditability.

[0201] 2. Two-layer Rollup anchoring mechanism.

[0202] The evidence chain summary is submitted to a Layer 2 Rollup contract after being hashed by Poseidon. The Rollup periodically anchors the aggregated hash to the Ethereum mainnet in batches. The original data of the complete evidence chain is stored in the InterPlanetary File System (IPFS) and forms an immutable binding with the on-chain hash through content addressing. The cost of a single anchoring is controlled within $0.001.

[0203] 3. Cross-chain mutual recognition.

[0204] Achieve bidirectional mutual recognition between Ethereum and Polygon through a cross-chain bridging protocol: Evidence chain hashes generate equivalent credentials on both chains respectively. After the bridging contract verifies the consistency between the two, the credentials can be independently verified on either chain, reducing the risk of single-chain failure and improving availability.

[0205] 4. Update delay and rollback protection.

[0206] Evidence chain update operations must be processed by an on-chain verifiable delay function (VDF) with a 10-minute delay period. During this period, any rollback attempt must recalculate the VDF result, and the computational cost increases with the on-chain difficulty parameter to ensure that the update cannot be quickly revoked.

[0207] 5. Voucher output and verification interface.

[0208] The final verifiable credentials output include:

[0209] Certificate identifier, issuance date, and validity period;

[0210] Zero-knowledge proof bytecode and corresponding public parameters;

[0211] On-chain anchored transaction hash and IPFS content identifier.

[0212] The verification interface provides two calling methods: RESTful and on-chain contract. It supports batch verification and real-time verification of single credentials, and returns boolean results and verification logs.

[0213] Through steps S1 to S7 above, an end-to-end automated, highly reliable, and traceable AI knowledge base is finally constructed. This knowledge base can achieve efficient integration and dynamic updating of multi-source heterogeneous data, support cross-modal retrieval and reasoning, and ensure the immutability and traceability of data through blockchain technology, meeting the needs of large-scale knowledge services.

[0214] Example 2

[0215] A multi-source data integration AI knowledge base construction system includes:

[0216] A heterogeneous data access framework with a built-in reversible crawler orchestrator, used to automatically generate replayable crawler scripts and output raw unstructured data sets on target sites without public APIs.

[0217] A pattern drift detector and protocol adaptive plugin are used to monitor field additions and deletions and perform hot-swap updates, outputting a continuously synchronized zero-downtime data stream;

[0218] The conflict resolution engine generates a version stamp with a digital watermark and lineage URI for each record based on the decay of the source credibility index, and outputs a set of resolved records with version stamps.

[0219] A multimodal contrastive learning encoder is used to map text, images, and time-series signals to a shared semantic space, outputting a multidimensional continuous vector with time decay awareness.

[0220] An incremental hypergraph knowledge graph builder uses multidimensional continuous vectors as nodes and probabilistically weighted, evolving hyperedges as connections. It supports hyperedge splitting and merging and outputs a hypergraph knowledge graph.

[0221] Generative adversarial reinforcement learning subgraph extractor is used to train the optimal subgraph extraction policy in the hypergraph knowledge graph and output candidate subgraphs.

[0222] The temporal knowledge repository and blockchain solidification module are used to write the optimal subgraph instance into the temporal knowledge repository, generate a zero-knowledge evidence chain that conforms to the W3C Verifiable Credentials data model, and solidify its immutability through blockchain timestamp service, outputting verifiable credentials.

[0223] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0224] Therefore, the present invention adopts the above-mentioned multi-source data integration AI knowledge base construction method and system to realize end-to-end automated, highly reliable, and traceable large-scale knowledge base construction.

[0225] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for constructing an AI knowledge base through multi-source data integration, characterized in that, Includes the following steps: Step S1: Through the reversible crawler orchestrator in the heterogeneous data access framework, automatically generate a replayable crawler script on the target site without a public API, pull unstructured documents, and output the original unstructured data set. Step S2: Using the original unstructured data set as input, the pattern drift detector monitors the addition or deletion of fields and triggers the protocol adaptive plugin to perform hot-plug updates, outputting a continuously synchronized zero-downtime data stream; Step S3: Taking the continuously synchronized zero-downtime data stream as input, perform conflict resolution based on the decay of the source credibility index, generate a version stamp with digital watermark and lineage URI for each record, and output a set of resolved records with version stamp; Step S4: Taking text, image, and time-series signal modalities from the resolved record set with version stamps as input, the unified contrastive learning loss function is used to map them to the shared semantic space, and the output is a multi-dimensional continuous vector with time decay awareness. Step S5: Using multidimensional continuous vectors as the node set and hyperedges with evolving probability weights that drift over time as connections, construct an incremental hypergraph knowledge graph that supports hyperedge splitting and merging, and output the hypergraph knowledge graph. Step S6: Using the hypergraph knowledge graph as the environment, train the optimal subgraph extraction strategy using the generative adversarial reinforcement learning framework. The generator simulates the query distribution and outputs candidate subgraphs, while the discriminator gives a reward based on information entropy gain and delay penalty. Step S7: Write the optimal subgraph instance extracted by the optimal subgraph extraction strategy into the time-series knowledge warehouse, generate a zero-knowledge evidence chain that conforms to the W3CVerifiable Credentials data model, and solidify its immutability through the blockchain timestamp service to output verifiable credentials.

2. The method for constructing an AI knowledge base by integrating multi-source data according to claim 1, characterized in that, Step S1 specifically includes: Using the target site's page as input, the script is recorded and played back in a sandbox environment, and the updated CSS / XPath selectors are automatically generated when the target site is redesigned. Taking CSS / XPath selectors and the target site's anti-crawling verification mechanism as input, it adopts a dual-channel parallel approach of headless browser and traffic replay, maintains crawling through behavioral fingerprint perturbation, and outputs an uninterrupted raw data pull stream. The script instructions from the fetching process are used as input, and the generated script is stored in the form of reversible bytecode. Using historical anti-scraping rule change data as input, the model is trained through a reinforcement learning-based anti-scraping strategy prediction module to generate behavioral perturbation schemes to cope with new anti-scraping mechanisms in advance. Taking the target site's IP reputation score and the proxy node status as input, the system dynamically switches proxy nodes through intelligent scheduling of a distributed proxy pool, and finally outputs a set of raw unstructured data.

3. The method for constructing an AI knowledge base by integrating multi-source data according to claim 1, characterized in that, Step S2 specifically includes: Using the field features of the original unstructured dataset as input, a bivariate statistic of field occurrence frequency and field type entropy is calculated using a sliding window approach. A dynamic threshold model is constructed based on the historical drift features of the data source, and a drift event is triggered when the joint deviation exceeds the corresponding threshold. Using drift events and original data protocol features as input, the protocol adaptive plugin generates and registers a new protocol template within 500ms after detecting a drift event by online matching and Bayesian updating based on a protocol template library. Using the output data of the old and new protocol templates as input, after verifying the consistency difference between the two outputs in parallel through the shadow channel to be ≤2%, switch to production traffic; Taking the warning signal that the field statistics are close to the threshold as input, lightweight preprocessing is triggered in advance, and finally the continuous synchronous zero-downtime data stream is output.

4. The method for constructing an AI knowledge base by integrating multi-source data according to claim 1, characterized in that, In step S3, conflict resolution based on the decay of the source credibility index, taking a continuously synchronized zero-downtime data stream as input, performs the following processing and outputs a set of resolved records with version stamps, specifically including: Using the historical records of each data source in the data stream as input, an accurate prior based on beta-Bernoulli dynamic updates is established for each data source, where the prior parameters decay with the real-time feedback error in an exponentially weighted moving average manner. The attenuated accuracy is taken as input and mapped to source credibility weights, which are used for weighted voting to resolve field-level conflicts. The source credibility weights are calculated as follows: oh source (t)=ω source (t-1)×λ Δt ; Where, ω source (t) represents the source credibility weight at time t, ω source (t-1) represents the source credibility weight at time t-1, Δt represents the time interval, and λ represents the decay coefficient; Using the conflict resolution results and data source reliability as inputs, ω source (t) is written as the version stamp of the corresponding record as metadata, and a version stamp with digital watermark and lineage URI is generated for each record. Using the data on changes in the accuracy of the data source as input, when the accuracy drops by more than 10%, the attenuation coefficient is temporarily adjusted to 0.8; By combining the domain prior weights of the data source, the final output is a set of resolved records with version stamps.

5. The method for constructing an AI knowledge base by integrating multi-source data according to claim 4, characterized in that, In step S3, the generation and embedding of the digital watermark takes the field data from the output de-watermarked record set as input, performs the following processing, and outputs records with digital watermarks, specifically including: Using the text content of the record field as input, a hidden watermarking technique based on zero-width Unicode characters is employed to embed a 128-bit hash digest into the blank space of the field value. Using watermarked field data as input, during the data leakage tracing phase, zero-width character sequences are extracted and decoded using regular expressions to achieve lossless tracing of the leakage source; Using the field sensitivity rating as input, the watermark embedding strength is dynamically adjusted to maintain a decoding accuracy of ≥99.9%, and the final output is a record with a digital watermark.

6. The method for constructing an AI knowledge base by integrating multi-source data according to claim 1, characterized in that, Step S4 specifically includes: Using three modal data as input, cross-modal anchor-positive sample-negative sample triples are constructed, and text, images, and time-series signals associated with the same business entity are marked as positive sample triples; Perform cross-modal hard negative sample mining within the same training batch; Using sample features and time decay parameters as input, a loss function L is used to make the gradient contribution of old samples decrease exponentially with time. The loss function L is as follows: L=L txt-img +L txt-ts +L img-ts +αL temporal-decay ; Among them, L txt-img L represents the contrastive learning loss between text and image modalities. txt-ts L represents the contrastive learning loss between text and temporal signal modes. img-ts L represents the contrastive learning loss between the image and the temporal signal modes, α represents the time decay regularization coefficient, and L represents the time decay regularization coefficient. temporal-decay This represents the time decay loss term; After training, the cosine distance between synonymous concepts of different modalities in the shared semantic space is ≤0.05, and the cosine distance between heteronymous concepts is ≥0.

35. Emerging modalities such as audio and 3D models are accessed through a standardized feature extraction protocol of the modal adaptation interface layer. The weight of modalities with high noise is reduced to 0.8 times the original weight, while the weight of stable modalities is maintained.

7. The method for constructing an AI knowledge base by integrating multi-source data according to claim 1, characterized in that, Step S5 specifically includes: The dynamic changes of the superedge weights follow a stochastic process driven by drift rate and volatility, where the drift rate is set by the domain prior and the volatility is updated in real time through online maximum likelihood estimation; A merge operation is triggered when the weight of the superedge drops to the first threshold of 0.05, and a split operation is triggered when the weight of the superedge rises to the second threshold of 0.95 and the number of associated nodes exceeds 10. A dynamic adjacency tensor storage structure accelerated by a graphics processor is used to achieve millisecond-level sparse updates based on a coordinate list format. Record the triggering factors for weight changes in the hyperedge attribute; Perform automatic aggregation on redundant superedges with a weight less than 0.

02.

8. The method for constructing an AI knowledge base by integrating multi-source data according to claim 1, characterized in that, In step S6, the adversarial reinforcement learning framework performs the following operations: The generator progressively selects nodes or superedges to add to the candidate subgraph using a pointer network structure. The discriminator reward function is defined as: the information entropy gain value minus 0.01 times the response delay value; In a federated training scenario, inject parameters ε=1 and δ=10 that satisfy privacy parameters. -5 Differential privacy noise is introduced, and training is terminated before the privacy budget is exhausted.

9. The method for constructing an AI knowledge base by integrating multi-source data according to claim 1, characterized in that, In step S7, the zero-knowledge evidence chain performs the following operations: By embedding STARK zero-knowledge proofs in the main fields of the evidence chain, we can achieve inference path verification without exposing the original nodes; A two-layer Rollup architecture is used to anchor the digest hash to the blockchain, while the complete chain of evidence is stored in the InterPlanetary File System. The cost of a single anchor is kept below $0.001, and cross-chain bridging protocol supports mutual recognition between Ethereum and Polygon. Evidence chain updates are implemented based on on-chain verifiable delay functions, with a 10-minute delay period set to defend against rollback attacks.

10. A multi-source data integration AI knowledge base construction system, characterized in that, include: A heterogeneous data access framework with a built-in reversible crawler orchestrator, used to automatically generate replayable crawler scripts and output raw unstructured data sets on target sites without public APIs. A pattern drift detector and protocol adaptive plugin are used to monitor field additions and deletions and perform hot-swap updates, outputting a continuously synchronized zero-downtime data stream; The conflict resolution engine generates a version stamp with a digital watermark and lineage URI for each record based on the decay of the source credibility index, and outputs a set of resolved records with version stamps. A multimodal contrastive learning encoder is used to map text, images, and time-series signals to a shared semantic space, outputting a multidimensional continuous vector with time decay awareness. An incremental hypergraph knowledge graph builder uses multidimensional continuous vectors as nodes and probabilistically weighted, evolving hyperedges as connections. It supports hyperedge splitting and merging and outputs a hypergraph knowledge graph. Generative adversarial reinforcement learning subgraph extractor is used to train the optimal subgraph extraction policy in the hypergraph knowledge graph and output candidate subgraphs. The temporal knowledge repository and blockchain solidification module are used to write the optimal subgraph instance into the temporal knowledge repository, generate a zero-knowledge evidence chain that conforms to the W3CVerifiable Credentials data model, and solidify its immutability through blockchain timestamp service, outputting verifiable credentials.

Images