Multi-source heterogeneous information fusion and analysis method for AI large model and data weaving

Abstract

The invention belongs to the technical field of multi-source heterogeneous data fusion and intelligent decision, and discloses a multi-source heterogeneous information fusion and analysis method for an AI large model and data knitting, which comprises the following steps: accessing multi-source heterogeneous data through data knitting, calibrating time and space by taking a key event as an anchor point, generating time and space alignment data, and synchronously constructing a dynamic metadata network. Generating a metadata graph; constructing a dynamic hypergraph, and calling a modal exclusive coding cluster to generate a multi-modal fusion feature; edge-cloud collaborative reasoning is started, a collaborative reasoning intermediate result is generated, confidence data is output by the general large model to arbitrate cross-domain conflicts, and a final decision packet is generated; triggering two-way feedback closed-loop optimization, driving data source optimization and knowledge completion of a metadata graph by an AI large model, further dynamically scheduling the priority of a data stream, and encrypting and updating large model parameters; and iterating the metadata atlas, and feeding back to a link of constructing a dynamic metadata network and generating a multi-modal fusion feature.

Description

Technical Field

[0001] The present invention relates to the field of multi-source heterogeneous data fusion and intelligent decision-making technology, and more specifically, to a multi-source heterogeneous information fusion and analysis method involving AI big models and data weaving. Background Art

[0002] Against the backdrop of accelerated digital transformation in areas such as the Industrial Internet, FinTech, and smart healthcare, multi-source heterogeneous data (such as sensor streams, text reports, images, and structured tables) are experiencing explosive growth. Cross-domain data fusion and real-time intelligent decision-making have become core requirements for improving business efficiency. However, current data integration faces problems such as data format differences, inconsistent spatiotemporal benchmarks, and semantic conflicts, making it difficult for traditional methods to fully tap the value of data.

[0003] To solve the above-mentioned data fusion and decision-making problems, existing technologies mostly adopt single-modal processing, static rule fusion or simple distributed reasoning solutions, but there are obvious limitations: existing technologies mostly focus on single-type data processing, lack unified access to multi-source heterogeneous data and high-precision spatiotemporal calibration capabilities, and find it difficult to eliminate time deviations and spatial positioning differences across devices; metadata management mostly relies on fixed ontologies and manual rules, and cannot automatically identify and resolve cross-domain semantic conflicts, resulting in low credibility of entity associations; the collaborative mechanism between the edge and the cloud is rigid, either sacrificing accuracy for speed, or relying on centralized processing resulting in excessive latency, and lacking dynamic adjustment capabilities, making it difficult to cope with new situations and new demands in dynamic scenarios. Overall, it is difficult to meet the requirements for data fusion depth, decision-making efficiency and system robustness in complex scenarios. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned objectives, the present invention provides the following technical solutions: a multi-source heterogeneous information fusion and analysis method for AI big model and data weaving, comprising: S1: Access multi-source heterogeneous data through data weaving, calibrate time and space with key events as anchors, and generate time-space aligned data packets; simultaneously build a dynamic metadata network, combine LLM to expand the domain ontology and eliminate semantic conflicts, and generate a metadata map; S2: Build a dynamic hypergraph based on spatiotemporal alignment data and metadata graphs; simultaneously call modality-specific coding clusters to process various types of data and generate multimodal fusion features; S3: Based on multimodal fusion features and a dynamic hypergraph, edge-cloud collaborative reasoning is initiated. Through edge reasoning and cloud-based AI large-scale model reasoning, intermediate collaborative reasoning results are generated. Simultaneously, the general large-scale model outputs confidence data to arbitrate cross-domain conflicts and generate the final decision package. S4: Based on the final decision package, a two-way feedback closed-loop optimization is triggered. The cloud-based AI big model drives data source optimization and metadata graph knowledge completion, dynamically scheduling data flow priorities and encrypting and updating cloud-based big model parameters. S5: Based on the edge-cloud learning results and the final decision package, the metadata graph is iterated and fed back to the construction of the dynamic metadata network and the generation of multimodal fusion features.

[0005] Furthermore, the method of spatiotemporally aligning data packets includes: Perform data classification and interface adaptation on the accessed multi-source heterogeneous data, define global data identification rules, generate a globally unique ID for each accessed data, and then build a metadata registry; Based on the metadata registry, time and space calibration is performed on multi-source heterogeneous data with key events as anchor points; Time calibration is as follows: by defining anchor trigger conditions to identify key events in the metadata registry as anchor points, and by calculating the time offset based on the timestamp of the anchor point, the original timestamp of the non-anchor point data is corrected; Spatial calibration involves defining a global target coordinate system, extracting physical device spatial metadata from multi-source heterogeneous data, and then uniformly converting the original coordinates of each physical device into the absolute coordinates of the global target coordinate system. Integrate all the data after spatiotemporal calibration and encapsulate them into a spatiotemporal alignment data package.

[0006] Furthermore, the metadata map is generated in the following manner: Extract technical metadata and business metadata based on spatiotemporal alignment of data packets to form an initialized dynamic metadata network; Preset the domain ontology based on business scenarios and define the concepts and relationships of the domain ontology; Call LLM to perform semantic parsing on the unstructured data in the spatiotemporal alignment data package to obtain entities, match the parsed entities with concepts, associate them with domain ontology, and then add the domain ontology to the dynamic metadata network; Detect cross-source semantic conflicts between entities through a dynamic metadata network and mark them as conflicting entities; Then, dynamic business weights are assigned to conflicting entities, and conflict resolution logic is defined to form conflict resolution rules; Entities are regarded as nodes, relationships as edges, and conflict resolution rules are associated with the corresponding nodes of conflicting entities as attribute constraints to construct a metadata graph.

[0007] Furthermore, the dynamic hypergraph is constructed in the following manner: Based on the spatiotemporal alignment of data packets and metadata graphs, we extract multimodal entities, assign hypergraph vertices to each entity, and add vertex attributes. Based on the entity relationships in the metadata graph, the hyperedge association rules are defined as follows: each hyperedge is associated with at least two or more hypergraph vertices; Then, add hyperedges to the hypergraph vertices according to the hyperedge association rules and label the corresponding attributes for the hyperedges; The hyperedge structure is dynamically optimized through predefined dynamic adjustment rules, and a dynamic hypergraph is generated based on the hypergraph vertices and hyperedges.

[0008] Furthermore, the multimodal fusion feature is generated by: Based on a dynamic hypergraph, the multimodal data to be processed is extracted by associating the hypergraph vertices with the spatiotemporal aligned data packets; According to the hypergraph vertex attributes, multimodal data is assigned to corresponding processing channels, and modality-specific coding clusters are called to perform feature extraction on different types of data to obtain dimensional features of the corresponding modality; The dimensional features of different modalities are processed in a unified dimension and fused to form multimodal fusion features.

[0009] Furthermore, the generation method of the collaborative reasoning intermediate result includes: Based on multimodal fusion features and dynamic hypergraph, fusion features are obtained through hyperedge confidence screening, and low-complexity association rules and high-complexity association rules are extracted from the dynamic hypergraph; Edge-cloud collaborative reasoning is defined as including edge reasoning and cloud reasoning; The edge is responsible for inferring and generating the intermediate feature vector for preliminary judgment through the filtered fusion features and low-complexity association rules; The cloud is responsible for combining multimodal fusion features, highly complex association rules, and intermediate feature vectors, performing deep reasoning through the cloud-based AI large model, and generating fine-grained decision recommendations; Integrate intermediate feature vectors and fine-grained decision recommendations to generate collaborative reasoning intermediate results containing reasoning basis.

[0010] Furthermore, the final decision package is generated in the following manner: The cloud-based AI model generates preliminary decisions based on the intermediate results of collaborative reasoning, integrating intermediate feature vectors, reasoning paths, and association rules. Call the general large model to perform cross-domain verification on the preliminary decision, output the decision confidence and conflict analysis report, and then perform hierarchical arbitration: If the decision confidence meets expectations and there is no conflict, the preliminary decision is directly adopted, supplemented by the verification conclusion of the general large model, to form the final decision conclusion; If the decision confidence does not meet expectations or there is a conflict, a secondary reasoning correction is performed based on the conflict analysis to re-form the final decision conclusion; Integrate decision conclusions and decision confidence to generate the final decision package.

[0011] Furthermore, the two-way feedback closed-loop optimization method includes: data quality detection optimization and metadata graph integrity detection optimization; Data quality detection and optimization: Based on the final decision package, the decision conclusion is associated with the data source, and data quality indicators are calculated by type. For data sources that do not meet the data quality indicators, an alternative data source is matched from the metadata map and switched. Optimize metadata graph integrity detection: extract new entities from decision conclusions, perform semantic matching with existing entities in the metadata graph through LLM, and obtain entity matching degree; If the entity matching degree does not meet expectations, it is determined that the graph is missing an entity. The associated features of the new entity are mined from the historical spatiotemporal alignment data package and added to the metadata graph.

[0012] Furthermore, the method of dynamically scheduling data stream priorities and encrypting and updating cloud-based AI model parameters includes: Based on the results of the two-way feedback closed-loop optimization, the actual bandwidth utilization of each data source is quantified, and task processing priorities are defined to divide all data into corresponding data streams; Then dynamically adjust the bandwidth allocation of all data flows according to the task processing priority; At the same time, based on the optimization results of metadata graph integrity detection, through the edge-cloud learning process, local training and gradient encryption are performed on the edge, and gradient aggregation and model updates are performed on the cloud; Then, the edge-cloud learning results are integrated to update the reasoning process of edge-cloud collaborative reasoning.

[0013] Furthermore, the method of iterating the metadata graph includes: Extract structured triples from edge-cloud learning results and final decision packages, write them into the metadata graph, compress redundant relationships through knowledge distillation, and iteratively update the metadata graph; The updated content of the metadata graph after iteration is converted into specific rules and fed back to the construction process of the dynamic metadata network and the generation of multimodal fusion features, forming a closed-loop optimization.

[0014] The technical effects and advantages of the multi-source heterogeneous information fusion and analysis method of the present invention's AI big model and data weaving are as follows: The present invention achieves efficient fusion and intelligent decision-making of multi-source heterogeneous data through a full-process collaborative architecture. First, unified access and standardized processing of multi-source heterogeneous data are achieved through intelligent weaving and spatiotemporal calibration. By constructing a dynamic metadata network and aligning spatiotemporal benchmarks, combined with LLM, the domain ontology is expanded and semantic conflicts are resolved, laying a high-quality data foundation for subsequent fusion and reasoning.

[0015] Secondly, a dynamic hypergraph and cross-modal fusion are constructed. The hyperedges can be dynamically adjusted according to the strength of data association. At the same time, through modality-specific encoding and unified dimensional processing, the features of different types of data are fused into a unified semantic space representation, realizing deep semantic association of cross-domain data.

[0016] Next, by building an edge-cloud collaborative reasoning mechanism, the edge quickly responds to simple associations, and the cloud deeply processes complex reasoning. Combined with the verification and conflict arbitration of general large models, the accuracy and reliability of decisions are improved while ensuring real-time performance, achieving a balance between efficiency and precision.

[0017] Furthermore, based on the decision results, data source optimization and metadata graph completion are triggered. At the same time, data flow priorities are dynamically scheduled through reinforcement learning, and federated learning is used to encrypt and update model parameters, enabling the system to adapt to dynamic scene changes and improve resource utilization efficiency and model adaptability.

[0018] Ultimately, the new entities and relationships generated in the decision-making process are precipitated into a knowledge graph and fed back to the dynamic metadata network construction and multimodal fusion links, forming a closed-loop evolution mechanism of data-decision-knowledge, and continuously improving the system's intelligence level and scenario adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A flowchart of the multi-source heterogeneous information fusion and analysis method for the AI ​​large model and data weaving of the present invention; Figure 2 A schematic diagram of the process of multimodal entity extraction and vertex initialization in the multi-source heterogeneous information fusion and analysis method of the present invention for the AI ​​big model and data weaving; Figure 3 Schematic diagram of the multi-source heterogeneous information fusion and analysis system flow for the AI ​​big model and data weaving of the present invention. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention. Example 1

[0021] See also Figure 1 As shown, the multi-source heterogeneous information fusion and analysis method of the AI ​​big model and data weaving described in this embodiment includes: S1: Access multi-source heterogeneous data through data weaving, calibrate time and space with key events as anchors, and generate time-space aligned data packets; simultaneously build a dynamic metadata network, combine LLM to expand the domain ontology and eliminate semantic conflicts, and generate a metadata map; S2: Build a dynamic hypergraph based on spatiotemporal alignment data and metadata graphs; simultaneously call modality-specific coding clusters to process various types of data and generate multimodal fusion features; S3: Based on multimodal fusion features and a dynamic hypergraph, edge-cloud collaborative reasoning is initiated. Through edge reasoning and cloud-based AI large-scale model reasoning, intermediate collaborative reasoning results are generated. Simultaneously, the general large-scale model outputs confidence data to arbitrate cross-domain conflicts and generate the final decision package. S4: Based on the final decision package, a two-way feedback closed-loop optimization is triggered. The cloud-based AI big model drives data source optimization and metadata graph knowledge completion, dynamically scheduling data flow priorities and encrypting and updating cloud-based big model parameters. S5: Based on the edge-cloud learning results and the final decision package, the metadata graph is iterated and fed back to the construction of the dynamic metadata network and the generation of multimodal fusion features.

[0022] Ways to align packets in time and space include: Deploy a unified access gateway that supports multiple protocols. Use a data weaving engine (such as the Denodo platform) to scan multiple heterogeneous data sources and label the protocol types (including structured data, time series data, and unstructured data). Classify the incoming data and adapt the interface, define global data identification rules (format: data source type - device / table name - timestamp - random number), and generate a globally unique ID for each incoming data; Among them, classification and interface adaptation include: Structured data (structured databases such as MySQL / Oracle, synchronized using the JDBC protocol to generate structured data identifiers), time series data (time series data from sensors / PLCs, accessed via the MQTT protocol, pre-processed by the edge gateway, and encapsulated in the "device ID - timestamp - value" format), and unstructured data (text reports, images, etc., connected to file servers via HTTP APIs to extract file metadata and generate unique identifiers). Deploy a distributed metadata service (such as Apache Atlas) and establish a metadata registry containing "global ID, data source type, physical storage address, logical view path, access status, and last synchronization time" to implement transparent call management of "query logical path → return physical data"; The metadata registry contains the following contents: field name, description, and field value, such as (field name: logic_path, corresponding description: logical view path, corresponding exemplary field value: industrial / sensor / vibrationdata_id); Time calibration involves defining anchor trigger conditions based on usage scenarios (e.g., industrial or financial) through a rule engine (e.g., Drools). For example, in industrial scenarios, if the device is powered on and the current is greater than 10A, the "Device Power On" anchor is triggered; in financial scenarios, if the payment is successful and the amount is greater than 0, the "Transaction Completed" anchor is triggered. Key events (e.g., device power on, transaction completed) are identified from the metadata registry as anchors. Anchor timestamps are recorded based on a high-precision clock synchronization protocol. Flink's sliding window technology (defined window size and sliding step, such as window size = 10s, sliding step = 1s) is used to match the preceding and following anchor points of non-anchor data (for example, if the non-anchor timestamp is T, match the nearest anchor point T1 before T and the nearest anchor point T2 after T). Calculate the time offset: Calculate the ratio of the physical time difference (T2-T1) to the device's local time difference (T2'-T1' recorded by the device, where T1' and T2' are the anchor timestamps before and after the device). This is used as the calibration coefficient K. This is then used to calculate the time offset of the non-anchor data (time offset = K × (T-T1')) and correct the non-anchor timestamp (i.e., add the time offset to the original timestamp). Spatial calibration involves extracting device spatial metadata (original coordinate system and semantic description, such as "machine tool No. 3 in workshop area A") through the IoT platform. Identify the type of the device's original coordinate system (e.g., relative coordinates, semantic spatial description) and determine its reference system (e.g., device local coordinate system, regional division of text description); The global target reference system is defined as the WGS84 latitude and longitude coordinate system (or an absolute coordinate system specified according to scenario requirements). The conversion model is adapted according to the original coordinate type: for relative coordinates (such as the offset of the device relative to the installation point), the affine transformation method is used; for local coordinate system coordinates (such as the X / Y axis coordinates in the workshop), they are converted to absolute longitude and latitude using the Gauss-Krüger projection method; for semantic spatial descriptions, the LLM (i.e., AI large language model) is used to map them to a preset spatial coding table (for example, the longitude and latitude range for "workshop area A" is: N30.1234°~N30.1236°, E120.5678°~E120.5680°). Collect device control points with known absolute coordinates (such as ground fixed markers and engineering surveying points), obtain their coordinate values ​​in the device's original coordinate system, and calculate the parameters of the corresponding transformation method (such as affine transformation matrix and projection coefficients) using the least squares method; After the device's original coordinates are converted using a corresponding conversion method to obtain a preliminary conversion result, they are compared with the absolute coordinates of the device's control points to calculate the error value (such as the root mean square error). If the error exceeds a preset coordinate error threshold (such as 0.5 meters), the conversion parameters are iteratively optimized using the gradient descent method until the error meets the standard. Ultimately, the absolute coordinates are unified into the global target coordinate system, ensuring that the coordinates of different types of devices can be correlated and compared within the same spatial framework. Integrate all the data after spatiotemporal calibration and encapsulate them into a spatiotemporal alignment data package (which needs to include core fields: globally unique ID, data source type, original data content, calibrated timestamp, latitude and longitude of the calibrated coordinates, and data status (valid / invalid)).

[0023] The metadata graph is generated in the following ways: Extract technical metadata and business metadata based on spatiotemporal alignment of data packets to form an initialized dynamic metadata network; Specifically, the method of extracting technical metadata is to parse the data source type, device / table name, timestamp and other information from the spatiotemporal alignment data packet, and then record the data format, field (such as JSON / CSV), field type (such as numeric / string), update frequency (such as real-time streaming / daily synchronization) and storage location (i.e., the physical storage address in the metadata registry) as technical metadata; Business metadata is extracted by mapping field names (e.g., "order_id" corresponds to "order number") and parsing data content (e.g., extracting "equipment model" and "transaction type" from text) to determine the business domain to which the data belongs (e.g., "industrial equipment monitoring" and "financial transactions"). The extracted metadata is organized into a hierarchy of "data source - business domain - field" to form the basic framework of the dynamic metadata network, which serves as the initialization of the dynamic metadata network. Preset the domain ontology based on business scenarios and define the concepts and relationships of the domain ontology; The concepts of domain ontology need to be customized according to different actual business scenarios, such as equipment, faults, and sensors in industrial scenarios, and users, transactions, and risk levels in financial scenarios. The definition of relationships is similar, such as "equipment-container-sensor" in industrial scenarios and "transaction-association-user" in financial scenarios; The domain ontology structure can then be described using the OWL language. Concepts are abstracted into classes, relationships are defined as object attributes, and the attribute features in concepts and relationships (such as "sensor accuracy" and "transaction amount") are defined as data attributes. These generated OWL files are stored in the ontology knowledge base. LLM is called to perform semantic parsing on the unstructured data in the spatiotemporal alignment data package to obtain entities. The parsed entities are matched with concepts (for example, "bearing noise" is associated with the "fault" class) and associated with the domain ontology. If the entity is new (such as "new sensor fault"), a subclass is automatically created in the domain ontology and associated with the parent class (such as "fault"), and a corresponding relationship is added. The domain ontology is then added to the dynamic metadata network, forming a two-layer structure of domain ontology and dynamic instances (the ontology ensures stability, and the instances support flexibility). The semantic parsing method for unstructured data includes: performing entity recognition (e.g., temperature surge, abnormal bearing noise) and relationship extraction (e.g., temperature surge leads to equipment shutdown) on text logs; parsing scenario-based entities (e.g., "equipment oil leak area belongs to machine tool No. 3") on image tags; Detect cross-source semantic conflicts between entities through a dynamic metadata network and mark them as conflicting entities (i.e., entities with conflicts); Specifically, the cross-source semantic conflict detection method is to compare the semantic descriptions of cross-source data through the metadata network to define and identify conflict scenarios; where conflict scenarios are defined as homonyms with different meanings and homonyms with different names; Then, dynamic business weights are assigned to conflicting entities, and conflict resolution logic is defined to form conflict resolution rules; Specifically, based on historical data (such as the accuracy of association decisions) or expert experience, conflicting entities are assigned dynamic business weights in different business scenarios (for example, "customer value" in the e-commerce business domain has a weight of 0.8 in the consumer analysis scenario and 0.2 in the financial scenario). The conflict resolution logic is defined to prioritize business domain-specific definitions. Integrate dynamic business weights and conflict resolution logic into conflict resolution rules; Build a metadata graph by treating entities as nodes (adding corresponding node attributes: name, type, dynamic business weight), relationships as edges, and associating conflict resolution rules as attribute constraints to the corresponding nodes of conflicting entities. Based on domain knowledge, we preset the basic weight of the relationship (for example, the initial weight of "sensor-monitoring-device" is 0.9). We then calculate the frequency of entities corresponding to the relationship in the spatiotemporal alignment data package as the co-occurrence frequency. This is then dynamically adjusted with the time decay factor to obtain the optimal relationship weight. This relationship weight is then added as an attribute to the corresponding edge. Relationship weight = decay factor × base weight + (1-decay factor) × co-occurrence frequency; Among them, the time decay factor is a preset value, usually 0.3, and its setting logic is that recent data has a higher weight and early data has a lower weight.

[0024] The construction methods of dynamic hypergraph include: Based on the spatiotemporal alignment of data packets and metadata graphs, we extract multimodal entities, assign hypergraph vertices to each entity, and add vertex attributes, including spatiotemporal coordinates (timestamps and longitude and latitude after spatiotemporal alignment), data types (such as text, images, and sensors), and ontology associations (mapping relationships with domain ontologies in the metadata graph (such as “lesions” being associated with the “pathological features” class)). Among them, the extraction method of multimodal entities is as follows: all data modalities in the spatiotemporal alignment data package are marked by the metadata registry, and the corresponding candidate entities are extracted by calling the modality-specific feature extraction tool; For example, for text classes (documents, logs), NLP tools (such as spaCy) can be used to extract noun phrases and professional terms as candidate entities; For visual classes (images, video frames), target detection / segmentation models (such as Faster R-CNN and Mask R-CNN) can be used to extract ROI regions or keyframe objects as candidate entities. It should be noted that different modal data require different tools to extract candidate entities. There are many options, and specific selection and adjustment can be made according to actual conditions. When new data types are added and the corresponding feature extractors are expanded, the scalability of entity recognition can be achieved. After extracting all candidate entities, LLM is used to unify the names of candidate entities (e.g., "device" is merged into "equipment"), annotate the entity's category (e.g., "fault", "component"), and associate it with the domain ontology (e.g., "lesion" is associated with the "pathological feature" class), thereby forming a multimodal entity. Assign a hypergraph vertex ID to each multimodal entity. After assigning the hypergraph vertices, merge the duplicate or synonymous multimodal entities into a unique hypergraph vertex. Based on the entity relationships in the metadata graph, the hyperedge association rules are defined as follows: each hyperedge is associated with at least two or more hypergraph vertices, and the association logic directly maps the relationships in the metadata graph; For example, in the metadata graph, "vibration sensor-monitoring-equipment bearing" and "equipment bearing-existence-wear fault" → hyperedges associate "vibration sensor vertex", "equipment bearing vertex" and "wear fault vertex"; Then, add hyperedges to the hypergraph vertices according to the hyperedge association rules and label the corresponding attributes for the hyperedges; Among them, the attributes of hyperedge annotation include association type (based on the relationship type definition of the metadata graph (such as "causal relationship", "temporal relationship", and "monitoring relationship"), confidence (directly using the relationship weight in the metadata graph as the hyperedge confidence), and spatiotemporal constraints (based on the spatiotemporal calibrated data in the spatiotemporal alignment data package, defining the time window (such as "within 10 minutes after the equipment is started") and spatial range (such as "workshop area A") of the hyperedge association as spatiotemporal constraints); Monitoring data spatiotemporal distribution: Use statistical methods (such as the 3σ principle) to detect whether the incoming multi-source heterogeneous data exceeds the normal range; Monitor association confidence: Regularly calculate the confidence of the hyperedge. If it is lower than a certain proportion of the edge base weight in the metadata graph (e.g., <80%, i.e., less than 80% of the edge base weight, the specific proportion is flexibly adjusted according to actual conditions), it is marked as a low-confidence association; if it is higher than a certain proportion of the edge base weight (e.g., >120%, i.e., greater than the edge base weight, and the excess is higher than 20% of the base weight), it is marked as a high-confidence association. Dynamically optimize the hyperedge structure through predefined dynamic adjustment rules, and then generate a dynamic hypergraph based on the hypergraph vertices and hyperedges, including vertices, hyperedges and dynamic adjustment rules, to provide a cross-domain semantic association framework for subsequent multimodal fusion; The dynamic adjustment rules are as follows: when a hyperedge is marked as a low-confidence association, the hyperedge is automatically deleted and the reason for deletion is recorded; if there is new data and the new data has a corresponding relationship in the metadata graph, a new hyperedge is added.

[0025] The generation methods of multimodal fusion features include: Based on a dynamic hypergraph, the hypergraph vertex ID is associated with the spatiotemporal alignment data packet to extract the multimodal data to be processed (such as text reports, image files, sensor time series values, structured tables, etc.); According to the data type in the hypergraph vertex attributes, multimodal data is assigned to the corresponding processing channel. At the same time, the association type and confidence level of the hypergraph hyperedge are associated (such as the causal relationship between "vibration sensor and temperature sensor" with a confidence level of 0.85) as the prior weight for subsequent fusion. Then, the modality-specific encoding cluster is called to perform feature extraction on different types of data to obtain the dimensional features of the corresponding modality (i.e., the extracted high-dimensional features or low-dimensional features); Among them, the modality-specific coding cluster is a preset coding tool set that includes multiple modality data feature extraction tools. This set can be increased or decreased according to actual conditions; Specifically, common multimodal data types are usually structured data, text data, image data, time series data, etc. For these types of data, there are currently very mature feature extraction methods; For example, for structured data, the attention mechanism of the TabNet tool can be used to screen key features and output a 128-dimensional feature vector. For text, the semantic vector can be extracted using the RoBERTa tool in combination with the industry's terminology library, outputting a 768-dimensional semantic feature vector. For images, the ViT tool can be used to extract spatial features and output a 1024-dimensional feature vector. For time series data, the TFT tool can be used to capture trends and output a 512-dimensional feature vector. It should be noted that the tools mentioned here are mature and can be used directly; Perform unified dimension processing on dimensional features of different modalities and fuse them to form multimodal fusion features; Among them, because the feature dimensions extracted from data of different modalities are different, for example, structured data extracts 128-dimensional features, while image data extracts 1024-dimensional features, it is necessary to unify the dimensions to facilitate subsequent use; The unified dimension processing method is: Compress high-dimensional features of different modalities to 256 dimensions through VAE (Variational Autoencoder); expand low-dimensional features to 256 dimensions through GAN (Generative Adversarial Network); After integrating all 256-dimensional features with unified dimensions, Z-score normalization is performed (mean = 0, standard deviation = 1) to eliminate dimensional differences; The method of fusing to form multimodal fusion features is as follows: input the 256-dimensional normalized features of each modality, use the hyperedge confidence as the initial weight, and fuse them through the cross-attention mechanism, including: The inter-modal similarity matrix (such as the cosine similarity between vibration features and temperature features) is calculated and normalized using the Softmax function to obtain the final attention weight (reflecting the real-time correlation strength between modalities, such as "when vibration is abnormal, the weight of the temperature feature increases to 0.9"). The 256-dimensional features of each modality are weightedly fused according to the attention weight to obtain the fused intermediate features. The intermediate features are then mapped to a 512-dimensional unified semantic space through a fully connected layer, and the fused features of the 512-dimensional unified semantic space are used as the final multimodal fusion features.

[0026] The generation methods of collaborative reasoning intermediate results include: Based on multimodal fusion features and dynamic hypergraph, fusion features are obtained by hyperedge confidence screening; One way to filter features by hyperedge confidence is to associate the vertices and hyperedges of the dynamic hypergraph with entities in the feature vector and extract the corresponding association information (such as the causal relationship between "vibration feature-temperature feature", with a confidence of 0.85); Through multimodal feature selection technology, multimodal fusion features are weighted based on hyperedge confidence. Among them, 90% of the dimensions of features associated with high confidence are retained, and only the core 30% of the dimensions of features associated with low confidence are retained. Finally, the filtered fusion features are formed to reduce redundant calculations; For example, in industrial scenarios, a high-confidence association between a vibration sensor and equipment failure will retain key dimensions such as vibration frequency and peak value, while a low-confidence association with humidity data will only retain the mean and variance; Then, low-complexity association rules and high-complexity association rules are extracted from the dynamic hypergraph; Specifically, low-complexity association rules refer to associations that can be judged through simple logic. A hyperedge that only associates two hypergraph vertices (e.g., a single sensor anomaly → preliminary warning) is defined as a low-complexity association rule. The Drools rule engine is used to parse the rules into "if-else" logic and store them in the local rule base at the edge. High-complexity association rules refer to associations that require cross-domain fusion reasoning. Hyperedges that associate three or more hypergraph vertices (e.g., abnormal vibration frequency + sudden temperature rise + historical fault records → comprehensive fault type) are defined as high-complexity association rules and retained as hypergraph structured data. Based on the latency requirements of business scenarios (e.g., industrial fault warnings must be ≤100ms, financial risk control must be ≤50ms), the filtered feature data and association rules are marked with priorities (such as high, medium, and low), giving priority to high-priority tasks. Edge-cloud collaborative reasoning is defined as including edge reasoning and cloud reasoning; The edge is responsible for inferring and generating the intermediate feature vector for preliminary judgment through the filtered fusion features and low-complexity association rules; Specifically, by deploying pre-trained lightweight models (such as TinyBERT and MobileNetV3) on the edge, the fused features and low-complexity association rules are used as input. The PCA algorithm is used to compress the fused features to 64 dimensions. Based on the low-complexity association rules, the basic associations in the dynamic hypergraph are matched and a preliminary judgment result (such as "anomaly type: temperature is too high; confidence level: 0.7") is output. The result is encapsulated as an intermediate feature vector containing basic information such as timestamp and device ID. It should be noted that this basic association is for low-complexity association rules. For example, if the low-complexity association rule is: single sensor anomaly → preliminary warning, then the basic association may be: single sensor value > threshold → single sensor anomaly"; The cloud is responsible for combining multimodal fusion features, highly complex association rules, and intermediate feature vectors, performing deep reasoning through the cloud-based AI large model, and generating fine-grained decision recommendations and reasoning paths. Specifically, a pre-trained cloud-based AI model (such as LLaMA 3-70B fine-tuned for business scenarios) is loaded into the cloud. Intermediate feature vectors are then concatenated with multimodal fusion features. Entity relationships within highly complex association rules (such as "Equipment S001 - Association - Historical Failures 3 Times") are converted into text prompts. The text prompts and concatenated features are then fed into the AI ​​model, which then outputs fine-grained decision recommendations (such as "Fault Type: Bearing Wear; Repair Priority: High; Association Evidence: Vibration Frequency Matches Historical Failure Pattern + Temperature Anomaly Intensifies") and generates an inference path (recording the logical chain from features to conclusions). Integrate intermediate feature vectors and fine-grained decision recommendations to generate collaborative reasoning intermediate results with reasoning evidence, including intermediate feature vectors at the edge, fine-grained decision recommendations, confidence scores, association rules in the dynamic hypergraph, and reasoning evidence; Among them, if the conclusions of the edge and cloud are consistent (for example, both are marked as "bearing wear"), the results are merged and the highest confidence level is retained; if there are differences (for example, the edge marks "temperature is too high" and the cloud judges "bearing wear"), the results and reasoning basis of both sides are retained.

[0027] The final decision package is generated by: The cloud-based AI model generates preliminary decisions based on the intermediate results of collaborative reasoning, integrating intermediate feature vectors, reasoning paths, and association rules, and simultaneously generates decision-making basis. The logic for generating the preliminary decision is as follows: The cloud-based AI model uses a multi-round attention mechanism to focus on key features and outputs specific decision conclusions based on pre-set business constraints (such as the "fault warning classification standard" for industrial scenarios: abnormal probability ≥ 0.8 → level 1 warning). For example, in an industrial scenario: "Equipment S001 fault warning (type: bearing wear), maintenance priority: high"; Decision-making basis is annotated using a chain structure of "feature-association rule-conclusion", for example: "Basis 1: Vibration frequency 50Hz (threshold 45Hz) → matches historical fault mode A (confidence 0.92); Basis 2: Hypergraph association shows 'abnormal vibration - bearing wear' confidence 0.85", thus ensuring traceability; Calling a general large model (such as GPT-4o, Gemini Ultra) to perform cross-domain verification on the preliminary decision, outputting a decision confidence and conflict analysis report, and then performing hierarchical arbitration: Specifically, it integrates preliminary decisions, association rules of dynamic hypergraphs, multimodal fusion features, and business constraint rules to generate natural language prompts. Then, the general large model performs operations based on the natural language prompts, including generating decision confidence and conflict analysis reports. Among them, the decision confidence level (0-100 points) is generated by comparing feature matching, rule compliance, and historical case consistency (e.g., "the vibration feature matches but the temperature is normal, the confidence level is 75 points"); Generate a conflict analysis report: clarify the conflict dimension (e.g., "current temperature is normal but the fault warning is inconsistent"), conflict severity (e.g., "low impact: single feature deviation; high impact: core rule conflict"), and potential causes (e.g., "sensor temperature data delay / vibration false alarm"); If the decision confidence meets expectations (e.g., ≥80) and there are no conflicts, the preliminary decision is directly adopted, supplemented with the verification conclusion of the general large model (e.g., "Verification passed: feature association meets business rules"), and a final decision conclusion is formed (e.g., "Equipment S001 fault warning: slight vibration interference, re-inspection recommended in 30 minutes"); If the decision confidence does not meet expectations (e.g., <80) or there is a conflict, a secondary reasoning correction is performed based on the conflict analysis to re-form the final decision conclusion; Among them, the secondary reasoning correction is as follows: based on the missing information pointed out by the conflict analysis (such as "normal temperature conflicts with fault warning → additional pressure sensor data is needed"), supplementary features are extracted from the spatiotemporal aligned data packets associated with the dynamic hypergraph (such as "the last 5 minutes of data of the pressure sensor of device S001"); Call a lightweight and fast inference model (such as DistilBERT), input the preliminary decision, supplementary features, and conflict analysis results, and recalculate the decision conclusion (for example, "After adding the normal pressure feature, the fault type is corrected to 'minor vibration interference' and the risk level is reduced to medium"); Record the correction logic (e.g., "Supplementary pressure data is normal → Bearing wear is ruled out, supporting interference judgment"), update the confidence level (e.g., from 75 to 82), and then use the recalculated decision conclusion as the final conclusion; Integrate decision conclusions and decision confidence to generate a final decision package, including decision conclusions, a complete basis chain (including the decision basis of the preliminary decision, the verification results of the general large model, and the correction logic of the secondary reasoning), decision confidence, timestamp and associated entity ID.

[0028] The two-way feedback closed-loop optimization methods include: data quality detection optimization and metadata graph integrity detection optimization; Data quality detection and optimization: Based on the final decision package, the decision conclusion is associated with the data source, and data quality indicators are calculated by type. For data sources that do not meet the data quality indicators, an alternative data source is matched from the metadata map and switched. Specifically, for all data sources associated with the decision conclusion, data quality indicators are defined for time series data, structured data, and unstructured data in the spatiotemporal alignment data package: Time series data: Calculate the noise rate (the degree to which data fluctuations deviate from the normal trend), missing rate (the proportion of data not sampled as expected per unit time), and timeliness (the delay from data generation to data access). If any indicator exceeds the corresponding threshold, it is marked as an anomaly. Structured data: Calculate the noise rate (noise rate = (null values ​​+ values ​​outside the reasonable range + logically contradictory values) / total number of records), consistency (consistency = 1 - (number of records with mismatched associated fields / total number of associated records)), and timeliness (timeliness = current time - last data update time). If the noise rate or timeliness exceeds the corresponding threshold, or the consistency is less than the corresponding threshold, an anomaly is marked. Unstructured data: Calculate the noise rate (percentage of invalid content), completeness (completeness = actual number of key information items / preset number of key information items), and timeliness (timeliness = data generation time - corresponding event occurrence time). If the noise rate or timeliness exceeds the corresponding threshold, or the completeness is less than the corresponding threshold, the data is marked as abnormal. If a data source is marked as abnormal and determined to not meet data quality indicators, a backup data source of the same type and accuracy level is retrieved from the metadata map. The link is switched through the interface and retested after the switch. If the data quality indicators are met, a switch log is recorded (including the cause of the abnormality, source IDs before and after the switch, and test results). Metadata graph integrity check optimization: New entities are extracted from the decision conclusions, standardized using LLM (unified naming, category labeling, and attribute extraction), and then semantically matched with existing entities in the metadata graph using the TransE entity linking algorithm. New entities and similar entities in the metadata graph are encoded into vectors using BERT. The cosine similarity formula is then used to calculate the entity matching degree. If the entity matching degree does not meet expectations (e.g., <0.7), it is determined that the graph is missing an entity. LLM is used to mine the associated features of the new entity from the historical spatiotemporal alignment data package and add them to the metadata graph. The method for mining the association features of newly added entities is to take the newly added entity name, historical spatiotemporal alignment data (such as data on the same device in the past six months), and the relationship in the metadata graph as input, and output the association features of the newly added entity through LLM mining; The mining logic of LLM is as follows: LLM extracts “feature-condition-result” associations through multiple rounds of reasoning and generates triples; Verify the associated features generated by LLM through the domain expert rule base (such as industrial fault manual). If the verification passes, it is confirmed to be valid; Add the verified new entities and associated features to the metadata graph, and update entity attributes and relationships.

[0029] Methods for dynamically scheduling data stream priorities and encrypting and updating AI model parameters include: Based on the results of the two-way feedback closed-loop optimization, the actual bandwidth utilization of each data source is quantified, and task processing priorities are defined to divide all data into corresponding data streams; Specifically, a reinforcement learning scheduler (using the Deep Q-Learning framework) is launched. The real-time transmission rate of each data source is quantified by the amount of data per second, which serves as the actual bandwidth utilization. Task processing priorities are then divided into three levels (e.g., high, medium, and low) based on business urgency and impact, and mapped to data labels (e.g., high = 3, medium = 2, low = 1). Among them, the logic of high-task processing priority division is urgent data, such as financial risk control (such as real-time anti-fraud) and industrial fault warning (such as emergency equipment failure); The logic of task processing priority is routine data, such as routine business reports (such as daily transaction statistics) and equipment status inspections; The logic for dividing low-task processing priority into statistical analysis data, such as historical data archiving and non-core log analysis; Then, according to the task processing priority, the data streams are divided into high-priority data streams (such as high-frequency trading streams), medium-priority data streams (such as device status streams), and low-priority data streams (such as historical logs); The actual bandwidth utilization and task processing priority are input into the reinforcement learning scheduler in numerical form as the state space of the scheduler; Expert experience is used to define the adjustment range (0% to 100%) of the bandwidth allocation ratio for each data source. This range serves as the scheduler's action space, and allows the action space to adjust the bandwidth of a single data stream in steps (e.g., from 80% to 85%) according to a preset ratio (e.g., ±5%). With task completion timeliness and bandwidth utilization as the core, the scheduler's reward function is defined as follows: the ratio of task completion time to planned completion time is taken as the task completion timeliness; the ratio of actual bandwidth utilization to planned target utilization is taken as the bandwidth utilization; the task completion timeliness and bandwidth utilization are weighted and integrated to obtain the reward; In the weighted fusion process, the sum of the weights of task completion time and bandwidth utilization is 1. The specific weight distribution can be adjusted based on expert experience. Then, based on the reinforcement learning scheduler, the bandwidth allocation of all data flows is dynamically adjusted according to the task processing priority; Bandwidth allocation rules are as follows: High-priority data streams: 70%-90% of the bandwidth (adjusted according to real-time rate fluctuations, 90% at peak); Medium priority data flow: 10%-25% bandwidth is allocated; Low-priority data streams: allocated ≤10% of the bandwidth and transmitted only when the tasks of high / medium-priority data streams are idle; Define and monitor the total bandwidth allocation of data flows in real time: If the task delay of high-priority data flows meets the expected level (e.g., ≤50ms), the bandwidth utilization meets the expected level (e.g., ≥85%), and there is no data loss for tasks of low-priority data flows, the total bandwidth allocation is considered to have met the requirements. If not, the scheduler is triggered to re-determine the allocation. At the same time, based on the optimization results of metadata graph integrity detection, through the edge-cloud learning process based on federated learning, local training and gradient encryption are performed on the edge, and gradient aggregation and model updates are performed on the cloud. Then, the edge-cloud learning results are integrated and the reasoning process of edge-cloud collaborative reasoning is updated. Local training and gradient encryption involves screening local data on each edge node (such as industrial gateways and regional servers), selecting high-quality data from the recent period, and using this data to calculate the gradients of the parameters of each layer of the local model. Laplace noise is added to the gradients (the noise intensity is inversely proportional to the amount of local data; the larger the data volume, the smaller the noise, ensuring that the gradient does not leak the local original data). The gradient data is encrypted using the TLS 1.3 protocol, and node identity authentication information (such as the edge device ID + digital signature) is attached before uploading to the cloud-based federated learning server. Gradient aggregation and model update are as follows: verify the identity legitimacy and gradient validity of each edge node, then use the FedAvg algorithm to assign weights according to the local data volume of each edge node (the larger the data volume, the higher the weight), and perform weighted aggregation on the gradients of each edge node. The parameter distillation mechanism retains more than 95% of historical knowledge, and then weightedly fuses the newly aggregated gradients with the original model parameters (for example, new gradient weight = 0.3, original parameter weight = 0.7) to avoid catastrophic forgetting; finally, the updated model parameters are sharded and compressed and pushed to each edge node through the edge synchronization protocol.

[0030] Ways to iterate the metadata graph and provide feedback include: Extract structured triples from edge-cloud learning results and final decision packages; Specifically, the core entities and relationships are extracted from the final decision package. For example, the decision conclusion is “Vibration frequency > 50 Hz leads to bearing wear (confidence 92 points)” → entities: “Vibration frequency > 50 Hz”, “Bearing wear”; relationship: “leads to”; The decision confidence is then used as the base score, and the feature importance of the updated model parameters in the edge-cloud learning results is quantified as a correction score. The triple confidence is calculated using the base score and the correction score. The correction score can be adjusted according to actual conditions. For example, if the vibration feature weight in the model parameters is increased by 15%, the confidence level will increase by 5%, and the correction score will be 0.05. The final confidence level = basic score × (1 + correction score) (for example, 0.92 × 1.05 = 0.966); Use entity linking tools (such as the BERT-based entity matching model) to align new entities with existing entities in the metadata graph. If synonymous entities exist, they are uniformly mapped to standard entity names. If the entity is newly added, the existing entity in the metadata graph is used as the standard annotation to add the new entity type. Integrate new entities, relations, and confidence, decompose new entities and relations into subject, predicate, and object, and form structured triples including subject, predicate, object, and confidence; Write structured triples into the metadata graph, compress redundant relationships through knowledge distillation, and iteratively update the metadata graph; Specifically, knowledge distillation compresses redundant relations by calling a rule-based redundancy detection algorithm, defining the rules of the redundancy detection algorithm including similarity determination and merging rules; Similarity determination: When the subject and object entity similarity of two triples is ≥0.85 and the relationship type is consistent (for example, "lead to" and "initiate" are considered synonymous), they are determined to be redundant; Merging rule: retain the triple with the highest confidence, perform weighted fusion on the confidences of other similar triplets, and then add the fusion result to the confidence of the retained triple to obtain the final confidence of the merged triple. The updated content of the metadata graph after iteration is converted into specific rules and fed back into the construction process of the dynamic metadata network and the generation of multimodal fusion features, forming a closed-loop optimization; The method of feeding back to the dynamic metadata network construction process is: converting the updated content into a domain ontology extension package, including a new entity list (the entity list includes existing entities, new entities, and the type definitions attached to the new entities), relationship type extension (such as adding subtypes "strong cause" and "weak cause" under the "cause" relationship (strong cause and weak cause are divided according to confidence through the threshold division method)) and updating the semantic parsing rules (updating the LLM prompt word template (such as "prioritize the identification of entity pairs with 'strong cause' relationship")); The method for feeding back to the generation of multimodal fusion features is as follows: the updated content is converted into fusion weight adjustment rules in the format of (entity pair, associated weight correction value): high-confidence triplets and low-confidence triplets are divided by threshold segmentation method. For high-confidence triplets, the fusion weight of the corresponding feature in the multimodal fusion feature is increased (for example, the weight of vibration features in bearing fault fusion is increased by 15%); for low-confidence triplets, the fusion weight is reduced. The fusion weight adjustment rule is loaded into the cross-attention mechanism to dynamically adjust the correlation strength between modalities (for example, the attention weights of vibration features and fault features are updated in real time with the triple confidence). Example 2

[0031] See also Figure 3 As shown, for the parts not described in detail in this embodiment, please refer to the description of Example 1. A multi-source heterogeneous information fusion and analysis system for AI large models and data weaving is provided, including: Intelligent weaving and space-time alignment unit: This unit accesses multi-source heterogeneous data through data weaving, calibrates space-time using key events as anchors, and generates space-time aligned data. It also simultaneously constructs a dynamic metadata network, combines LLM to expand domain ontology, resolves semantic conflicts, and generates a metadata graph. Cross-modal semantic fusion unit: Builds a dynamic hypergraph based on spatiotemporal alignment data and metadata graphs; simultaneously calls modality-specific coding clusters to process various types of data and generate multimodal fusion features; Real-time decision-making and cognitive arbitration unit: Based on multimodal fusion features and dynamic hypergraphs, it initiates edge-cloud collaborative reasoning and generates intermediate results. Meanwhile, a general large model outputs confidence data to arbitrate cross-domain conflicts and generate the final decision package. Dynamic Optimization and Bidirectional Feedback Unit: Based on the final decision package, a bidirectional feedback closed-loop optimization is triggered. The AI ​​big model drives data source optimization and metadata graph knowledge completion, dynamically scheduling data flow priorities and encrypting and updating big model parameters. Knowledge accumulation and capability evolution unit: Based on the edge-cloud learning results and the final decision package, it iterates the metadata graph and feeds back to the construction of the dynamic metadata network and the generation of multimodal fusion features. Example 3

[0032] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the multi-source heterogeneous information fusion and analysis method of the AI ​​big model and data weaving provided above is realized.

[0033] Since the electronic device introduced in this embodiment is an electronic device used to implement the multi-source heterogeneous information fusion and analysis method of the AI ​​big model and data weaving in the embodiment of this application, based on the multi-source heterogeneous information fusion and analysis method of the AI ​​big model and data weaving introduced in the embodiment of this application, technical personnel in this field can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application will not be introduced in detail here. As long as technical personnel in this field implement the electronic device used by the multi-source heterogeneous information fusion and analysis method of the AI ​​big model and data weaving in the embodiment of this application, they all fall within the scope of protection of this application.

[0034] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field according to actual conditions.

[0035] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that for users of ordinary skill in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A multi-source heterogeneous information fusion and analysis method based on AI big models and data weaving, characterized by: include: S1: Access multi-source heterogeneous data through data weaving, calibrate time and space with key events as anchor points, and generate time-space aligned data packets; Simultaneously build a dynamic metadata network, combine LLM to expand the domain ontology and resolve semantic conflicts to generate a metadata graph; S2: Build a dynamic hypergraph based on spatiotemporal alignment data and metadata graphs; simultaneously call modality-specific coding clusters to process various types of data and generate multimodal fusion features; S3: Based on multimodal fusion features and a dynamic hypergraph, edge-cloud collaborative reasoning is initiated. Through edge reasoning and cloud-based AI large-scale model reasoning, intermediate collaborative reasoning results are generated. Simultaneously, the general large-scale model outputs confidence data to arbitrate cross-domain conflicts and generate the final decision package. S4: Based on the final decision package, a two-way feedback closed-loop optimization is triggered. The cloud-based AI big model drives data source optimization and metadata graph knowledge completion, dynamically scheduling data flow priorities and encrypting and updating cloud-based big model parameters. S5: Based on the edge-cloud learning results and the final decision package, the metadata graph is iterated and fed back to the construction of the dynamic metadata network and the generation of multimodal fusion features.

2. The multi-source heterogeneous information fusion and analysis method for AI big model and data weaving according to claim 1 is characterized in that: The method of spatiotemporally aligning data packets includes: Perform data classification and interface adaptation on the accessed multi-source heterogeneous data, define global data identification rules, generate a globally unique ID for each accessed data, and then build a metadata registry; Based on the metadata registry, time and space calibration is performed on multi-source heterogeneous data with key events as anchor points; Time calibration is as follows: by defining anchor trigger conditions to identify key events in the metadata registry as anchor points, and by calculating the time offset based on the timestamp of the anchor point, the original timestamp of the non-anchor point data is corrected; Spatial calibration involves defining a global target coordinate system, extracting physical device spatial metadata from multi-source heterogeneous data, and then uniformly converting the original coordinates of each physical device into the absolute coordinates of the global target coordinate system. Integrate all the data after spatiotemporal calibration and encapsulate them into a spatiotemporal alignment data package.

3. The multi-source heterogeneous information fusion and analysis method for AI big model and data weaving according to claim 2 is characterized in that: The metadata graph is generated in the following manner: Extract technical metadata and business metadata based on spatiotemporal alignment of data packets to form an initialized dynamic metadata network; Preset the domain ontology based on business scenarios and define the concepts and relationships of the domain ontology; Call LLM to perform semantic parsing on the unstructured data in the spatiotemporal alignment data package to obtain entities, match the parsed entities with concepts, associate them with domain ontology, and then add the domain ontology to the dynamic metadata network; Detect cross-source semantic conflicts between entities through a dynamic metadata network and mark them as conflicting entities; Then, dynamic business weights are assigned to conflicting entities, and conflict resolution logic is defined to form conflict resolution rules; Entities are regarded as nodes, relationships as edges, and conflict resolution rules are associated with the corresponding nodes of conflicting entities as attribute constraints to construct a metadata graph.

4. The multi-source heterogeneous information fusion and analysis method for AI big model and data weaving according to claim 3 is characterized by: The dynamic hypergraph is constructed in the following manner: Based on the spatiotemporal alignment of data packets and metadata graphs, we extract multimodal entities, assign hypergraph vertices to each entity, and add vertex attributes. Based on the entity relationships in the metadata graph, the hyperedge association rules are defined as follows: each hyperedge is associated with at least two or more hypergraph vertices; Then, add hyperedges to the hypergraph vertices according to the hyperedge association rules and label the corresponding attributes for the hyperedges; The hyperedge structure is dynamically optimized through predefined dynamic adjustment rules, and a dynamic hypergraph is generated based on the hypergraph vertices and hyperedges.

5. The multi-source heterogeneous information fusion and analysis method for AI big model and data weaving according to claim 4 is characterized in that: The generation method of the multimodal fusion feature includes: Based on a dynamic hypergraph, the multimodal data to be processed is extracted by associating the hypergraph vertices with the spatiotemporal aligned data packets; According to the hypergraph vertex attributes, multimodal data is assigned to corresponding processing channels, and modality-specific coding clusters are called to perform feature extraction on different types of data to obtain dimensional features of the corresponding modality; The dimensional features of different modalities are processed in a unified dimension and fused to form multimodal fusion features.

6. The multi-source heterogeneous information fusion and analysis method for AI big model and data weaving according to claim 5 is characterized in that: The generation method of the collaborative reasoning intermediate result includes: Based on multimodal fusion features and dynamic hypergraph, fusion features are obtained through hyperedge confidence screening, and low-complexity association rules and high-complexity association rules are extracted from the dynamic hypergraph; Edge-cloud collaborative reasoning is defined as including edge reasoning and cloud reasoning; The edge is responsible for inferring and generating the intermediate feature vector for preliminary judgment through the filtered fusion features and low-complexity association rules; The cloud is responsible for combining multimodal fusion features, highly complex association rules, and intermediate feature vectors, performing deep reasoning through the cloud-based AI large model, and generating fine-grained decision recommendations; Integrate intermediate feature vectors and fine-grained decision recommendations to generate collaborative reasoning intermediate results containing reasoning basis.

7. The multi-source heterogeneous information fusion and analysis method for AI big model and data weaving according to claim 6 is characterized in that: The final decision package is generated in the following manner: The cloud-based AI model generates preliminary decisions based on the intermediate results of collaborative reasoning, integrating intermediate feature vectors, reasoning paths, and association rules. Call the general large model to perform cross-domain verification on the preliminary decision, output the decision confidence and conflict analysis report, and then perform hierarchical arbitration: If the decision confidence meets expectations and there is no conflict, the preliminary decision is directly adopted, supplemented by the verification conclusion of the general large model, to form the final decision conclusion; If the decision confidence does not meet expectations or there is a conflict, a secondary reasoning correction is performed based on the conflict analysis to re-form the final decision conclusion; Integrate decision conclusions and decision confidence to generate the final decision package.

8. The multi-source heterogeneous information fusion and analysis method for AI big model and data weaving according to claim 7 is characterized in that: The two-way feedback closed-loop optimization method includes: data quality detection optimization and metadata graph integrity detection optimization; Data quality detection and optimization: Based on the final decision package, the decision conclusion is associated with the data source, and data quality indicators are calculated by type. For data sources that do not meet the data quality indicators, an alternative data source is matched from the metadata map and switched. Optimize metadata graph integrity detection: extract new entities from decision conclusions, perform semantic matching with existing entities in the metadata graph through LLM, and obtain entity matching degree; If the entity matching degree does not meet expectations, it is determined that the graph is missing an entity. The associated features of the new entity are mined from the historical spatiotemporal alignment data package and added to the metadata graph.

9. The multi-source heterogeneous information fusion and analysis method for AI big model and data weaving according to claim 8 is characterized in that: The method of dynamically scheduling data stream priorities and encrypting and updating cloud-based AI model parameters includes: Based on the results of the two-way feedback closed-loop optimization, the actual bandwidth utilization of each data source is quantified, and task processing priorities are defined to divide all data into corresponding data streams; Then dynamically adjust the bandwidth allocation of all data flows according to the task processing priority; At the same time, based on the optimization results of metadata graph integrity detection, through the edge-cloud learning process, local training and gradient encryption are performed on the edge, and gradient aggregation and model updates are performed on the cloud; Then, the edge-cloud learning results are integrated to update the reasoning process of edge-cloud collaborative reasoning.

10. The multi-source heterogeneous information fusion and analysis method for AI big model and data weaving according to claim 9 is characterized in that: The method of iterating the metadata graph includes: Extract structured triples from edge-cloud learning results and final decision packages, write them into the metadata graph, compress redundant relationships through knowledge distillation, and iteratively update the metadata graph; The updated content of the metadata graph after iteration is converted into specific rules and fed back to the construction process of the dynamic metadata network and the generation of multimodal fusion features, forming a closed-loop optimization.

Images