Multi-source heterogeneous data fusion method and system

Abstract

The invention relates to the technical field of data processing and information fusion, and provides a multi-source heterogeneous data fusion method and system, and the method comprises the steps: carrying out the multi-dimensional quality evaluation of at least three heterogeneous data sources; on the basis of a quality evaluation result, feature decoupling and cross-modal correlation analysis of the heterogeneous data source are executed, and an intermediate feature set with a space-time alignment characteristic is generated; constructing a nonlinear dynamic weight distribution model, and calculating a multi-dimensional credibility weight of each data source; performing fusion calculation on the intermediate feature set through a three-level joint optimization architecture; and dynamically correcting the fusion result by adopting a feedback type self-adaptive calibration mechanism, and outputting an optimized fusion data cube. According to the invention, the accuracy, consistency and reliability of fused data can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing and information fusion, and more specifically, to a multi-source heterogeneous data fusion method and system. Background Art

[0002] In today's digital age, data sources are increasingly diverse, including structured databases, unstructured text streams, real-time sensor networks, and image acquisition devices. Each of these data sources has unique data formats and characteristics. For example, structured data is stored in tabular form, making it easy to query and analyze; unstructured text streams contain large amounts of natural language information, requiring complex text processing techniques to extract useful information; real-time sensor networks can provide continuous environmental monitoring data, but the data may be noisy and uncertain; and image acquisition devices generate visual information, requiring image processing and analysis to extract features. Existing data fusion techniques primarily focus on the fusion of a single modality or a few modalities, such as simply correlating sensor data with image data or performing basic semantic analysis on text data. While these techniques demonstrate some effectiveness when processing single or a few data sources, they often suffer from significant limitations when dealing with multi-source, heterogeneous data. For example, single-modality fusion techniques fail to fully leverage the complementary information between different modalities, limiting the accuracy and completeness of the fusion results. Furthermore, most existing fusion methods lack a comprehensive assessment of data quality and are unable to effectively address issues such as spatiotemporal inconsistency, semantic differences, and data freshness across data sources.

[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the existing technology: the existing technology cannot effectively handle the complex correlation relationships between multi-source heterogeneous data, and it is difficult to generate fusion results with spatiotemporal alignment and semantic consistency; there is a lack of comprehensive evaluation of the multi-dimensional quality of data sources, resulting in the inability to accurately assign the credibility weights of each data source during the fusion process; the fusion process lacks a dynamic correction mechanism and cannot adapt to the dynamic changes and uncertainties of the data source, thereby affecting the accuracy and reliability of the fusion results. Summary of the Invention

[0004] The present invention provides a multi-source heterogeneous data fusion method and system.

[0005] In a first aspect of the present invention, a multi-source heterogeneous data fusion method is provided, comprising:

[0006] S1. Perform multi-dimensional quality assessment on at least three heterogeneous data sources, including structured databases, unstructured text streams, real-time sensor networks, and image acquisition devices;

[0007] S2. Based on the quality assessment results, perform feature decoupling and cross-modal correlation analysis of heterogeneous data sources to generate an intermediate feature set with spatiotemporal alignment characteristics;

[0008] S3. Construct a nonlinear dynamic weight distribution model to calculate the multi-dimensional credibility weight of each data source, including spatiotemporal consistency, semantic entropy, and data freshness;

[0009] S4. Fusion computation of intermediate feature sets is performed through a three-level joint optimization architecture, where the first level uses feature-level bidirectional interaction, the second level performs spatiotemporal joint modeling, and the third level enforces semantic consistency constraints.

[0010] S5. Use the feedback adaptive calibration mechanism to dynamically correct the fusion results and output the optimized fusion data cube.

[0011] Furthermore, the dynamic weight allocation model of step S3 includes the following operations:

[0012] S31. Calculate the spatiotemporal consistency weight factor of the data source:

[0013]

[0014] Where W st is the spatiotemporal consistency weight factor, ΔT is the data timestamp difference (unit: milliseconds), L i , L j are the spatial coordinates (three-dimensional vectors) of data sources i and j, θ1 and θ2 are the learnable parameter matrices;

[0015] S32. Evaluate semantic entropy weighting factor:

[0016]

[0017] Where W se is the semantic entropy weight factor, N unique is the number of unique semantic entities, p k is the actual semantic distribution probability, is the theoretical benchmark distribution, M is the total number of semantic categories;

[0018] S33. Calculate the data freshness decay coefficient:

[0019]

[0020] Where W fresh is the freshness attenuation coefficient, λ is the attenuation rate parameter, t current is the current system time, t gen is the data generation time, t max is the preset maximum effective time window (unit: seconds), α is the trend gain coefficient, Strend is the slope of the data change trend;

[0021] S34. Generate the final weight through nonlinear combination:

[0022]

[0023] Where W final is the final weight vector, represents the element-wise product of tensors, Represents a feature concatenation operation.

[0024] Furthermore, the step S2 includes:

[0025] S21. Decouple multiple tables from the structured database to generate independent subject feature subsets while preserving foreign key constraints.

[0026] S22. Perform context-aware segmented vectorization on unstructured text streams, using a sliding window mechanism to extract local semantic features, with the window size dynamically adjusted based on the text entropy.

[0027] S23. Perform spatiotemporal joint noise reduction on real-time sensor data, using a bidirectional LSTM network to predict the normal value interval and perform nonlinear interpolation on data points outside the interval;

[0028] S24. Implement cross-modal feature anchoring on image data, align the visual feature vector with the text description vector, and generate a joint embedding representation.

[0029] Furthermore, the three-level joint optimization architecture of step S4 includes:

[0030] S41. In the first level feature fusion, a bidirectional gated interaction mechanism is adopted, and its calculation formula is:

[0031] F gate =σ(W g ·[F i ,F j ])⊙tanh(W f ·[F i ,F j ])

[0032] Where, F gate is the gated fusion feature, F i 、F j are different source feature vectors (dimension ≥ 256), W g 、W f is the trainable parameter matrix (dimension: N×2N), σ is the sigmoid function, and ⊙ represents the Hadamard product;

[0033] In the second-level spatiotemporal modeling, a spatiotemporal hypergraph network is constructed, where nodes represent feature vectors and hyperedges connect sets of nodes with spatiotemporal associations. The feature representation is updated through hypergraph convolution.

[0034] S43. In the third-level semantic constraint, knowledge graph embedding is introduced as a regularization term to minimize the KL divergence between the fusion result and the prior semantic relationship.

[0035] Furthermore, the feedback calibration mechanism of step S5 includes:

[0036] S51. Construct a dynamic residual learning model, whose residual calculation formula is:

[0037]

[0038] Where R t is the residual vector at the current moment, E t-1 is the error vector of the previous moment (dimension: 128), is the current feature gradient matrix, Q(F t ) is the query vector, K(H hist ) is the history state key vector, × represents matrix multiplication;

[0039] S52. Implement multi-scale calibration to calibrate the fusion results at the macro, micro, and semantic levels simultaneously;

[0040] Macro-level calibration includes: adjusting feature distributions to match global statistics;

[0041] Micro-level calibration includes: correcting the deviation of individual data points from the local cluster center;

[0042] Semantic level calibration includes: constraining the consistency of output with the domain knowledge graph.

[0043] Furthermore, the bidirectional LSTM network in step S23 adopts an asymmetric structure:

[0044] The number of layers in the forward network is greater than that in the reverse network, and the hidden layer dimension satisfies:

[0045]

[0046] Where, d forward is the hidden layer dimension of the feedforward network, d backward is the hidden layer dimension of the reverse network, N sensor is the total number of sensor nodes;

[0047] The network output passes through the threshold filter function:

[0048]

[0049] Where, is the corrected sensor data, P(x t ) is the confidence probability of the data point, τ is the dynamic threshold (range: 0.6-0.9), μ t|t-1 is the predicted mean, σ t|t-1 is the predicted standard deviation, ∈ is a random noise vector (consistent with N(0,1) distribution).

[0050] Furthermore, the nonlinear combination operation in step S34 is implemented by a differentiable decision tree:

[0051] The tree node splitting condition is determined by the statistical characteristics of the weight factor. The leaf node outputs the combined weight, and the decision path satisfies:

[0052]

[0053] Where Path is the decision path function, f k is the characteristic function designed based on the information gain ratio (k=1,...,K), γ k is the dynamic splitting threshold, and I is the indicator function.

[0054] Furthermore, the hypergraph convolution operation adopts an anisotropic diffusion strategy:

[0055] The node feature update formula is:

[0056]

[0057] Where, F (l+1) is the updated feature matrix, F (l) is the current layer feature, w(e) is the hyperedge weight (range: 0-1), |e| is the number of nodes contained in the hyperedge, MLP is a multi-layer perceptron (including 3 hidden layers), A e is the adjacency matrix block (dimension: n×n).

[0058] Furthermore, the quality assessment in step S1 includes:

[0059] S11. Calculate protocol compatibility score: Verify whether the data interface complies with the ISO / IEC 21878 standard extension protocol;

[0060] S12. Evaluate the data completeness index: Calculate the topological impact of missing fields using the following formula:

[0061]

[0062] Where C comp is the completeness index (range: 0-1), ρ i is the network centrality (PageRank value) of field i, δ iis the missing state indicator (δ = 1 for missing, δ = 0 for present);

[0063] S13. Measure the timing continuity index: Calculate the inverse of the product of the jitter coefficient and the retransmission rate of the valid data packet.

[0064] In a second aspect of the present invention, a multi-source heterogeneous data fusion system is provided, comprising:

[0065] A quality assessment module, configured to perform multi-dimensional quality assessment on at least three heterogeneous data sources, including a structured database, an unstructured text stream, a real-time sensor network, and an image acquisition device;

[0066] A correlation analysis module, configured to perform feature decoupling and cross-modal correlation analysis of heterogeneous data sources based on the quality assessment results, and generate an intermediate feature set with spatiotemporal alignment characteristics;

[0067] A weight assignment module is used to calculate the multi-dimensional credibility weight of each data source, including spatiotemporal consistency, semantic entropy, and data freshness;

[0068] A fusion computing module is used to perform fusion computing on the intermediate feature set through a three-level joint optimization architecture, where the first level uses feature-level bidirectional interaction, the second level performs spatiotemporal joint modeling, and the third level implements semantic consistency constraints;

[0069] The dynamic correction module is used to dynamically correct the fusion results using a feedback adaptive calibration mechanism and output the optimized fusion data cube.

[0070] According to the above-mentioned embodiments of the present invention, there are at least the following beneficial effects: the multi-source heterogeneous data fusion method and system of the present invention can effectively improve the accuracy and reliability of data fusion. Through multi-dimensional quality assessment, it is possible to comprehensively analyze the spatiotemporal consistency, semantic entropy and data freshness of each data source, thereby providing an accurate basis for weight distribution for subsequent fusion calculations. In the feature decoupling and cross-modal association analysis stage, it is possible to effectively extract and align features of different types of data sources, generate an intermediate feature set with spatiotemporal alignment characteristics, and lay the foundation for high-quality data fusion. The three-level joint optimization architecture further enhances the fusion effect. The first-level feature-level two-way interaction can make full use of the complementary information between different data sources. The second-level spatiotemporal joint modeling can effectively process the spatiotemporal correlation of data. The third-level semantic consistency constraint ensures the consistency of the fusion result with the prior semantic knowledge, thereby comprehensively improving the quality of the fused data.

[0071] Furthermore, the feedback-based adaptive calibration mechanism employed in this invention dynamically corrects the fusion results, adapting them to the dynamic changes and uncertainties of the data source. This dynamic correction mechanism allows for real-time adjustment of the fusion strategy, ensuring the accuracy and reliability of the fused data in different scenarios. Furthermore, this method generates an optimized fused data cube, providing a high-quality data foundation for subsequent data analysis and decision support. This method is particularly suitable for multi-source heterogeneous data fusion tasks in complex scenarios, and has broad application prospects and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which:

[0073] Figure 1 A flowchart of a multi-source heterogeneous data fusion method provided by one embodiment of the present invention;

[0074] Figure 2 A schematic diagram of the structure of a multi-source heterogeneous data fusion system provided by one embodiment of the present invention;

[0075] Figure 3 The figure schematically shows the structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0076] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0077] Those skilled in the art will appreciate that the embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0078] It should be noted that any number of elements in the drawings is for illustration only and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0079] Reference below Figure 1 , Figure 1 The following is a flow chart of a multi-source heterogeneous data fusion method provided by an embodiment of the present invention. Figure 1As shown, a multi-source heterogeneous data fusion method 100 includes:

[0080] S1. Perform multi-dimensional quality assessment on at least three heterogeneous data sources, including structured databases, unstructured text streams, real-time sensor networks, and image acquisition devices;

[0081] S2. Based on the quality assessment results, perform feature decoupling and cross-modal correlation analysis of heterogeneous data sources to generate an intermediate feature set with spatiotemporal alignment characteristics;

[0082] S3. Construct a nonlinear dynamic weight distribution model to calculate the multi-dimensional credibility weight of each data source, including spatiotemporal consistency, semantic entropy, and data freshness;

[0083] S4. Fusion computation of intermediate feature sets is performed through a three-level joint optimization architecture, where the first level uses feature-level bidirectional interaction, the second level performs spatiotemporal joint modeling, and the third level enforces semantic consistency constraints.

[0084] S5. Use the feedback adaptive calibration mechanism to dynamically correct the fusion results and output the optimized fusion data cube.

[0085] It should be noted that the multi-source heterogeneous data fusion method of the present invention first performs a multi-dimensional quality assessment on at least three heterogeneous data sources. Heterogeneous data sources here refer to data sources with different data formats, origins, or characteristics, such as structured databases, unstructured text streams, real-time sensor networks, and image acquisition devices. Multi-dimensional quality assessment refers to a quantitative analysis of the quality of data sources from multiple aspects, including but not limited to the temporal and spatial consistency of the data, semantic entropy, and data freshness. This assessment can provide an accurate basis for weight allocation for subsequent fusion calculations, ensuring that the fusion process can fully utilize high-quality data source information.

[0086] Specifically, multi-dimensional quality assessment involves quantitative analysis of multiple key characteristics of different data sources. For example, spatiotemporal consistency assessment can measure the degree of temporal and spatial alignment of different data sources by calculating the differences in data timestamps and spatial coordinates. Semantic entropy assessment measures the semantic richness and uncertainty of data by analyzing the distribution of semantic information in the data. Data freshness assessment measures the timeliness of data by calculating the difference between the data generation time and the current system time. The specific parameters of these evaluation indicators can be adjusted according to the application scenario. For example, in an environmental monitoring scenario, the weight of spatiotemporal consistency may be higher, while in a news analysis scenario, the weight of semantic entropy may be more important. In addition, quality assessment can also include protocol compatibility verification of data interfaces to ensure that data sources comply with specific standard extension protocols, such as the ISO / IEC21878 standard.

[0087] Preferably, when conducting multi-dimensional quality assessment, a dynamic parameter adjustment mechanism can be introduced to dynamically adjust the weights of the evaluation indicators according to the actual characteristics of the data source. For example, for application scenarios with high real-time requirements, the weight of data freshness can be increased, while for application scenarios that require high-precision spatiotemporal alignment, the weight of spatiotemporal consistency can be increased. In addition, a machine learning algorithm can be introduced to automatically learn the optimal weights of evaluation indicators through historical data, thereby further improving the accuracy of quality assessment. During the evaluation process, the network topology of the data source can also be considered, for example, by calculating the network centrality of the field (such as the PageRank value) to measure its importance in the network, thereby more accurately evaluating the completeness of the data.

[0088] In some embodiments, the dynamic weight allocation model of step S3 includes the following operations:

[0089] S31. Calculate the spatiotemporal consistency weight factor of the data source:

[0090]

[0091] Where W st is the spatiotemporal consistency weight factor, ΔT is the data timestamp difference (unit: milliseconds), L i , L j are the spatial coordinates (three-dimensional vectors) of data sources i and j, θ1 and θ2 are the learnable parameter matrices;

[0092] S32. Evaluate semantic entropy weighting factor:

[0093]

[0094] Where W se is the semantic entropy weight factor, N unique is the number of unique semantic entities, p k is the actual semantic distribution probability, is the theoretical benchmark distribution, M is the total number of semantic categories;

[0095] S33. Calculate the data freshness decay coefficient:

[0096]

[0097] Where W fresh is the freshness attenuation coefficient, λ is the attenuation rate parameter, t current is the current system time, t gen is the data generation time, t max is the preset maximum effective time window (unit: seconds), α is the trend gain coefficient, S trend is the slope of the data change trend;

[0098] S34. Generate the final weight through nonlinear combination:

[0099]

[0100] Where W final is the final weight vector, represents the element-wise product of tensors, Represents a feature concatenation operation.

[0101] It should be noted that the dynamic weight assignment model is a key component of this invention, used to calculate the multi-dimensional credibility weights of each data source. These dimensions include spatiotemporal consistency, semantic entropy, and data freshness. The spatiotemporal consistency weight factor measures the degree of alignment of data sources in time and space, the semantic entropy weight factor reflects the semantic richness and uncertainty of the data, and the data freshness decay coefficient assesses the timeliness of the data. Generating the final weights through nonlinear combination ensures that the contribution of each data source during the fusion process matches its quality and relevance.

[0102] Specifically, the spatiotemporal consistency weight factor is calculated based on the differences in data timestamps and spatial coordinates. Timestamp differences measure the temporal synchronization of data, while spatial coordinate differences measure the spatial alignment of data. The semantic entropy weight factor measures the semantic richness of the data by analyzing the number and distribution of semantic entities in the data. The data freshness decay coefficient measures the timeliness of the data by calculating the difference between the data generation time and the current system time. The specific parameters of these weight factors can be adjusted according to the application scenario. For example, in environmental monitoring, spatiotemporal consistency may be given a higher weight, while in news analysis, semantic entropy may be given a higher weight. In addition, the dynamic weight allocation model can automatically learn these parameters through machine learning algorithms to adapt to different application scenarios.

[0103] Preferably, when calculating the spatiotemporal consistency weight factor, a learnable parameter matrix can be introduced. These parameters can be automatically adjusted through training data to better adapt to the characteristics of different data sources. When evaluating the semantic entropy weight factor, natural language processing technology can be used to extract semantic entities and calculate their distribution probabilities. For the data freshness decay coefficient, a trend gain coefficient can be introduced to take into account the changing trend of the data, thereby more accurately evaluating the timeliness of the data. In addition, nonlinear combination operations can be implemented through differentiable decision trees. The node splitting conditions of the decision tree are determined by the statistical characteristics of the weight factors, and the leaf nodes output the combined weights, thereby achieving flexible weight distribution.

[0104] In some embodiments, step S2 includes:

[0105] S21. Decouple multiple tables from the structured database to generate independent subject feature subsets while preserving foreign key constraints.

[0106] S22. Perform context-aware segmented vectorization on unstructured text streams, using a sliding window mechanism to extract local semantic features, with the window size dynamically adjusted based on the text entropy.

[0107] S23. Perform spatiotemporal joint noise reduction on real-time sensor data, using a bidirectional LSTM network to predict the normal value interval and perform nonlinear interpolation on data points outside the interval;

[0108] S24. Implement cross-modal feature anchoring on image data, align the visual feature vector with the text description vector, and generate a joint embedding representation.

[0109] It should be noted that the feature decoupling and cross-modal association analysis in the present invention are key steps to achieve multi-source heterogeneous data fusion. This process involves extracting and aligning features from different types of data sources to generate an intermediate feature set with spatiotemporal alignment characteristics. Specifically, the feature decoupling of structured databases aims to generate independent topic feature subsets while retaining foreign key constraint relationships; the processing of unstructured text streams extracts local semantic features through context-aware segmented vectorization; the spatiotemporal joint denoising of real-time sensor data ensures data accuracy and consistency; and the cross-modal feature anchoring of image data generates a joint embedding representation by aligning visual features with text descriptions. These operations provide a high-quality feature foundation for subsequent data fusion.

[0110] Specifically, for the decoupling of multi-table associations in structured databases, database query languages (such as SQL) can be used to generate independent subsets of topic features by analyzing the foreign key relationships between tables. Context-aware segmented vectorization of unstructured text streams can be achieved through natural language processing technology, such as using a sliding window mechanism to extract local semantic features, where the window size can be dynamically adjusted based on the complexity of the text. Spatiotemporal joint denoising of real-time sensor data can be achieved through a bidirectional long short-term memory (Bi-LSTM) network, which can predict the normal value interval and perform nonlinear interpolation replacement on data points outside the interval. Cross-modal feature anchoring of image data can be achieved through an attention mechanism, aligning the visual feature vector with the text description vector to generate a joint embedding representation. The combination of these technologies ensures effective association and alignment between data of different modalities.

[0111] Preferably, when processing unstructured text streams, the size of the sliding window can be dynamically adjusted according to the text entropy to better adapt to the semantic complexity of different texts. For example, in areas with more complex text content or a large amount of information, the window size can be appropriately increased to extract richer semantic features. For the denoising of real-time sensor data, the structure of the bidirectional LSTM network can be further optimized, such as adopting an asymmetric structure in which the number of forward network layers is greater than that of the reverse network, so as to better capture the time series characteristics of the data. In addition, cross-modal feature anchoring of image data can be achieved through a multimodal learning framework, such as using a pre-trained visual-language model to enhance the effect of feature alignment. These optimization measures can further improve the performance of feature decoupling and cross-modal association analysis, providing a more solid foundation for high-quality data fusion.

[0112] In some embodiments, the three-level joint optimization architecture of step S4 includes:

[0113] S41. In the first level feature fusion, a bidirectional gated interaction mechanism is adopted, and its calculation formula is:

[0114] F gate =σ(W g ·[F i ,F j ])⊙tanh(W f ·[F i ,F j ])

[0115] Where, F gate is the gated fusion feature, F i 、F j are different source feature vectors (dimension ≥ 256), W g 、W f is the trainable parameter matrix (dimension: N×2N), σ is the sigmoid function, and ⊙ represents the Hadamard product;

[0116] In the second-level spatiotemporal modeling, a spatiotemporal hypergraph network is constructed, where nodes represent feature vectors and hyperedges connect sets of nodes with spatiotemporal associations. The feature representation is updated through hypergraph convolution.

[0117] S43. In the third-level semantic constraint, knowledge graph embedding is introduced as a regularization term to minimize the KL divergence between the fusion result and the prior semantic relationship.

[0118] It should be noted that the three-level joint optimization architecture of the present invention is the core link of fusion computing, aiming to achieve high-quality data fusion through a multi-level optimization strategy. The first level adopts feature-level bidirectional interaction, fusing feature vectors from different sources through a gating mechanism; the second level performs spatiotemporal joint modeling, using a spatiotemporal hypergraph network to capture the spatiotemporal correlations between features; and the third level implements semantic consistency constraints, ensuring the consistency of the fusion results with prior semantic knowledge through knowledge graph embedding. This hierarchical optimization strategy can fully utilize the characteristics of each data source while ensuring the accuracy and consistency of the fusion results.

[0119] Specifically, the bidirectional gated interaction mechanism in the first-level feature fusion achieves dynamic feature fusion by calculating the weighted sum of feature vectors from different sources and applying an activation function. The dimension of the trainable parameter matrix can be set according to the size of the feature vector. For example, when the feature vector dimension is 256, the parameter matrix dimension can be set to N×2N. In the second-level spatiotemporal modeling, the spatiotemporal hypergraph network connects a set of nodes with spatiotemporal associations through hyperedges and updates the feature representation through hypergraph convolution. The hyperedge weights can be dynamically adjusted based on the spatiotemporal distance between nodes, typically ranging from 0 to 1. In the third-level semantic constraint, the knowledge graph embedding acts as a regularization term to ensure the semantic consistency of the fusion result by minimizing the KL divergence between the fusion result and the prior semantic relationship. The settings of these parameters and mechanisms can be optimized according to the specific application scenario.

[0120] Preferably, in the first-level feature fusion, an attention mechanism can be introduced to dynamically adjust the gating weights according to the importance of the feature vector, thereby further improving the fusion effect. In the second-level spatiotemporal modeling, the hypergraph convolution operation can adopt an anisotropic diffusion strategy to update the node features through a multi-layer perceptron (MLP) and an adjacency matrix block to better capture spatiotemporal associations. In the third-level semantic constraint, a pre-trained language model (such as BERT) can be used to generate knowledge graph embeddings to enhance the effect of semantic consistency constraints. In addition, a multi-scale calibration mechanism can be introduced to adjust the feature distribution to match the global statistics at the macro level, correct the deviation of individual data points from the local cluster center at the micro level, and constrain the consistency of the output with the domain knowledge graph at the semantic level. These optimization measures can further improve the performance of the three-level joint optimization architecture and ensure the high quality of the fusion results.

[0121] In some embodiments, the feedback calibration mechanism of step S5 includes:

[0122] S51. Construct a dynamic residual learning model, whose residual calculation formula is:

[0123]

[0124] Where R t is the residual vector at the current moment, Et-1 is the error vector of the previous moment (dimension: 128), is the current feature gradient matrix, Q(F t ) is the query vector, K(H hist ) is the history state key vector, × represents matrix multiplication;

[0125] S52. Implement multi-scale calibration to calibrate the fusion results at the macro, micro, and semantic levels simultaneously;

[0126] Macro-level calibration includes: adjusting feature distributions to match global statistics;

[0127] Micro-level calibration includes: correcting the deviation of individual data points from the local cluster center;

[0128] Semantic level calibration includes: constraining the consistency of output with the domain knowledge graph.

[0129] It should be noted that the feedback adaptive calibration mechanism of the present invention is a key link in the dynamic correction of fusion results, aiming to improve the accuracy and reliability of fusion data through dynamic residual learning and multi-scale calibration. The dynamic residual learning model calculates the residual vector at the current moment, combines the historical state and the current feature gradient, and makes real-time corrections to the fusion results. Multi-scale calibration adjusts at the macro, micro, and semantic levels to ensure the consistency and accuracy of the fusion results at different levels. This mechanism can effectively cope with the dynamic changes and uncertainties of the data source and further optimize the fusion results.

[0130] Specifically, the residual calculation of the dynamic residual learning model is based on the error vector at the previous moment, the current feature gradient matrix, the query vector, and the historical state key vector. These parameters can be optimized using training data. For example, the dimension of the error vector can be set to 128 to accommodate most application scenarios. The query vector and historical state key vector can be dynamically generated using an attention mechanism to better capture the correlation between features. In the multi-scale calibration mechanism, macro-level calibration can adjust the mean and variance of the feature distribution to match the global statistics; micro-level calibration can reduce bias by correcting the distance between individual data points and local cluster centers; and semantic-level calibration ensures semantic accuracy by constraining the output to be consistent with the domain knowledge graph. The settings of these parameters and mechanisms can be adjusted according to the specific application scenario.

[0131] Preferably, in the dynamic residual learning model, an adaptive learning rate mechanism can be introduced to dynamically adjust the learning rate according to the size of the current error, thereby accelerating the convergence of the model. In the multi-scale calibration mechanism, the micro-level calibration can be further refined, for example, by introducing local sensitive hashing (LSH) technology to quickly locate and correct deviated individual data points. In the semantic level calibration, domain-specific knowledge graphs can be used to enhance the effect of semantic constraints through graph neural networks (GNNs). In addition, a multi-model fusion mechanism can be introduced to combine the outputs of multiple residual learning models to further improve the accuracy and robustness of the calibration. These optimization measures can further improve the performance of the feedback adaptive calibration mechanism and ensure the high quality and reliability of the fusion results.

[0132] In some embodiments, the bidirectional LSTM network in step S23 adopts an asymmetric structure:

[0133] The number of layers in the forward network is greater than that in the reverse network, and the hidden layer dimension satisfies:

[0134]

[0135] Where, d forward is the hidden layer dimension of the feedforward network, d backward is the hidden layer dimension of the reverse network, N sensor is the total number of sensor nodes;

[0136] The network output passes through the threshold filter function:

[0137]

[0138] Where, is the corrected sensor data, P(x t ) is the confidence probability of the data point, τ is the dynamic threshold (range: 0.6-0.9), μ t|t-1 is the predicted mean, σ t|t-1 is the predicted standard deviation, ∈ is a random noise vector (consistent with N(0,1) distribution).

[0139] It should be noted that the bidirectional LSTM network mentioned in this invention adopts an asymmetric structure and is a technical means for processing real-time sensor data for spatiotemporal noise reduction. This structural design enables the network to better capture the time series characteristics of the data and more effectively identify and handle abnormal data points when processing sensor data. The asymmetric structure means that the number of forward network layers is greater than the number of reverse network layers, and the hidden layer dimension is adjusted according to the total number of sensor nodes. In addition, the network output is corrected by a threshold filter function to ensure the accuracy and reliability of the data.

[0140] Specifically, in the asymmetric design of the bidirectional LSTM network, the number of forward network layers is greater than the number of backward network layers. This is because the forward network is better able to capture the time series characteristics of the data, while the backward network is primarily used to process short-term data dependencies. The hidden layer dimension is adjusted based on the total number of sensor nodes. The specific formula is: the hidden layer dimension of the forward network equals the hidden layer dimension of the backward network plus the logarithm (base 2) of the total number of sensor nodes. This setting ensures that the network has sufficient capacity to process large amounts of sensor data. The threshold filter function at the network output determines whether to correct a data point based on the confidence probability of the data point. If the confidence probability exceeds a dynamic threshold, the data point is directly output; otherwise, the data point is corrected using the predicted mean plus a random noise vector. The dynamic threshold is typically set between 0.6 and 0.9 to accommodate different data characteristics and application scenarios.

[0141] To further enhance the performance of bidirectional LSTM networks, additional hidden layers can be introduced into the feedforward network to enhance the network's ability to capture long-term dependencies. Furthermore, more complex threshold filtering functions can be used, such as those combining multiple thresholds and different noise distributions, to better handle data uncertainty and outliers. Furthermore, an attention mechanism can be introduced to enable the network to focus more on important time steps and sensor nodes, thereby improving the accuracy of data denoising. In practical applications, the network structure and parameters can be dynamically adjusted based on the scale of the sensor network and the characteristics of the data to achieve optimal denoising results.

[0142] In some embodiments, the nonlinear combination operation in step S34 is implemented by a differentiable decision tree:

[0143] The tree node splitting condition is determined by the statistical characteristics of the weight factor. The leaf node outputs the combined weight, and the decision path satisfies:

[0144]

[0145] Where Path is the decision path function, f k is the characteristic function designed based on the information gain ratio (k=1,...,K), γ k is the dynamic splitting threshold, and I is the indicator function.

[0146] It should be noted that the nonlinear combination operation mentioned in the present invention is implemented through a differentiable decision tree, which is a technical means for dynamic weight allocation. The differentiable decision tree combines the hierarchical structure of the traditional decision tree with the differentiable characteristics of the neural network, and can automatically learn the optimal decision path and weight allocation during the training process. The node splitting condition of the decision tree is determined by the statistical characteristics of the weight factor, the leaf node outputs the combined weight, and the decision path is implemented by a combination of the indicator function and the characteristic function. This structure can flexibly adjust the weight allocation to adapt to the characteristics of different data sources.

[0147] Specifically, the node splitting conditions of a differentiable decision tree are designed based on the statistical characteristics of the weight factors, which can include the mean, variance, maximum, or minimum value of the weight factors. A feature function is designed based on the information gain ratio to assess the importance of each feature when splitting a node. A dynamic splitting threshold can be automatically adjusted based on training data to optimize the performance of the decision tree. An indicator function is used to determine whether a feature value meets the splitting conditions, thereby determining the flow of data within the decision tree. In this way, a differentiable decision tree can adaptively learn the optimal weight allocation strategy, ensuring that the contribution of each data source in the fusion process matches its quality and relevance.

[0148] Preferably, to further improve the performance of differentiable decision trees, regularization terms, such as weight decay or dropout, can be introduced during training to prevent overfitting. At the same time, more complex feature functions can be used, such as nonlinear combinations of multiple statistical features, to more accurately assess feature importance. Furthermore, ensemble learning methods using multiple differentiable decision trees can be introduced to combine the outputs of multiple trees through voting or weighted averaging to further improve the accuracy and robustness of weight allocation. In practical applications, the structure and parameters of the decision tree can be dynamically adjusted based on the number and characteristics of the data sources to achieve optimal weight allocation.

[0149] In some embodiments, the hypergraph convolution operation adopts an anisotropic diffusion strategy:

[0150] The node feature update formula is:

[0151]

[0152] Where, F (l+1) is the updated feature matrix, F (l) is the current layer feature, w(e) is the hyperedge weight (range: 0-1), |e| is the number of nodes contained in the hyperedge, MLP is a multi-layer perceptron (including 3 hidden layers), A e is the adjacency matrix block (dimension: n×n).

[0153] It should be noted that the hypergraph convolution operation mentioned in the present invention adopts an anisotropic diffusion strategy, which is a technical means for spatiotemporal joint modeling. Hypergraph convolution is a convolution operation based on a hypergraph structure that can process data with complex associations. The anisotropic diffusion strategy allows node features to propagate in different directions and speeds in the hypergraph, thereby better capturing the spatiotemporal associations between nodes. Under this strategy, the node feature update formula realizes dynamic update of features by considering the hyperedge weight, the number of nodes contained in the hyperedge, and the output of the multi-layer perceptron (MLP), thereby improving the quality of the fusion result.

[0154] Specifically, the anisotropic diffusion strategy in the hypergraph convolution operation involves multiple key parameters and concepts. The hyperedge weight represents the strength of the association between the nodes connected by the hyperedge. Its value range is usually between 0 and 1, and it can be dynamically adjusted according to the spatiotemporal distance between the nodes or other correlation metrics. The number of nodes contained in the hyperedge determines the range of feature propagation, and the multi-layer perceptron (MLP) is used to perform nonlinear transformations on node features to enhance the expressiveness of features. In the node feature update formula, the node feature is updated by calculating the weighted sum of the current layer feature and the hyperedge weight, and combining it with the output of the MLP. This method can effectively handle the problem of feature propagation in complex network structures and is suitable for spatiotemporal joint modeling scenarios.

[0155] Preferably, in the hypergraph convolution operation, the calculation method of the hyperedge weight can be further optimized, for example, by introducing a weight distribution based on the attention mechanism, so that the weight can be dynamically adjusted according to the similarity of the node features, thereby more accurately reflecting the association between the nodes. In addition, the number of MLP layers can be increased or the dimension of its hidden layer can be adjusted to enhance the nonlinear ability of feature transformation. In practical applications, other graph neural network technologies, such as the graph attention network (GAT), can also be combined to further improve the performance of hypergraph convolution. At the same time, the structure and parameters of the hypergraph can be dynamically adjusted according to the specific data characteristics and application scenarios to achieve the best spatiotemporal joint modeling effect.

[0156] In some embodiments, the quality assessment in step S1 includes:

[0157] S11. Calculate protocol compatibility score: Verify whether the data interface complies with the ISO / IEC 21878 standard extension protocol;

[0158] S12. Evaluate the data completeness index: Calculate the topological impact of missing fields using the following formula:

[0159]

[0160] Where C comp is the completeness index (range: 0-1), ρ iis the network centrality (PageRank value) of field i, δ i is the missing state indicator (δ = 1 for missing, δ = 0 for present);

[0161] S13. Measure the timing continuity index: Calculate the inverse of the product of the jitter coefficient and the retransmission rate of the valid data packet.

[0162] It should be noted that the multi-source heterogeneous data fusion method mentioned in this invention performs a multi-dimensional assessment of heterogeneous data sources during the quality assessment phase, including protocol compatibility scores, data completeness indices, and temporal continuity indicators. These evaluation indicators are designed to comprehensively measure the quality of the data sources and ensure good consistency and availability before fusion. The protocol compatibility score verifies whether the data interface conforms to a specific extension protocol, the data completeness index assesses data integrity, and the temporal continuity indicator measures the continuity and stability of the data over time.

[0163] Specifically, the calculation of the protocol compatibility score is achieved by verifying whether the data interface complies with the ISO / IEC21878 standard extended protocol. This standard provides clear specifications for the compatibility of data interfaces, ensuring that data between different data sources can be seamlessly connected. The calculation of the data completeness index is based on the network centrality of the field (such as the PageRank value) and the missing status indicator. The network centrality reflects the importance of the field in the network, while the missing status indicator is used to identify whether the field is missing. The calculation of the time series continuity index is achieved by analyzing the inverse of the product of the jitter coefficient of the valid data packet (that is, the degree of jitter in the data packet transmission) and the retransmission rate. This indicator can effectively measure the continuity and stability of the data in the time series.

[0164] Preferably, during the quality assessment process, the calculation method of the evaluation indicators can be further refined. For example, when calculating the data completeness index, the weight of the network centrality can be dynamically adjusted according to the importance of the field to more accurately reflect the impact of field missing on data integrity. For the time series continuity indicator, more time series features can be introduced, such as the average transmission time of the data packet or the data packet loss rate, to more comprehensively evaluate the time series characteristics of the data. In addition, it is possible to combine machine learning algorithms to automatically learn the optimal parameters of the evaluation indicators through historical data, thereby further improving the accuracy and reliability of the quality assessment. In practical applications, the evaluation indicators and their parameters can be dynamically adjusted according to the specific characteristics of the data source and the application scenario to achieve the best quality assessment effect.

[0165] The above-mentioned embodiments of the present invention have the following beneficial effects: the multi-source heterogeneous data fusion method of the present invention can effectively improve the accuracy and reliability of data fusion. Through multi-dimensional quality assessment, the spatiotemporal consistency, semantic entropy and data freshness of each data source can be comprehensively analyzed, thereby providing an accurate weight distribution basis for subsequent fusion calculations. In the feature decoupling and cross-modal association analysis stage, the effective extraction and alignment of features of different types of data sources can be achieved, and an intermediate feature set with spatiotemporal alignment characteristics can be generated, laying the foundation for high-quality data fusion. The three-level joint optimization architecture further enhances the fusion effect. The first-level feature-level two-way interaction can make full use of the complementary information between different data sources. The second-level spatiotemporal joint modeling can effectively process the spatiotemporal correlation of data. The third-level semantic consistency constraint ensures the consistency of the fusion result with the prior semantic knowledge, thereby comprehensively improving the quality of the fused data. In addition, the feedback adaptive calibration mechanism can dynamically correct the fusion result so that it can adapt to the dynamic changes and uncertainties of the data source, further improving the accuracy and reliability of the fused data.

[0166] In terms of the dynamic weight assignment model, the spatiotemporal consistency weight factor, semantic entropy weight factor, and data freshness decay coefficient of each data source can be calculated, and the final weight can be generated through nonlinear combination, thereby achieving dynamic credibility assessment of each data source. In feature decoupling and cross-modal association analysis, multi-table association decoupling can be performed on structured databases, context-aware segmented vectorization can be performed on unstructured text streams, spatiotemporal joint denoising can be performed on real-time sensor data, and cross-modal feature anchoring can be implemented on image data, thereby generating high-quality intermediate feature sets. In the three-level joint optimization architecture, a bidirectional gated interaction mechanism can be used for feature fusion, a spatiotemporal hypergraph network can be constructed for spatiotemporal modeling, and knowledge graph embedding can be introduced as a regularizer to impose semantic constraints, thereby achieving effective fusion of multi-source heterogeneous data. In the feedback calibration mechanism, a dynamic residual learning model can be constructed and multi-scale calibration can be implemented to further optimize the fusion results. In the bidirectional LSTM network, an asymmetric structure and threshold filter function can be used to effectively predict normal value intervals and replace anomalous data points. In the nonlinear combination operation, differentiable decision trees are used to implement flexible weight assignment. In hypergraph convolution operations, anisotropic diffusion strategies are used to effectively update node features. In quality assessment, protocol compatibility scores, data completeness indices, and temporal continuity metrics can be calculated to comprehensively assess the quality of data sources.

[0167] like Figure 2 As shown, a multi-source heterogeneous data fusion system 200 of some embodiments includes:

[0168] A quality assessment module 201 is configured to perform multi-dimensional quality assessment on at least three heterogeneous data sources, including a structured database, an unstructured text stream, a real-time sensor network, and an image acquisition device;

[0169] The association analysis module 202 is configured to perform feature decoupling and cross-modal association analysis of heterogeneous data sources based on the quality assessment results, and generate an intermediate feature set with spatiotemporal alignment characteristics;

[0170] The weight assignment module 203 is used to calculate the multi-dimensional credibility weight of each data source, wherein the multi-dimensionality includes spatiotemporal consistency, semantic entropy, and data freshness;

[0171] A fusion calculation module 204 is used to perform fusion calculation on the intermediate feature set through a three-level joint optimization architecture, where the first level adopts feature-level bidirectional interaction, the second level performs spatiotemporal joint modeling, and the third level implements semantic consistency constraints;

[0172] The dynamic correction module 205 is used to dynamically correct the fusion result by adopting a feedback adaptive calibration mechanism and output an optimized fusion data cube.

[0173] It is understandable that the modules described in the multi-source heterogeneous data fusion system 200 are similar to those in the reference Figure 1 Therefore, the operations, features and beneficial effects described above for the multi-source heterogeneous data fusion method are also applicable to the multi-source heterogeneous data fusion system 200 and the modules contained therein, and will not be repeated here.

[0174] Reference below Figure 3 , which shows a schematic structural diagram of an electronic device 300 suitable for implementing some embodiments of the present invention. The electronic devices in some embodiments of the present invention may include, but are not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 3 The terminal device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0175] like Figure 3As shown, the electronic device 300 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 302 or a program loaded from a storage device 308 into a random access memory (RAM) 303. Various programs and data required for the operation of the electronic device 300 are also stored in the RAM 303. The processing device 301, the ROM 302, and the RAM 303 are connected to each other via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0176] Typically, the following devices may be connected to the I / O interface 305: an input device 306 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 307 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 308 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 309. The communication device 309 may allow the electronic device 300 to communicate with other devices wirelessly or by wire to exchange data. Figure 3 The electronic device 300 is shown with various devices, but it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed instead. Figure 3 Each block shown in the figure may represent one device, or may represent multiple devices as needed.

[0177] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or a terminal device such as a computer, a server, a mobile phone, or a tablet.

[0178] The above descriptions are merely some preferred embodiments of the present invention and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also encompass other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the above-mentioned inventive concept. For example, a technical solution formed by mutually replacing the above-mentioned features with (but not limited to) technical features having similar functions disclosed in the embodiments of the present invention.

Claims

1. A multi-source heterogeneous data fusion method, characterized in that: The following steps are involved: S1. Perform a multi-dimensional quality assessment on at least three heterogeneous data sources to obtain a quality assessment result, wherein the heterogeneous data sources include a structured database, an unstructured text stream, a real-time sensor network, and an image acquisition device; S2. Based on the quality assessment results, perform feature decoupling and cross-modal correlation analysis of heterogeneous data sources to generate an intermediate feature set with spatiotemporal alignment characteristics; S3. Construct a nonlinear dynamic weight distribution model to calculate the multi-dimensional credibility weight of each data source, including spatiotemporal consistency, semantic entropy, and data freshness; S4. Fusion computation of intermediate feature sets is performed through a three-level joint optimization architecture, where the first level uses feature-level bidirectional interaction, the second level performs spatiotemporal joint modeling, and the third level enforces semantic consistency constraints. S5. Use a feedback adaptive calibration mechanism to dynamically correct the fusion calculation results and output the optimized fusion data cube.

2. The method according to claim 1, characterized in that The dynamic weight allocation model of step S3 includes the following operations: S31. Calculate the spatiotemporal consistency weight factor of the data source, as shown in the following formula: Where W st is the spatiotemporal consistency weight factor, ΔT is the data timestamp difference, L i , L j are the spatial coordinates of data sources i and j, θ1 and θ2 are the learnable parameter matrices; S32. Evaluate the semantic entropy weight factor as shown in the following formula: Where W se is the semantic entropy weight factor, N unique is the number of unique semantic entities, p k is the actual semantic distribution probability, is the theoretical benchmark distribution, M is the total number of semantic categories; S33. Calculate the data freshness decay coefficient, as shown in the following formula: Where W fresh is the freshness attenuation coefficient, λ is the attenuation rate parameter, t current is the current system time, t gen is the data generation time, t max is the preset maximum effective time window, α is the trend gain coefficient, S trend is the slope of the data change trend; S34. Generate the final weight through nonlinear combination, as shown in the following formula: Where W final is the final weight vector, represents the element-wise product of tensors, Represents a feature concatenation operation.

3. The method according to claim 1, characterized in that The step S2 comprises: S21. Decouple multiple tables from the structured database to generate independent subject feature subsets while preserving foreign key constraints. S22. Perform context-aware segmented vectorization on unstructured text streams, using a sliding window mechanism to extract local semantic features, with the window size dynamically adjusted based on the text entropy. S23. Perform spatiotemporal joint noise reduction on real-time sensor data, using a bidirectional LSTM network to predict the normal value interval and perform nonlinear interpolation on data points outside the interval; S24. Implement cross-modal feature anchoring on image data, align the visual feature vector with the text description vector, and generate a joint embedding representation.

4. The method according to claim 1, wherein The three-level joint optimization architecture of step S4 includes: S41. In the first-level feature fusion, a bidirectional gated interaction mechanism is adopted as shown in the following formula: F gate =σ(W g ·[F i ,F j ])⊙tanh(W f ·[F i ,F j ]) Where, F gate is the gated fusion feature, F i 、F j are different source eigenvectors, W g 、W f is the trainable parameter matrix, σ is the sigmoid function, and ⊙ represents the Hadamard product; In the second-level spatiotemporal modeling, a spatiotemporal hypergraph network is constructed, where nodes represent feature vectors and hyperedges connect sets of nodes with spatiotemporal associations. The feature representation is updated through hypergraph convolution. S43. In the third-level semantic constraint, knowledge graph embedding is introduced as a regularization term to minimize the KL divergence between the fusion result and the prior semantic relationship.

5. The method according to claim 1, wherein The feedback calibration mechanism of step S5 includes: S51. Construct a dynamic residual learning model, whose residual calculation formula is: Where R t is the residual vector at the current moment, E t-1 is the error vector at the previous moment, is the current feature gradient matrix, Q(F t ) is the query vector, K(H hist ) is the history state key vector, × represents matrix multiplication; S52. Implement multi-scale calibration to calibrate the fusion results at the macro, micro, and semantic levels simultaneously; Macro-level calibration includes: adjusting feature distributions to match global statistics; Micro-level calibration includes: correcting the deviation of individual data points from the local cluster center; Semantic level calibration includes: constraining the consistency of output with the domain knowledge graph.

6. The method according to claim 3, characterized in that The bidirectional LSTM network in step S23 adopts an asymmetric structure: The number of layers in the forward network is greater than that in the reverse network, and the hidden layer dimension satisfies: Where, d forward is the hidden layer dimension of the feedforward network, d backward is the hidden layer dimension of the reverse network, N sensor is the total number of sensor nodes; The network output passes through the threshold filter function: Where, is the corrected sensor data, P(x t ) is the confidence probability of the data point, τ is the dynamic threshold, μ t|t-1 is the predicted mean, σ t|t-1 is the prediction standard deviation, ∈ is the random noise vector.

7. The method according to claim 2, characterized in that The nonlinear combination in step S34 is implemented by generating the final weight through the nonlinear combination through a differentiable decision tree, including: The tree node splitting condition is determined by the statistical characteristics of the weight factor. The leaf node outputs the combined weight, and the decision path satisfies: Where Path is the decision path function, f k is the characteristic function designed based on the information gain ratio, γ k is the dynamic splitting threshold, and I is the indicator function.

8. The method according to claim 4, characterized in that The hypergraph convolution adopts an anisotropic diffusion strategy; The node feature update formula is: Where, F (l+1) is the updated feature matrix, F (l) is the current layer feature, w(e) is the hyperedge weight, |e| is the number of nodes included in the hyperedge, MLP is the multi-layer perceptron, A e is the adjacency matrix block.

9. The method according to claim 1, characterized in that The quality assessment in step S1 includes: S11. Calculate protocol compatibility score: Verify whether the data interface complies with the ISO / IEC 21878 standard extension protocol; S12. Evaluate the data completeness index: Calculate the topological impact of missing fields using the following formula: Where C comp is the completeness index (range: 0-1), ρ i is the network centrality of field i, δ i is a missing status indicator; S13. Measure the timing continuity index: Calculate the inverse of the product of the jitter coefficient and the retransmission rate of the valid data packet.

10. A multi-source heterogeneous data fusion system, used to use the method described in any one of claims 1 to 9, characterized in that: include: A quality assessment module, configured to perform multi-dimensional quality assessment on at least three heterogeneous data sources, including a structured database, an unstructured text stream, a real-time sensor network, and an image acquisition device; A correlation analysis module, configured to perform feature decoupling and cross-modal correlation analysis of heterogeneous data sources based on the quality assessment results, and generate an intermediate feature set with spatiotemporal alignment characteristics; A weight assignment module is used to calculate the multi-dimensional credibility weight of each data source, including spatiotemporal consistency, semantic entropy, and data freshness; A fusion computing module is used to perform fusion computing on the intermediate feature set through a three-level joint optimization architecture, where the first level uses feature-level bidirectional interaction, the second level performs spatiotemporal joint modeling, and the third level implements semantic consistency constraints; The dynamic correction module is used to dynamically correct the fusion results using a feedback adaptive calibration mechanism and output the optimized fusion data cube.

Images