An intelligent decision data processing method and system for fusing multi-source heterogeneous data

Abstract

The application relates to the technical field of decision data processing, in particular to an intelligent decision data processing method and system fusing multi-source heterogeneous data; the method comprises the following steps: allocating a source identifier to the heterogeneous data and constructing an anchor point structure body, obtaining comparable anchor point representation and source feature representation through standardized mapping; calculating a similarity factor and a mutual exclusion and complement factor between the sources based on the comparable anchor point representation, combining a historical fusion quality record to construct a source pair bridging constraint set and solve a source pair bridging vector, and performing bridging fusion on the source feature representation to obtain a fused feature representation; on the basis, iterative decision calculation and error measurement are performed, bridging mismatch and data and timing problems are distinguished according to error attribution, and bridging quality updating or data isolation and timing realignment is triggered, and a decision result or abnormal state data is output under the control of iterative counting. The method improves the robustness of the fusion result and the convergence ability of the decision error in the complex multi-source heterogeneous data scene.

Description

Technical Field

[0001] This invention relates to the field of decision data processing technology, and specifically to an intelligent decision data processing method and system that integrates multi-source heterogeneous data. Background Technology

[0002] In intelligent decision-making scenarios driven by multi-source heterogeneous data, information systems typically need to process multiple data sources simultaneously, including business databases, log streams, sensor data, third-party interfaces, and text archives. These data sources exhibit significant differences in data structure, time granularity, sampling frequency, and semantic expression. Existing technologies generally align different data sources to a unified feature space through methods such as format conversion, feature extraction, and vectorization mapping. Then, decision calculations are performed using rule engines, machine learning models, or deep learning models. In this process, multi-source data fusion and online decision feedback have become important foundations for supporting applications such as intelligent risk control, intelligent operation and maintenance, and production scheduling optimization. In intelligent decision-making applications oriented towards multi-source heterogeneous data, there are still two technical challenges: multi-source data fusion and online decision correction. Firstly, existing multi-source fusion methods often rely solely on similarity or relevance indicators to weight or allocate attention to each data source, lacking explicit modeling of conflict and complementarity relationships between different data sources, and lacking structured records and constraint mapping of historical fusion quality. This results in the fusion strategy failing to promptly suppress the influence of highly conflicting sources when data sources with inconsistent definitions, measurement biases, or accumulated noise participate in the decision-making process, and also failing to stably amplify the contributions of data sources with long-term reliable performance. Consequently, the multi-source heterogeneous data fusion results are ineffective. Firstly, the decision output is not stable enough due to its sensitivity to anomalies. Secondly, existing online decision feedback mechanisms usually only evaluate the deviation between the decision result and the reference result based on a single error index. They lack an attribution mechanism that breaks down the decision error into errors at the fusion strategy level and errors at the data quality and time series level. They also lack technical means to select different correction paths for different error sources. They often only passively correct the error by uniformly adjusting the model parameters or the overall weights, which makes it difficult to converge the decision error in a timely manner and to trace the error source. After multiple iterations, invalid fusion and decision calculations may still be repeatedly executed under the condition of data quality problems or time series misalignment. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent decision-making data processing method and system that integrates multi-source heterogeneous data, so as to solve the technical problems of multi-source data fusion and online decision correction mentioned in the background art.

[0004] To achieve the above objectives, the technical solution of the present invention is: an intelligent decision-making data processing method that integrates multi-source heterogeneous data, comprising: S1. Obtain the original heterogeneous data and assign a one-to-one corresponding source identifier. Based on the source identifier, extract the anchor points of the original heterogeneous data to generate an anchor point structure. Perform anchor point normalization mapping on the anchor point structure to obtain a comparable anchor point representation. Generate the source feature representation based on the comparable anchor point representation. Among them, the anchor structure is a data structure formed by performing field-level aggregation on the original heterogeneous data associated with the source identifier based on the preset anchor fields; the comparable anchor representation is a multi-dimensional numerical representation generated based on the anchor structure according to the preset anchor field order, unified encoding rules and unified units. S2. Based on the comparable anchor point representation, calculate the source similarity factor and the source mutually exclusive complementary factor. Combine the source similarity factor, historical fusion quality record and the source mutually exclusive complementary factor to construct the source pair bridging constraint set. Solve the constraints of the source pair bridging constraint set to generate the source pair bridging vector. Based on the source pair bridging vector, perform bridging fusion on the source feature representation to generate the fusion feature representation. Among them, the inter-source mutual exclusion complementary factor is a numerical factor calculated based on the difference set and intersection set of the comparable anchor point representations corresponding to different sources; the source pair bridging constraint set is a set of constraint conditions composed of the inter-source similarity factor, the inter-source mutual exclusion complementary factor and the historical fusion quality record for the same source pair; the source pair bridging vector is a multi-dimensional vector representation that uses the source pair as an index to numerically encode the weight allocation and bridging state of any pair of source feature representations in the preset bridging space. S3. Set an iteration counter, perform decision calculations based on the fused feature representation to obtain decision result data, calculate the decision error metric between the decision result data and the preset verification data, and increment the iteration counter by one; when the decision error metric meets the threshold condition, output the decision result data and terminate the current decision data processing; when the decision error metric does not meet the threshold condition and the iteration counter has not reached the maximum number of iterations, perform error attribution classification on the decision error metric and the fused feature representation to obtain the attribution category identifier. Among them, the attribution category identifiers include bridging mismatch identifiers and non-bridging mismatch identifiers; S4. When the attribution category identifier is a bridging mismatch identifier, update the historical fusion quality record and return to execute S2. When the attribution category identifier is a non-bridging mismatch identifier, perform data isolation and temporal realignment on the source identifier and return to execute S2. When the iteration counter reaches the maximum number of iterations, output the abnormal status data and terminate the current decision data processing.

[0005] Preferably, in step S1, the anchor point of the original heterogeneous data is a set of field name and field value pairs extracted from the original heterogeneous data associated with the source identifier based on a preset anchor point field set, used to represent different types of original heterogeneous data with a unified anchor point field dimension; the fields of the anchor point structure include at least structural anchor point fields, statistical anchor point fields, and time series anchor point fields; the anchor point normalization mapping process is a mapping operation performed on different types of anchor point fields in the anchor point structure; the comparable anchor point representation is also used to uniformly measure and compare anchor point fields between different sources and as input for calculating source feature representation, inter-source similarity factor, and inter-source mutually exclusive complementary factor; the source feature representation is a feature vector representation obtained based on the comparable anchor point representation through a preset feature transformation mapping operation.

[0006] Preferably, in S2, the inter-source similarity factor is calculated based on the comparable anchor representations corresponding to different sources and is a numerical factor used to characterize the degree of similarity between different sources at the anchor field level; the inter-source mutual exclusion and complementarity factor includes an inter-source mutual exclusion factor and an inter-source complementarity factor. The inter-source mutual exclusion factor is calculated based on the difference set of the comparable anchor representations corresponding to different sources and is a numerical factor used to represent the degree of conflict between different sources at the anchor field. The inter-source complementarity factor is calculated based on the intersection set of the comparable anchor representations corresponding to different sources and is a numerical factor used to represent the degree of complementarity between different sources at the anchor field. The inter-source mutual exclusion and complementarity factor and the inter-source similarity factor are used together to distinguish between conflict-dominated source pairs and complementarity-dominated source pairs when constructing the source pair bridging constraint set.

[0007] Preferably, in S2, the source pair is a combination consisting of two different source identifiers and source feature representations corresponding one-to-one with the two source identifiers; the source pair bridging constraint set is a set of multiple constraints established for the same source pair, used to limit the value range and value relationship of each component of the source pair bridging vector under the combined effect of the inter-source similarity factor, the inter-source mutual exclusion and complementarity factor, and the historical fusion quality record; the constraint types of the source pair bridging constraint set specifically include weight normalization constraints, mutual exclusion constraints, and historical quality constraints.

[0008] Preferably, in step S2, the constraint solving strategy adopts a numerical solution method based on the source pair bridging constraint set to construct a constraint optimization model. The source pair bridging vector is used as the optimization variable, and the preset objective function is optimized under the premise of satisfying the source pair bridging constraint set. The preset objective function includes a similarity term based on the inter-source similarity factor and a quality term based on historical fusion quality records. The numerical solution method includes convex optimization algorithm and iterative projection algorithm, and the source pair bridging vector is iteratively updated.

[0009] Preferably, in S2, the source pair bridging vector includes a weight component for numerically weighting the source feature representations corresponding to each source in the source pair, and a bridging state component for characterizing the stability and confidence level of the source pair bridging. The weight component is used to control the contribution ratio of each source feature representation corresponding to each source in the source pair in the fused feature representation. The preset bridging space is a vector space that carries the source feature representations and the source pair bridging vector. The preset bridging space corresponds to the source feature representation space in terms of dimension, and is extended to include the bridging state component on the basis of the preset bridging space, for uniformly representing the source feature representations and the source pair bridging vector. The bridging fusion is an operation of performing component-weighted combination on the source feature representations belonging to the same source pair based on the source pair bridging vector.

[0010] Preferably, in step S3, the decision error metric is a composite numerical index calculated based on the decision result data and the preset verification data. It is used to determine whether the threshold condition is met in each iteration and serves as the input for error attribution classification. The decision error metric is specifically composed of a main error component and a bridging regularization component. The main error component is used to measure the deviation between the decision result data and the preset verification data in the output space, while the bridging regularization component is used to measure the magnitude change, direction change, and number of sources participating in bridging in the current iteration round.

[0011] Preferably, in step S3, the attribution category identifier is a classification identifier determined based on the decision error metric, used to indicate that the decision error mainly originates from the source pair bridging vector configuration and from the original heterogeneous data and its temporal relationship; the error attribution classification is an operation that performs classification operations on the decision error metric and its decomposed principal error component and bridging regularization component, and determines the attribution category identifier by combining the fusion feature representation corresponding to the iteration, the statistical value of the mutually exclusive complementary factors between sources, and the historical fusion quality record change trend; the judgment logic of the error attribution classification is specifically as follows: when the ratio of the bridging regularization component to the principal error component is greater than a first preset threshold, the attribution category identifier is determined as a bridging mismatch identifier; when the ratio of the bridging regularization component to the principal error component is less than or equal to a second preset threshold, the attribution category identifier is determined as a non-bridging mismatch identifier.

[0012] On the other hand, the present invention provides an intelligent decision data processing system that integrates multi-source heterogeneous data, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the aforementioned intelligent decision data processing method that integrates multi-source heterogeneous data.

[0013] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: 1. In this invention, a source pair bridging constraint set is constructed based on the anchor point structure, comparable anchor point representation, and source similarity factor and source mutually exclusive complementarity factor. The source pair bridging vector is generated by solving the constraint optimization problem. Bridging fusion of multi-source heterogeneous data can be performed in a unified bridging space. During the fusion process, similarity, conflict, complementarity and historical fusion quality records are considered at the same time to suppress the influence of poor quality or conflict-dominated data sources on the fusion results and improve the stability and reliability of multi-source fusion results. 2. In this invention, the decision error metric is split into a main error component and a bridging regularization component. Based on the error component ratio, fusion feature representation, statistical values ​​of inter-source mutually exclusive complementary factors, and the trend of historical fusion quality record changes, error attribution classification is performed. The error is divided into two categories: bridging mismatch and non-bridging mismatch. The bridging quality update path and the data isolation and temporal realignment path are triggered respectively. Different correction strategies can be selected for different error sources, which can accelerate the convergence of decision errors and form error source diagnostic information that can be used for subsequent analysis. Attached Figure Description

[0014] Figure 1 This is a flowchart of one embodiment of the present invention. Detailed Implementation

[0015] Example 1, as Figure 1 As shown, the present invention proposes an intelligent decision-making data processing method that integrates multi-source heterogeneous data. The specific implementation steps are as follows: S1. Obtain the original heterogeneous data and assign a one-to-one corresponding source identifier. Based on the source identifier, extract the anchor points of the original heterogeneous data to generate an anchor point structure. Perform anchor point normalization mapping on the anchor point structure to obtain a comparable anchor point representation. Generate the source feature representation based on the comparable anchor point representation. Among them, the anchor structure is a data structure formed by performing field-level aggregation on the original heterogeneous data associated with the source identifier based on the preset anchor fields; the comparable anchor representation is a multi-dimensional numerical representation generated based on the anchor structure according to the preset anchor field order, unified encoding rules and unified units. S2. Based on the comparable anchor point representation, calculate the source similarity factor and the source mutually exclusive complementary factor. Combine the source similarity factor, historical fusion quality record and the source mutually exclusive complementary factor to construct the source pair bridging constraint set. Solve the constraints of the source pair bridging constraint set to generate the source pair bridging vector. Based on the source pair bridging vector, perform bridging fusion on the source feature representation to generate the fusion feature representation. Among them, the inter-source mutual exclusion complementary factor is a numerical factor calculated based on the difference set and intersection set of the comparable anchor point representations corresponding to different sources; the source pair bridging constraint set is a set of constraint conditions composed of the inter-source similarity factor, the inter-source mutual exclusion complementary factor and the historical fusion quality record for the same source pair; the source pair bridging vector is a multi-dimensional vector representation that uses the source pair as an index to numerically encode the weight allocation and bridging state of any pair of source feature representations in the preset bridging space. S3. Set an iteration counter, perform decision calculations based on the fused feature representation to obtain decision result data, calculate the decision error metric between the decision result data and the preset verification data, and increment the iteration counter by one; when the decision error metric meets the threshold condition, output the decision result data and terminate the current decision data processing; when the decision error metric does not meet the threshold condition and the iteration counter has not reached the maximum number of iterations, perform error attribution classification on the decision error metric and the fused feature representation to obtain the attribution category identifier. Among them, the attribution category identifiers include bridging mismatch identifiers and non-bridging mismatch identifiers; S4. When the attribution category identifier is a bridging mismatch identifier, update the historical fusion quality record and return to execute S2. When the attribution category identifier is a non-bridging mismatch identifier, perform data isolation and temporal realignment on the source identifier and return to execute S2. When the iteration counter reaches the maximum number of iterations, output the abnormal status data and terminate the current decision data processing.

[0016] In this embodiment S1, the anchor point of the original heterogeneous data is a set of field name and field value pairs extracted from the original heterogeneous data associated with the source identifier based on a preset anchor point field set. This set is used to represent different types of original heterogeneous data with a unified anchor point field dimension. The fields of the anchor point structure include at least structural anchor point fields, statistical anchor point fields, and time series anchor point fields. The anchor point normalization mapping process is a mapping operation performed on different types of anchor point fields in the anchor point structure. The comparable anchor point representation is also used to uniformly measure and compare anchor point fields between different sources and serves as input for calculating source feature representation, inter-source similarity factor, and inter-source mutually exclusive complementary factor. The source feature representation is a feature vector representation obtained by a preset feature transformation mapping operation based on the comparable anchor point representation.

[0017] In this embodiment S1, the original heterogeneous data includes relational database table records, key-value logs, message queue events, device-collected time-series data, and structured or semi-structured data returned by the interface. Within a single intelligent decision-making task cycle, data sources belonging to the same data source, the same collection logic, or the same business role are divided into a source set. A unique source identifier is assigned to each source set. The source identifier can be formed by combining the data source category marker, the physical source marker, and the current decision cycle marker, and maintains a one-to-one correspondence with all the original heterogeneous data in the source set within the decision cycle. The preset anchor field set is selected and configured from the candidate field set based on the structural, statistical, and time information involved in the target decision-making task. The preset anchor field set can use different field combinations in different application scenarios, but remains fixed in the same decision-making task, and is used to limit the range of fields used when extracting anchors from the original heterogeneous data.

[0018] In this embodiment S1, for each source identifier, multiple original heterogeneous data records associated with that source identifier are aggregated within the current decision-making cycle. Field names and values ​​are extracted and paired within a predefined anchor field set to form anchors. Field-level aggregation is then performed on multiple records at the field dimension to obtain an anchor set containing a single field name and the aggregated field value. An anchor structure is constructed based on this anchor set. The fields of the anchor structure are functionally divided into structural anchor fields, statistical anchor fields, and time-series anchor fields. Structural anchor fields reflect the type and source characteristics of the data source itself, such as data type identifier, data source system identifier, and data acquisition channel identifier. The anchor fields are used to reflect the statistical behavior of the corresponding indicators of the source within the decision-making cycle, such as the mean of field values, the variance of field values, the missing rate, or the proportion of outliers. The time-series anchor fields are used to reflect the evolution information of the source data on the time axis, such as the most recent update timestamp, the interval between consecutive updates, and the current decision window span. For fields that exist in the preset anchor field set but are missing in the current source data, missing field placeholders are written when constructing the anchor structure. A uniformly agreed default flag or value is used to place the missing fields, so that the anchor structures of different sources are consistent in the field dimension, providing a complete field space for subsequent normalized mapping and comparison operations.

[0019] In this embodiment S1, for each anchor point structure, anchor point normalization mapping is performed according to the type of the anchor point field, mapping different types of anchor point fields to a unified anchor point coordinate system; for numerical anchor point fields, dimensional normalization or standardization processing is performed, mapping numerical values ​​of different dimensions to a preset interval or standardized space to eliminate dimensional differences; for category anchor point fields, encoding mapping is performed, mapping category labels to preset integer codes or sparse vector codes, so that category information can be represented in the same vector space as numerical information; for time anchor points, time stamping is performed. Standardization or periodic expansion mapping converts absolute time into relative time intervals, time interval indices, or periodic decomposition representations, thereby characterizing time features under a unified time reference. After completing the above mapping, the values ​​of each standardized anchor field are arranged in order according to the preset anchor field order to form a multidimensional numerical vector, which can then be used to compare anchor representations. Each comparable anchor representation corresponds one-to-one with the source identifier and is used to measure and compare anchor fields of different sources in a unified anchor coordinate system. It also serves as input for subsequent generation of source feature representations, calculation of inter-source similarity factors, and calculation of inter-source mutually exclusive complementary factors.

[0020] In this embodiment S1, the comparable anchor representation corresponding to each source identifier is indexed to the source identifier in the storage structure. Within a decision task cycle, a set of comparable anchor representations indexed by the source identifier is formed. The set of comparable anchor representations can be stored in the form of a vector list or a matrix, so that the corresponding representation can be quickly retrieved by the source identifier when constructing bridging constraints later. In this implementation, the comparable anchor representation only undertakes the functions of unified representation and comparison, without introducing task-related feature transformations, and maintains a one-to-one semantic correspondence with the anchor structure. That is, each anchor field has a stable dimensional position in the comparable anchor representation, avoiding the loss of field alignment relationships between different sources due to feature transformations. The set of comparable anchor representations can be used as input for the generation of source feature representations, and can also be used as the basic data set for constructing inter-source similarity factors and inter-source mutually exclusive complementary factors.

[0021] In this embodiment S1, the source feature representation is a feature vector representation obtained based on the comparable anchor representation through a preset feature transformation mapping operation. The preset feature transformation mapping can be a linear mapping, a dimensionality reduction mapping, or a nonlinear feature transformation mapping. In one implementation, the preset feature transformation mapping uses a linear transformation matrix with fixed parameters to perform matrix multiplication on the comparable anchor representation, mapping the high-dimensional anchor space to a lower-dimensional source feature space, which is used to compress redundant fields and highlight the combined features sensitive to the decision task. In another implementation, the preset feature transformation mapping uses a pre-trained feedforward neural network or an automatic... The encoder inputs the comparable anchor representation into the network to obtain the source feature representation, which is used to capture the nonlinear relationship between anchor fields. In any of the above implementations, the source feature representation corresponds one-to-one with the source identifier and is different from the comparable anchor representation in terms of dimension and semantics. The comparable anchor representation emphasizes cross-source comparability, while the source feature representation emphasizes the feature expression capability for intelligent decision-making tasks. As the object of subsequent source action on the bridging vector and one of the inputs for decision calculation, the source feature representation provides the feature expression basis for calculating and bridging fusion of inter-source similarity factors and inter-source mutually exclusive complementary factors based on the comparable anchor representation.

[0022] In this embodiment S2, the inter-source similarity factor is calculated based on the comparable anchor representations corresponding to different sources and is a numerical factor used to characterize the degree of similarity between different sources at the anchor field level. The inter-source mutual exclusion and complementarity factor includes an inter-source mutual exclusion factor and an inter-source complementarity factor. The inter-source mutual exclusion factor is calculated based on the difference set of the comparable anchor representations corresponding to different sources and is a numerical factor used to represent the degree of conflict between different sources at the anchor field. The inter-source complementarity factor is calculated based on the intersection set of the comparable anchor representations corresponding to different sources and is a numerical factor used to represent the degree of complementarity between different sources at the anchor field. The inter-source mutual exclusion and complementarity factor and the inter-source similarity factor are used together to distinguish between conflict-dominated source pairs and complementarity-dominated source pairs when constructing the source pair bridging constraint set.

[0023] In this embodiment S2, for all comparable anchor representations, firstly, a pairwise search is performed on the comparable anchor representations corresponding to each pair of candidate source identifiers within the unified anchor field space. Each comparable anchor representation corresponding to a source identifier is regarded as a numerical vector of the same dimension. In this vector space, a candidate list of source pairs is generated according to a preset pairing strategy. The pairing strategy may include pairing all sources in the same business domain, or it may include limited pairing between sources that are connected in physical topology or have a collaborative relationship in business logic. For each pair of sources in the candidate list, the inter-source similarity factor and the inter-source mutual exclusion and complementarity factor are calculated in sequence, and the calculation results are bound to the source identifiers to form an inter-source relationship measurement set indexed by the source pair, providing input for the subsequent construction of the source pair bridging constraint set.

[0024] In this embodiment S2, the calculation of the inter-source similarity factor takes the comparable anchor points corresponding to two different sources as input. Under the unified anchor point field order, a one-to-one similarity calculation is performed on the two vectors in the numerical dimension. For numerical anchor points, the numerical difference can be calculated by dimension and converted into a similarity contribution value through a distance function, such as using normalized Euclidean distance or correlation coefficient to obtain a single-dimensional similarity, and then weighted summation or weighted average is performed on each dimension to obtain the overall similarity factor. For categorical anchor points, the similarity factor can be calculated based on whether the categories are the same or whether they belong to a preset similar category set. Similar or dissimilar labels are generated, and the category matching results are converted into numerical contributions by statistically analyzing the proportion of similar labels or using a preset mapping table. These contributions are then normalized to the similarity contributions of numerical fields. For time-based anchor fields, a time similarity component is generated based on whether the time difference is within a preset time window and whether the time period positions match. Finally, the similarity components of numerical, categorical, and time-based anchor fields are combined to obtain a single scalar inter-source similarity factor. This factor takes values ​​within a preset range and is used to characterize the overall similarity between the two sources at the anchor field level.

[0025] In this embodiment S2, the inter-source mutually exclusive complementary factor consists of an inter-source mutually exclusive factor and an inter-source complementary factor. Both are calculated based on the difference set and intersection set of the comparable anchor points corresponding to different sources. The construction process of the difference set is as follows: under a unified anchor point field order, the values ​​of the two sources at each anchor point field are compared. For numerical fields, when the difference of the normalized numerical values ​​exceeds a preset difference threshold, the field is added to the difference set. For categorical fields, when the category labels are different or do not belong to the same similar category group, the field is added to the difference set. For time fields, when the time difference exceeds a preset tolerance time window, the field is added to the difference set. The mutually exclusive factor can be calculated by statistically analyzing the difference set. The proportion of the number of anchor fields to the total number of anchor fields, or obtained by weighting and normalizing the difference magnitudes of the difference fields, is used to represent the degree of conflict between the two sources on the anchor fields. The construction process of the intersection set is as follows: under the unified anchor field order, the values ​​of the two sources on each anchor field are compared. When the difference of the numerical field is within a preset close range or the categorical field matches and the time field is in the same period position, the field is added to the intersection set. The complementarity factor can be obtained by statistically analyzing the proportion of the number of fields in the intersection set to the total number of anchor fields, or by combining and normalizing the information content indicators of the intersection fields, and is used to represent the degree of complementarity between the two sources on the anchor fields.

[0026] In this embodiment S2, the inter-source similarity factor and the inter-source mutually exclusive complementary factor are used together to classify the source pairs in terms of bridging semantics. The source pairs are classified into conflict-dominant source pairs or complementary-dominant source pairs by a preset judgment rule. In one implementation, a combined scoring function is constructed based on the mutually exclusive factor, the complementary factor, and the inter-source similarity factor. When the mutually exclusive factor is greater than a first threshold and significantly greater than the complementary factor, the source pair is marked as a conflict-dominant source pair. When the complementary factor is greater than a second threshold and significantly greater than the mutually exclusive factor, the source pair is marked as a complementary-dominant source pair. When both are in the middle range, the bias can be determined according to the similarity factor or the pair can be marked as a neutral source pair. The judgment results of conflict-dominant source pairs and complementary-dominant source pairs have different influence weights when constructing constraints in the subsequent process. This is used to distinguish between source pairs that are mainly for conflict suppression and source pairs that are mainly for information enhancement in the source pair bridging constraint set.

[0027] In this embodiment S2, a source pair is a combination of two different source identifiers and source feature representations corresponding one-to-one with the two source identifiers. The source pair bridging constraint set is a set of multiple constraints established for the same source pair, used to limit the value range and value relationship of each component of the source pair bridging vector under the combined effect of the inter-source similarity factor, the inter-source mutual exclusion and complement factor, and the historical fusion quality record. The constraint types of the source pair bridging constraint set specifically include weight normalization constraints, mutual exclusion constraints, and historical quality constraints. Among them, the weight normalization constraint restricts each component of the source pair bridging vector to a preset value range and requires that the sum of each component satisfies the preset constraint condition. The mutual exclusion constraint sets an upper limit on the value of the component in the source pair bridging vector corresponding to the high conflict source according to the magnitude of the inter-source mutual exclusion factor. The historical quality constraint sets a lower limit or penalty weight on the component in the source pair bridging vector corresponding to the low-quality historical fusion result according to the historical fusion quality record, so that the inter-source similarity factor, the inter-source mutual exclusion and complement factor, and the historical fusion quality record are respectively mapped to the corresponding similarity constraints, conflict constraints, and quality constraints.

[0028] In this embodiment S2, a source pair is a combination consisting of two different source identifiers and source feature representations that correspond one-to-one with the two source identifiers. Within the current decision task cycle, a source pair set is first generated based on the source identifier set according to a preset pairing rule. The pairing rule may include performing pairwise combinations on all source identifiers to generate a complete source pair set, or it may include filtering source identifiers with potential bridging value based on physical topology, business grouping information, or similarity thresholds, and then performing pairwise combinations within the filtering results to generate a source pair set. In the storage structure, the source pair uses ordered or unordered source identifier pairs as index keys, and correspondingly stores information such as source feature representations, inter-source similarity factors, inter-source mutually exclusive and complementary factors, and historical fusion quality records related to the source pair. Within a decision task cycle, a bridging input data set oriented towards source pairs is formed, providing a complete context for constructing an independent source pair bridging constraint set for each source pair.

[0029] In this embodiment S2, the source pair bridging constraint set is established for the same source pair and consists of multiple constraint conditions. These constraints are used to limit the value range and value relationship of each component of the source pair bridging vector under the combined effect of the inter-source similarity factor, the inter-source mutual exclusion and complementarity factor, and the historical fusion quality record. During the construction process, a weight normalization constraint is first initialized for each source pair, restricting each weight component in the source pair bridging vector to a preset numerical range. For example, the lower bound of the range is zero or a preset non-negative lower bound, and the upper bound is one or a preset positive upper bound. It is also set that the sum of all weight components satisfies either equal to one or less than one. The normalization condition is used to ensure that the contribution ratio of each source in the source pair to the fused feature representation is within a reasonable range. Then, based on the source similarity factor corresponding to the source pair, additional similarity-related constraints are generated. For example, when the similarity factor is higher than the third threshold, an additional constraint is added to limit the difference of weight components to no more than a preset limit, so that the weight distribution of highly similar sources is more balanced when bridging and fusing. When the similarity factor is lower than the fourth threshold, the weight difference restriction can be relaxed to allow for greater weight adjustment space, so as to highlight more reliable data sources in scenarios with significant conflicts or complementarities.

[0030] In this embodiment S2, mutual exclusion constraints and historical quality constraints correspond to conflict constraints and quality constraints, respectively, in the source pair bridging constraint set. The construction process of mutual exclusion constraints is as follows: for source pairs marked as conflict-dominant or with mutual exclusion factors exceeding a preset mutual exclusion threshold, the upper limit of the weight component corresponding to the high-conflict source is calculated based on the magnitude of the mutual exclusion factor. The upper limit can be set as a combination of the basic upper limit and the mutual exclusion factor function, for example, multiplying the basic upper limit by one minus the function value of the mutual exclusion factor. The larger the mutual exclusion factor, the lower the upper limit. This method suppresses the contribution of sources with severe conflicts in bridging fusion. The construction process of historical quality constraints is as follows: the quality index corresponding to the source pair is read from the historical fusion quality record. The quality index can be generated based on the statistical results of decision errors, bridging stability, or confidence in multiple past decision cycles. For sources with low historical quality indexes, the upper or lower limit of their corresponding weight component is adjusted, or a penalty weight related to the weight component is introduced into the preset objective function, so that sources with poor historical quality are less likely to obtain excessive weight in subsequent solutions, while sources with good historical quality can obtain a larger weight space within the allowable range of constraints.

[0031] In this embodiment S2, the various constraints in the source pair bridging constraint set are organized with a unified constraint description structure. Similarity constraints, mutual exclusion constraints, and quality constraints are instantiated from parameter sets generated based on inter-source similarity factors, inter-source mutual exclusion and complementary factors, and historical fusion quality records, respectively. In one implementation, similarity constraints are stored as weight difference limit parameters, mutual exclusion constraints are stored as weight upper limit adjustment parameters, and quality constraints are stored as penalty coefficients or lower limit parameters. All constraints are associated with the source pair bridging vector component indexes corresponding to the source pair. After the constraint generation and organization are completed, a source pair bridging constraint set is formed for the source pair. The source pair bridging constraint set is used to provide constraint boundaries and value relationships when optimizing and updating the source pair bridging vectors in the future.

[0032] In this embodiment S2, the constraint solving strategy adopts a numerical solution method based on the source pair bridging constraint set to construct a constraint optimization model. The source pair bridging vector is used as the optimization variable, and the preset objective function is optimized under the premise of satisfying the source pair bridging constraint set. The preset objective function includes a similarity term based on the inter-source similarity factor and a quality term based on historical fusion quality records. The numerical solution method includes a convex optimization algorithm and an iterative projection algorithm, and the source pair bridging vector is iteratively updated.

[0033] In this embodiment S2, for the source pair bridging constraint set that has been constructed for each source pair, the constraint solution strategy adopts a numerical solution method based on the source pair bridging constraint set to construct a constraint optimization model. When constructing the constraint optimization model, all components of the source pair bridging vector corresponding to the source pair are used as optimization variables. The weight normalization constraint, mutual exclusion constraint, and historical quality constraint are respectively transformed into variable value range constraints, linear inequality constraints, or other forms of explicit constraints in the optimization problem. The source similarity factor, source mutual exclusion and complement factor, and historical fusion quality record are mapped as a set of parameters used to restrict or adjust the value relationship of each component. On this basis, a constraint optimization model is formed with the source pair bridging vector as the independent variable, the preset objective function as the optimization objective, and the source pair bridging constraint set as the constraint condition set. The source pair bridging vector update value that satisfies the constraint conditions is obtained by performing numerical solution on the constraint optimization model.

[0034] In this embodiment S2, using the source pair bridging vector as the optimization variable means that all adjustable components in the source pair bridging vector belonging to the same source pair are uniformly regarded as a set of variables that need to be determined through numerical calculation. This set of variables includes at least a weight component used to control the contribution ratio of each source feature representation in the fused feature representation and a bridging state component used to encode the current bridging state. When constructing the constraint optimization model, value range constraints and relational constraints derived from the source pair bridging constraint set are applied to this set of variables respectively, so that the optimization process always stays within the feasible region when adjusting the specific values ​​of the weight components and the bridging state components. On this basis, a preset objective function is used to rank the merits of different value combinations within the feasible region. The preset objective function includes similarity based on the inter-source similarity factor. The similarity term measures whether the current weight configuration is consistent with the similarity relationship between sources by measuring the degree of matching between the weight components corresponding to the bridging vector and the similarity factor between sources. The quality term measures whether the current weight configuration is biased towards sources with higher historical fusion quality by measuring the degree of matching between the weight components and the historical fusion quality records. In another implementation, the preset objective function may also include a smoothing term based on the magnitude of the change in the bridging state components and a constraint term based on the sparsity of the weight components, which are used to suppress the drastic fluctuations of the bridging state between adjacent iterations and control the number of effective sources participating in bridging. The preset objective function as a whole can be fixed when constructing the optimization model by weighted summation of the above multiple terms or other scalar combinations.

[0035] In this embodiment S2, the numerical solution method includes convex optimization algorithms and iterative projection algorithms. The convex optimization algorithm performs iterative updates on the source pair bridging vectors based on first- or second-order information when the preset objective function and the source pair bridging constraint set satisfy the convexity condition. For example, it uses gradient descent, quasi-Newton algorithms, or interior-point methods to search for the minimum point of the preset objective function within the feasible region. The iterative projection algorithm ensures that the constraint conditions are satisfied by projecting intermediate results back to the feasible region defined by the source pair bridging constraint set after each update step. In another implementation, the numerical solution method may also include coordinate descent algorithms, alternating direction multiplier methods, or other distribution methods. The algorithm updates the source pair bridging vector numerically by alternating updates between different variable blocks, alternating projections between different constraint sets, or parallel solutions between different source pairs. In any of the above implementations, iterative updates of the source pair bridging vector refer to starting from the initial bridging vector and adjusting the bridging vector once in each iteration based on the current gradient information, search direction, or projection rules. Under the premise of satisfying the constraints, the value of the preset objective function is gradually improved until the preset convergence threshold or the upper limit of the number of internal iterations is reached. Through this iterative update process, the optimal or suboptimal source pair bridging vector configuration is approximated within the feasible region.

[0036] In this embodiment S2, the initial bridging vector for constraint solving can be heuristically initialized based on the inter-source similarity factor and historical fusion quality records. For example, a larger initial weight component is assigned to sources with a high similarity factor and good historical fusion quality, while a smaller initial weight component is assigned to sources with a low similarity factor or poor historical fusion quality. The initial bridging state component reflects the stability of the historical bridging configuration, and the above numerical solution process is performed based on this. During the constraint optimization solution process, the constraint conditions ensure that the value of the source pair bridging vector after each update always meets the requirements of weight normalization, conflict suppression, and quality bias. The components of the preset objective function guide the weight component and the bridging state component to adjust in the feasible region towards similarity matching, historical quality priority, and state stability. Through multiple iterations, a compromise is formed between the constraint boundary and the objective optimization to obtain a source pair bridging vector that is compatible with the similarity relationship, mutual exclusion and complementarity relationship, and historical quality status of the current source pair.

[0037] In this embodiment S2, the source pair bridging vector includes a weight component for numerically weighting the source feature representations corresponding to each source in the source pair, and a bridging state component for characterizing the stability and confidence level of the source pair bridging. The weight component is used to control the contribution ratio of each source feature representation corresponding to each source in the source pair in the fused feature representation, and the bridging state component is used to provide a bridging state reference when subsequently attributing decision errors and updating historical fusion quality records. The preset bridging space is a vector space that carries the source feature representations and the source pair bridging vector. The preset bridging space corresponds to the source feature representation space in terms of dimension, and is extended to include the bridging state component, which is used to uniformly represent the source feature representations and the source pair bridging vector. The bridging fusion is an operation based on the source pair bridging vector to perform a component-weighted combination of the source feature representations belonging to the same source pair, including linear weighting of the source feature representations at the field level and the overall vector level within the preset bridging space to generate the corresponding fused feature representation.

[0038] In this embodiment S2, the source pair bridging vector structurally includes weight components for numerically weighting the source feature representations corresponding to each source in the source pair, and bridging state components for characterizing the bridging stability and confidence level of the source pair. The weight components are arranged according to the order of the source identifiers in the source pair, such that each weight component corresponds one-to-one with a source identifier and its corresponding source feature representation. In a source pair scenario containing two sources, the weight components may include two non-negative weight coefficients to control the contribution ratio of the two source feature representations in the fused feature representation. The bridging state components are used to encode the source pair in the current decision task cycle. The bridging state information during the period includes a stability index characterizing the magnitude of changes in the bridging configuration in adjacent iterations and a confidence index characterizing the overall performance of the source pair in the historical fusion process. The stability index can be calculated based on the statistical values ​​of the magnitude or direction of the changes in the source pair bridging vector in the most recent several iterations, and the confidence index can be calculated based on the statistical values ​​of the historical fusion quality records in multiple decision cycles. The weight component and the bridging state component together constitute the source pair bridging vector of the source pair, which is used to provide both numerical weighting basis and state reference information during the bridging fusion process.

[0039] In this embodiment S2, the preset bridging space is a vector space that carries the source feature representation and the source pair bridging vector. The principal dimension of the preset bridging space corresponds one-to-one with the dimension of the source feature representation. When embedding the source feature representation, each feature component of the source feature representation is mapped to the principal dimension coordinate position in the bridging space. On this basis, several additional dimensions are extended to store the bridging state components, so that the source pair bridging vector occupies the principal dimension position corresponding to the weight component and the extended dimension position corresponding to the bridging state component in the same bridging space. In one implementation, during the embedding process of the source feature representation in the bridging space, the extended dimension is filled with a default state or a fixed value that does not participate in the weighting operation. The weight component of the source pair bridging vector in the principal dimension and the bridging state component in the extended dimension are stored and operated according to the same coordinate system. In this way, the source feature representation and the source pair bridging vector are uniformly represented in a unified preset bridging space, providing a consistent spatial basis for subsequent field-level weighting operations and vector-level weighting operations.

[0040] In this embodiment S2, bridging fusion is based on the operation of weighted combination of source feature representations belonging to the same source pair. In the bridging fusion process, firstly, field-level linear weighting is performed on the main dimension of the preset bridging space. For each feature dimension in the source pair, the values ​​of the source feature representations corresponding to the two sources in that dimension and the weight components corresponding to the two sources are extracted. The two values ​​of the feature dimension are linearly combined according to the weight components to obtain the field-level fusion result of the fused feature representation in that dimension. This operation is repeated on the entire feature dimension set to form the feature vector after field-level fusion. In addition to field-level linear weighting, overall vector-level linear weighting can also be performed. In one implementation, overall vector-level linear weighting is uniformly scaled by the overall scaling factor determined by the bridging state components, or by first performing vector-level linear combination of the source feature representations according to the overall weight factor and then refining the weighting at the field level. This allows the bridging state components to adjust the comparability of the field-level weighting results between different iteration rounds or different decision tasks with a uniform scale.

[0041] In this embodiment S2, the combination of field-level linear weighting and overall vector-level linear weighting can be achieved through different implementation paths. In one implementation, the source feature representation is first weighted at the overall vector level, and then field-level fine-tuning weighting is performed on the main dimension, so that the overall weight factor and field-level weight components jointly determine the final value of each feature dimension. In another implementation, linear combination of each feature dimension by weight components is directly performed at the field level, and the validity label of the field-level combination result in different decision batches or the weight of participation in subsequent evaluation is adjusted by the bridging state component. In this embodiment, the fused feature representation generated by bridging fusion has the same dimensional structure as the source feature representation on the main dimension of the bridging space, and can carry additional information related to the bridging state on the extended dimension, which is used to simultaneously record the fusion result of the current source pair at the feature level and the state information at the bridging configuration level in the same bridging space.

[0042] In this embodiment S3, the decision error metric is a composite numerical index calculated based on the decision result data and the preset verification data. It is used to determine whether the threshold condition is met in each iteration and serves as the input for error attribution classification. The decision error metric is specifically composed of a main error component and a bridging regularization component. The main error component is used to measure the deviation between the decision result data and the preset verification data in the output space, and the bridging regularization component is used to measure the magnitude change, direction change, and number of sources participating in bridging in the current iteration round.

[0043] In this embodiment S3, after the fused feature representation is completed, the current value of the iteration counter and the upper limit of the maximum number of iterations are first initialized. The current value of the iteration counter is used to record the number of decision calculation rounds that have been executed for the same batch of fused feature representations. The upper limit of the maximum number of iterations is used to limit the termination boundary of the current decision process when the decision error metric does not meet the threshold condition. In each iteration, the fused feature representation is input into a preset decision model to obtain decision result data. The preset decision model can be a classification model, a regression model, or a rule-based decision engine. Its input is the numerical vector of the fused feature representation in the output space, and its output is the decision result data corresponding to the target business scenario. Then, the preset verification data corresponding to the current decision task is retrieved from the preset verification dataset. The preset verification data is structurally comparable to the decision result data in terms of output fields and dimensions. It can be manually annotated reference output, historical business results, or expected output generated according to business rules. Based on this, the decision error metric is calculated based on the decision result data and the preset verification data, and the current value of the iteration counter is incremented by one.

[0044] In this embodiment S3, the calculation process of the decision error metric includes two steps: the calculation of the principal error component and the calculation of the bridging regularization component. When calculating the principal error component, the difference measurement is performed on the corresponding fields of the decision result data and the preset verification data in the output space. In the classification output scenario, the consistency between the predicted category and the verification category can be statistically analyzed and category confidence information can be introduced to map the confidence deviation of the misclassified sample into a numerical error contribution. In the regression or numerical output scenario, the difference or deviation between the decision result data and the preset verification data can be calculated by field and normalized. In the multidimensional output scenario, the difference values ​​of each dimension can be weighted and combined according to preset weights. In any of the above implementation methods, the principal error component is a scalar used to reflect the overall deviation of the current round decision result data from the preset verification data. The larger the value of the principal error component, the higher the degree of deviation of the current round business decision from the expectation.

[0045] In this embodiment S3, the bridging regularization component is used to measure the perturbation degree of the source pair bridging vector at the bridging configuration level and the number of sources participating in bridging in the current iteration round. In specific implementation, the source pair bridging vector of the current iteration round can be compared with the source pair bridging vector of the previous iteration round or the historical reference round. The amplitude change and distribution pattern change of each source pair bridging vector on the weight component between the two rounds can be counted. The absolute value or relative rate of change of amplitude can be aggregated to obtain the bridging amplitude change index. At the same time, the number of sources participating in bridging can be counted based on the number of weight components of the source pair bridging vector that are non-zero or exceed the preset participation threshold in the current round, to obtain the bridging participation index. In another implementation, a state change index representing the degree of bridging state fluctuation can also be constructed for the bridging state component. The bridging regularization component forms a single scalar by weighting and combining one or more of the above indices according to preset rules. While maintaining the same numerical dimension as the main error component, it reflects the perturbation strength and complexity of the bridging configuration itself in the current iteration round.

[0046] In this embodiment S3, the decision error metric, as a composite numerical index, can be obtained by a weighted combination of the principal error component and the bridging regularization component. In one implementation, the decision error metric is constructed by multiplying the principal error component and the bridging regularization component by preset weight coefficients and then summing them. This allows the principal error component to dominate the deviation contribution between the decision output and the preset verification data, while the bridging regularization component dominates the fluctuation contribution of the source to the bridging vector configuration. In another implementation, the decision error metric can be composed only of the principal error component, and the bridging regularization component does not directly participate in the threshold determination but is used as a feature input for subsequent error attribution classification. In any of the above implementations, the decision error metric is compared with a preset threshold condition at the end of each iteration. The threshold condition can include a single upper limit threshold or an interval threshold. When the decision error metric is less than or falls into the normal interval specified by the threshold condition, the decision result data of the current iteration is considered to meet the accuracy requirements and the decision result data is output. When the decision error metric does not meet the threshold condition and the current value of the iteration counter has not reached the maximum number of iterations, the error attribution classification process is entered.

[0047] In this embodiment S3, the principal error component and the bridging regularization component correspond to two different sources: decision output bias and bridging configuration disturbance, respectively. The principal error component focuses on the degree of deviation of the fused feature representation from the preset verification data after being mapped to the output space by the decision model. It reflects the information loss or noise impact of the original heterogeneous data in the multi-source fusion and decision mapping link. The bridging regularization component focuses on the changes in the source to the bridging vector in terms of weight configuration and state encoding, as well as the changes in the number of sources participating in the bridging. It reflects the instability and complexity of the current iteration round at the bridging strategy level. By combining the principal error component and the bridging regularization component into a composite numerical index and analyzing their relative magnitude in the subsequent error attribution classification, the main sources of decision error can be divided into two categories: bias towards the source to the bridging vector configuration or bias towards the original heterogeneous data and its temporal relationship. Other sources of disturbance, such as slight drift of model parameters or environmental interference, are merged into the above two categories in this embodiment and indirectly reflected through the changes in the principal error component or the bridging regularization component.

[0048] In this embodiment S3, the attribution category identifier is a classification identifier determined based on the decision error metric, used to indicate that the decision error mainly originates from the source pair bridging vector configuration and from the original heterogeneous data and its temporal relationship; the error attribution classification is an operation that performs classification operation on the decision error metric and its decomposed principal error component and bridging regularization component, and combines the fusion feature representation corresponding to the iteration, the statistical value of the inter-source mutually exclusive complementary factors, and the historical fusion quality record change trend to determine the attribution category identifier; the specific judgment logic of the error attribution classification is as follows: when the ratio of the bridging regularization component to the principal error component is greater than a first preset threshold, the attribution category identifier is determined as a bridging mismatch identifier; when the ratio of the bridging regularization component to the principal error component is less than or equal to a second preset threshold, the attribution category identifier is determined as a non-bridging mismatch identifier.

[0049] In this embodiment S3, the attribution category identifier is a classification identifier determined based on the decision error metric and its decomposition results of the current iteration round. It is used to indicate the main source category of the decision error in the current round. In this technical solution, the attribution category identifier includes at least two categories: bridging mismatch identifier and non-bridging mismatch identifier. The bridging mismatch identifier is used to indicate that the decision error in the current round is mainly caused by unreasonable weight configuration of the source to the bridging vector and bridging state configuration. For example, some high-conflict sources obtain excessive weight in bridging fusion or the bridging state component fluctuates abnormally. The non-bridging mismatch identifier is used to indicate that the decision error in the current round is mainly caused by the quality problems of the original heterogeneous data itself, the accumulation of measurement noise, or the time alignment deviation. At this time, even if the source to the bridging vector configuration is adjusted, the decision error is difficult to be significantly reduced under the current data conditions. The attribution category identifier serves as the basis for the next round of bridging adjustment or data processing path selection in each iteration.

[0050] In this embodiment S3, error attribution classification is an operation that performs classification operations on the decision error metric and its decomposed principal error component and bridging regularization component, and combines the fusion feature representation corresponding to the current iteration, the statistical value of inter-source mutually exclusive complementary factors, and the trend of historical fusion quality records to determine the attribution category identifier. In specific implementation, the decision error metric of the current iteration is first decomposed into two sub-components: the principal error component and the bridging regularization component. Then, the ratio of the bridging regularization component to the principal error component is calculated to characterize the relative proportion of bridging-related perturbations in the overall decision error. At the same time, the overall score of the fusion feature representation of the current iteration is statistically analyzed. Features and sensitive field features are used to identify whether there are obvious abnormal data patterns. Then, the mutually exclusive and complementary factors between sources in the historical multiple iterations are summarized to form statistical indicators of conflict and complementarity. The time series changes of historical fusion quality records are analyzed to obtain quality change trend indicators that reflect the long-term performance of the bridging strategy. Based on this, the above values ​​are used as inputs to the error attribution classification module for classification. The classification operation can adopt a branch decision based on threshold rules or a lightweight discriminant model. In this embodiment, the error features are mapped to bridging mismatch indicators or non-bridging mismatch indicators through a predefined rule set.

[0051] In this embodiment S3, the error attribution classification judgment logic specifically includes two levels: preliminary judgment based on the ratio of the bridging regularization component to the main error component, and auxiliary judgment based on contextual statistical indicators. In the preliminary judgment level, when the ratio of the bridging regularization component to the main error component is greater than a first preset threshold, it is considered that the proportion of bridging-related disturbances in the overall error reaches a set proportion, thus determining the attribution category as a bridging mismatch identifier. When the ratio of the bridging regularization component to the main error component is less than or equal to a second preset threshold, it is considered that the decision output bias component is relatively dominant, thus determining the attribution category as a non-bridging mismatch identifier. The first preset threshold and the second preset threshold can be... Setting the same value to form a single dividing point can be used, or it can be set to the two ends of adjacent intervals to achieve a certain tolerance range. In the auxiliary judgment level, when the ratio is close to the threshold boundary or the values ​​of the main error component and the bridging regularization component are both large, the changes in the conflict sensitive field in the fusion feature representation, the proportion of conflict-dominant source pairs in the inter-source mutual exclusion complementary factor statistics, and the changing trend of historical fusion quality records can be referenced. When the proportion of conflict-dominant source pairs and the bridging quality fluctuation are both high, the judgment confidence of the bridging mismatch indicator is enhanced. When the abnormal distribution of the fusion feature representation in the input space is significant and the historical fusion quality records remain stable, the judgment confidence of the non-bridging mismatch indicator is enhanced.

[0052] In this embodiment S4, when the attribution category identifier is a bridging mismatch identifier, the update method for updating the historical fusion quality record is as follows: For the source pair corresponding to the attribution category identifier, the corresponding record entry is read from the historical fusion quality record, and a sliding time window statistical or exponential weighted update, or a combination of both, is used to map the decision error metric of the current iteration round and the bridging state component in the bridging vector of the source pair to a new fusion quality index and write it into the record entry, so as to form a fusion quality time series reflecting the change of the bridging quality of the source pair over time; when the attribution category identifier is a non-bridging mismatch identifier... When assigning identifiers, the triggering and execution strategy for data isolation and temporal realignment is as follows: when the attribution category identifiers corresponding to the same source identifier are all non-bridging mismatch identifiers within a preset number of consecutive iterations, the source identifier is added to the isolation set. When executing S2 in the subsequent steps, the source identifiers in the isolation set will no longer be included in the construction scope of the source pair bridging constraint set. The original heterogeneous data associated with the source identifier will be resampled or re-segmented according to a unified time reference to complete the temporal realignment. This is used to correct the input data and their temporal relationships participating in bridging fusion while keeping the historical fusion quality record unchanged.

[0053] In this embodiment S4, after the calculation of the decision error metric and the error attribution classification are completed, the subsequent processing path is determined based on the attribution category identifier obtained in the current iteration round and the current value of the iteration counter. If the iteration counter has not reached the preset maximum number of iterations, the bridging quality update path is executed when the attribution category identifier is a bridging mismatch identifier, and the data correction path is executed when the attribution category identifier is a non-bridging mismatch identifier. After completing the corresponding path processing, the process returns to S2 to recalculate the source pair bridging vector and fusion feature representation. When the iteration counter reaches the maximum number of iterations, the process no longer enters any of the above paths, but instead generates abnormal state data and terminates the current decision processing flow.

[0054] In this embodiment S4, after the calculation of the decision error metric and the error attribution classification are completed, the subsequent processing path is determined based on the attribution category identifier obtained in the current iteration round and the current value of the iteration counter. If the iteration counter has not reached the preset maximum number of iterations, the bridging quality update path is executed when the attribution category identifier is a bridging mismatch identifier, and the data correction path is executed when the attribution category identifier is a non-bridging mismatch identifier. After completing the corresponding path processing, the process returns to S2 to recalculate the source pair bridging vector and fusion feature representation. When the iteration counter reaches the maximum number of iterations, the process no longer enters any of the above paths, but instead generates abnormal state data and terminates the current decision processing flow.

[0055] In this embodiment S4, when the attribution category identifier is a non-bridging mismatch identifier, the error attribution classification result indicates that the decision error in the current iteration round is mainly caused by the quality of the original heterogeneous data or the temporal relationship between multiple source data. At this time, without changing the historical fusion quality record, priority is given to data isolation and temporal realignment processing of the data involved in the decision. Specifically, this includes: based on the main error component, fusion feature representation, and statistical values ​​of mutually exclusive complementary factors between sources in the current round, locating the source or source pair that makes a significant contribution to the main error in this round of decision, marking its corresponding source identifier and specific time segment as abnormal data segments, and removing or reducing the weight of these abnormal data from the input set of anchor point extraction, comparable anchor point representation calculation, and source feature representation generation in subsequent iterations, thereby... Data isolation is achieved. Simultaneously, using the temporal anchor field associated with the source identifier and a unified time reference system, the cross-source time deviation of the remaining multi-source data involved in the decision-making is re-estimated. Temporal realignment operations such as timestamp correction, interpolation resampling, or window realignment are performed, so that data from different clock sources are realigned on the unified reference time axis to the updated anchor structure and comparable anchor representation. After completing data isolation and temporal realignment, a new source feature representation is generated based on the corrected comparable anchor representation. The corrected representation and the unchanged historical fusion quality record are then re-entered into S2. At this time, S2 updates the source pair bridging vector under the same constraint solution framework, forming a data and temporal layer correction path that is different from the bridging quality update path.

[0056] In this embodiment S4, both the bridging quality update path and the data correction path are triggered before the iteration counter reaches the maximum number of iterations and return to execute S2 after processing. The difference between the two paths is that the bridging quality update path keeps the original heterogeneous data and its temporal structure unchanged, and only drives S2 to change the quality preference for different source pairs when constructing the source pair bridging constraint set and optimizing the objective function by adjusting the historical fusion quality records. This makes the source pair bridging vector converge to a more reasonable weight configuration and bridging state configuration in subsequent iterations. The data correction path keeps the historical fusion quality records unchanged, and directly changes the input basis of the anchor structure, comparable anchor representation and source feature representation by performing data isolation on abnormal sources or abnormal time periods and performing temporal realignment on the remaining data. This makes S2 act on the cleaned and aligned data set under the same constraint solving strategy. The two paths correct the same decision task from the bridging strategy layer and the data temporal layer, respectively, thus forming an alternating or collaborative error correction mechanism in subsequent iterations.

[0057] In S4 of this embodiment, when the iteration counter reaches the preset maximum number of iterations after executing any of the above correction paths and the decision error metric still does not meet the threshold condition, the current decision task is determined to be an abnormal task. In this case, it will not return to S2 to continue to execute the bridging quality update or data correction processing. Instead, it will construct abnormal state data and terminate the current decision processing flow. The abnormal state data includes at least the decision result data of the latest iteration, the corresponding decision error metric, the current iteration counter value, the most recent attribution category identifier, and diagnostic information related to the bridging configuration and data correction operation during the iteration process. The abnormal state data can be archived, alarmed, or triggered by manual intervention by the upper-layer business logic. This embodiment introduces a mechanism of combining the iteration counter with the maximum number of iterations to output abnormal states, ensuring that the decision process can be bounded and provide traceable diagnostic basis when the decision error cannot converge for a long time.

[0058] Example 2: The present invention proposes an intelligent decision data processing system that integrates multi-source heterogeneous data. It is applied to the intelligent decision data processing method that integrates multi-source heterogeneous data proposed in Example 1. It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the intelligent decision data processing method that integrates multi-source heterogeneous data in Example 1.

[0059] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for intelligent decision-making data processing that integrates multi-source heterogeneous data, characterized in that, Includes the following steps: S1. Obtain the original heterogeneous data and assign a one-to-one corresponding source identifier. Based on the source identifier, extract the anchor points of the original heterogeneous data to generate an anchor point structure. Perform anchor point normalization mapping on the anchor point structure to obtain a comparable anchor point representation. Generate the source feature representation based on the comparable anchor point representation. Among them, the anchor structure is a data structure formed by performing field-level aggregation on the original heterogeneous data associated with the source identifier based on the preset anchor fields; the comparable anchor representation is a multi-dimensional numerical representation generated based on the anchor structure according to the preset anchor field order, unified encoding rules and unified units. S2. Based on the comparable anchor point representation, calculate the source similarity factor and the source mutually exclusive complementary factor. Combine the source similarity factor, historical fusion quality record and the source mutually exclusive complementary factor to construct the source pair bridging constraint set. Solve the constraints of the source pair bridging constraint set to generate the source pair bridging vector. Based on the source pair bridging vector, perform bridging fusion on the source feature representation to generate the fusion feature representation. Among them, the inter-source mutual exclusion complementary factor is a numerical factor calculated based on the difference set and intersection set of the comparable anchor point representations corresponding to different sources; the source pair bridging constraint set is a set of constraint conditions composed of the inter-source similarity factor, the inter-source mutual exclusion complementary factor and the historical fusion quality record for the same source pair; the source pair bridging vector is a multi-dimensional vector representation that uses the source pair as an index to numerically encode the weight allocation and bridging state of any pair of source feature representations in the preset bridging space. S3. Set an iteration counter, perform decision calculations based on the fused feature representation to obtain decision result data, calculate the decision error metric between the decision result data and the preset verification data, and increment the iteration counter by one; when the decision error metric meets the threshold condition, output the decision result data and terminate the current decision data processing; when the decision error metric does not meet the threshold condition and the iteration counter has not reached the maximum number of iterations, perform error attribution classification on the decision error metric and the fused feature representation to obtain the attribution category identifier. Among them, the attribution category identifiers include bridging mismatch identifiers and non-bridging mismatch identifiers; S4. When the attribution category identifier is a bridging mismatch identifier, update the historical fusion quality record and return to execute S2. When the attribution category identifier is a non-bridging mismatch identifier, perform data isolation and temporal realignment on the source identifier and return to execute S2. When the iteration counter reaches the maximum number of iterations, output the abnormal status data and terminate the current decision data processing.

2. The intelligent decision-making data processing method for fusing multi-source heterogeneous data according to claim 1, characterized in that: In step S1, the anchor point of the original heterogeneous data is a set of field name and field value pairs extracted from the original heterogeneous data associated with the source identifier based on a preset anchor point field set. This set is used to represent different types of original heterogeneous data with a unified anchor point field dimension. The fields of the anchor point structure include at least structural anchor point fields, statistical anchor point fields, and time series anchor point fields. The anchor point normalization mapping process is a mapping operation performed on different types of anchor point fields in the anchor point structure. The comparable anchor point representation is also used to uniformly measure and compare anchor point fields between different sources and serves as input for calculating source feature representation, inter-source similarity factor, and inter-source mutually exclusive complementary factor. The source feature representation is a feature vector representation obtained by a preset feature transformation mapping operation based on the comparable anchor point representation.

3. The intelligent decision-making data processing method for fusing multi-source heterogeneous data according to claim 2, characterized in that: In S2, the inter-source similarity factor is calculated based on the comparable anchor representations corresponding to different sources and is a numerical factor used to characterize the degree of similarity between different sources at the anchor field level. The inter-source mutual exclusion and complementarity factor includes an inter-source mutual exclusion factor and an inter-source complementarity factor. The inter-source mutual exclusion factor is calculated based on the difference set of the comparable anchor representations corresponding to different sources and is a numerical factor used to represent the degree of conflict between different sources at the anchor field. The inter-source complementarity factor is calculated based on the intersection set of the comparable anchor representations corresponding to different sources and is a numerical factor used to represent the degree of complementarity between different sources at the anchor field. The inter-source mutual exclusion and complementarity factor and the inter-source similarity factor are used together to distinguish between conflict-dominated source pairs and complementarity-dominated source pairs when constructing the source pair bridging constraint set.

4. The intelligent decision-making data processing method for fusing multi-source heterogeneous data according to claim 3, characterized in that: In S2, a source pair is a combination of two different source identifiers and source feature representations that correspond one-to-one with the two source identifiers; the source pair bridging constraint set is a set of multiple constraints established for the same source pair, used to limit the value range and value relationship of each component of the source pair bridging vector under the combined effect of inter-source similarity factor, inter-source mutual exclusion and complementarity factor, and historical fusion quality record; the constraint types of the source pair bridging constraint set specifically include weight normalization constraint, mutual exclusion constraint, and historical quality constraint.

5. The intelligent decision-making data processing method for fusing multi-source heterogeneous data according to claim 4, characterized in that: In S2, the constraint solving strategy adopts a numerical solution method based on the source pair bridging constraint set to construct a constraint optimization model. The source pair bridging vector is used as the optimization variable, and the preset objective function is optimized under the premise of satisfying the source pair bridging constraint set. The preset objective function includes a similarity term based on the inter-source similarity factor and a quality term based on historical fusion quality records. The numerical solution method includes convex optimization algorithm and iterative projection algorithm, and the source pair bridging vector is iteratively updated.

6. The intelligent decision-making data processing method for fusing multi-source heterogeneous data according to claim 5, characterized in that: In S2, the source pair bridging vector includes a weight component for numerically weighting the source feature representations corresponding to each source in the source pair, and a bridging state component for characterizing the stability and confidence level of the source pair bridging. The weight component is used to control the contribution ratio of each source feature representation corresponding to each source in the source pair in the fused feature representation. The preset bridging space is a vector space that carries the source feature representations and the source pair bridging vector. The preset bridging space corresponds to the source feature representation space in terms of dimension, and is extended to include the bridging state component on the basis of the preset bridging space, for uniformly representing the source feature representations and the source pair bridging vector. The bridging fusion is an operation of performing component-weighted combination on the source feature representations belonging to the same source pair based on the source pair bridging vector.

7. The intelligent decision-making data processing method for fusing multi-source heterogeneous data according to claim 6, characterized in that: In S3, the decision error metric is a composite numerical index calculated based on the decision result data and the preset verification data. It is used to determine whether the threshold condition is met in each iteration and serves as the input for error attribution classification. The decision error metric is specifically composed of a principal error component and a bridging regularization component; wherein, the principal error component is used to measure the deviation between the decision result data and the preset verification data in the output space, and the bridging regularization component is used to measure the magnitude change, direction change and number of sources participating in bridging in the current iteration round.

8. The intelligent decision-making data processing method for fusing multi-source heterogeneous data according to claim 7, characterized in that: In step S3, the attribution category identifier is a classification identifier determined based on the decision error metric, used to indicate that the decision error mainly originates from the source pair bridging vector configuration and from the original heterogeneous data and its temporal relationship; the error attribution classification is an operation that performs classification operations on the decision error metric and its decomposed principal error component and bridging regularization component, and combines the fusion feature representation corresponding to the current iteration, the statistical value of the inter-source mutually exclusive complementary factors, and the historical fusion quality record change trend to determine the attribution category identifier; the specific judgment logic of the error attribution classification is as follows: when the ratio of the bridging regularization component to the principal error component is greater than a first preset threshold, the attribution category identifier is determined as a bridging mismatch identifier; when the ratio of the bridging regularization component to the principal error component is less than or equal to a second preset threshold, the attribution category identifier is determined as a non-bridging mismatch identifier.

9. An intelligent decision-making data processing system integrating multi-source heterogeneous data, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes a computer program to implement the intelligent decision data processing method for fusing multi-source heterogeneous data as described in any one of claims 1-8.

Images