Field mapping method of heterogeneous system

Abstract

The invention relates to the technical field of data migration, in particular to a field mapping method of a heterogeneous system. In the embodiment of the invention, the feature vectors of the fields of the new and old systems are determined through the feature extraction model, so that the mapping relation between the fields of the new and old systems is intelligently constructed, and the data migration process is greatly simplified. An immigration principal of a new system does not need to be familiar with a complex table structure of an old system, possible differences between different systems are ingeniously shielded, the whole migration process is efficient, the accuracy is powerfully guaranteed, the dependence of a traditional migration mode on artificially familiar with an old system architecture is eliminated, and the migration efficiency is improved. The migration risk caused by the cognitive difference of personnel is reduced, and a solid support is provided for stable transition of data among different systems.

Description

Technical Field

[0001] This application relates to the field of data migration techniques, and more particularly to a field mapping method for heterogeneous systems. Background Technology

[0002] In the complex process of switching between old and new core systems, data migration is a crucial and self-evident step. Data migration is not simply a matter of moving data; it requires comprehensive consideration. Among the key aspects, the mapping of business fields between the old and new systems is extremely critical. However, due to significant differences in business models, database design, and interface specifications between the old and new systems, field semantics and structures are often inconsistent, leading to the challenge of automatically identifying field mapping relationships during data migration.

[0003] Existing technologies primarily rely on manual analysis or rule-based and simple semantic similarity matching methods to generate candidate mappings by comparing field names, annotations, and data types. However, existing technologies struggle to handle complex scenarios such as field name drift, changes in units of measurement, and extended call chains caused by service splitting. Especially with continuous system iteration and dynamic changes in metadata, rule base maintenance costs are high, generalization ability is weak, and mismatches or missed matches are prone to occur. Furthermore, mapping accuracy and automation levels are limited, making it difficult to meet the requirements for highly reliable and efficient data migration. Summary of the Invention

[0004] This application provides a field mapping method for heterogeneous systems to adapt to various scenarios and improve the accuracy of field mapping.

[0005] This application provides a field mapping method for heterogeneous systems, the method comprising: A feature extraction model is used to extract feature vectors for each field in the system based on the basic information of each field, the field relationships corresponding to each business, and the sub-data information corresponding to each field within a set time period. The feature vectors for each field are then obtained. These feature vectors include at least semantic vectors, statistical vectors, and structural vectors. The system includes a new system and an old system. The feature vectors for each field include a first feature vector corresponding to each first field in the old system and a second feature vector corresponding to each second field in the new system. Based on the second feature vector of each second field and the first feature vector of each first field, determine the target first field that matches the second field; map the first data information of the target first field in the old system to the second data information of the second field corresponding to the target first field in the new system.

[0006] Furthermore, the feature extraction model includes a semantic feature extraction sub-model, a structural feature extraction sub-model, a statistical feature extraction sub-model, and a three-flow graph neural network; Before extracting the feature vector for each field in the system by using a feature extraction model, based on the basic information of each field in the system, the field relationships corresponding to each business, and the sub-data information corresponding to each field within a set time period, the method further includes: Based on the field association relationships corresponding to each business, determine the shortest path between each field and the set field; The feature extraction model is adopted to extract feature vectors for each field in the system based on the basic information of each field, the field association relationship corresponding to each business, and the sub-data information corresponding to each field within a set time period. The feature vectors corresponding to each field include: Using the semantic extraction sub-model, the semantic vector corresponding to each field is determined based on the basic information of each field in the system; Using the aforementioned structural feature extraction sub-model, the structural vector corresponding to each field is determined based on the shortest path between each field and a set field; Using the aforementioned statistical feature extraction sub-model, the statistical vector corresponding to each field is determined based on the sub-data information corresponding to each field within a set time period; The three-flow graph neural network is used to update the semantic vector, structural vector, and statistical vector corresponding to each field; Based on the weights corresponding to each field, the updated semantic vector, statistical vector, and structural vector of each field are weighted and fused to obtain the feature vector corresponding to each field.

[0007] Furthermore, the step of using the semantic extraction sub-model to determine the semantic vector corresponding to each field based on the basic information of each field in the system includes: For each field, the semantic extraction sub-model concatenates the basic information corresponding to the field into a text sequence in a preset order, and uses a pre-configured business domain vocabulary to segment the entities in the text sequence to obtain a segmented intermediate sequence; the semantic extraction sub-model performs semantic recognition on the intermediate sequence to obtain the semantic vector corresponding to the field.

[0008] Furthermore, the step of using the structural feature extraction sub-model to determine the structural vector corresponding to each field based on the field association relationship corresponding to each business includes: For each field, the structural feature extraction sub-model constructs a shortest path vector based on the shortest path corresponding to that field, where each dimension of the shortest path vector corresponds to a node in the shortest path; the structural feature extraction sub-model determines a position vector based on the position of each node in the shortest path; the structural feature extraction sub-model extracts structural features by concatenating the position vector with the shortest path vector to obtain a concatenated vector, thus obtaining the structural vector corresponding to that field.

[0009] Furthermore, the step of using the statistical feature extraction sub-model to determine the statistical vector corresponding to each field based on the sub-data information corresponding to each field within a set time period includes: For each field, the input layer of the statistical feature extraction sub-model determines the initial statistical vector corresponding to the field by counting the number of sub-data information corresponding to each field in each preset interval within the set time period. Multiple convolutional layers of the statistical feature extraction sub-model convolve the initial statistical vector. The pooling layer of the statistical feature extraction sub-model performs global max pooling on the output of each convolutional layer to obtain each intermediate statistical vector. The output layer of the statistical feature extraction sub-model concatenates each intermediate statistical vector into a joint vector and determines the joint vector as the statistical vector corresponding to the field.

[0010] Furthermore, the basic information of each field, the field association relationship corresponding to each business, and the sub-data information corresponding to each field within a set time period are stored in the heterogeneous graph; The step of using the three-flow graph neural network to update the semantic vector, structural vector, and statistical vector corresponding to each field includes: Add the semantic vector, structural vector, and statistical vector corresponding to each field to the node corresponding to each field in the heterogeneous graph; The updated heterogeneous graph is input into the three-flow graph neural network; the three-flow graph neural network updates the semantic vector of each field in the heterogeneous graph based on whether the two fields corresponding to each edge belong to the same business entity; updates the statistical vector corresponding to each field in the heterogeneous graph based on the difference vector between the statistical vector of each field and the preset global distribution primitive vector; and updates the structure vector of each field based on the hop count corresponding to the structure vector of each field.

[0011] Furthermore, the method also includes: Obtain the preset constraint vector corresponding to each field, which is determined based on the field's anomaly alarm label and / or business label; The preset constraint vector corresponding to each field is concatenated with the corresponding feature vector, and the feature vector is updated using the concatenated vector.

[0012] Further, determining the target first field that matches each second field based on the second feature vector of each second field and the first feature vector of each first field includes: For each second field, the similarity between the second field and each first field is determined based on the second feature vector of the second field and the first feature vector of each first field; the first field whose similarity exceeds the similarity threshold is determined as the target first field that matches the second field.

[0013] Furthermore, if there is no corresponding first field with a similarity exceeding the similarity threshold, the method further includes: Determine the difference between the highest similarity among the similarities between the second field and each of the first fields and the similarity threshold; If the difference is within a preset range, a work order awaiting review is generated, and a manual review notification is triggered.

[0014] Furthermore, the method also includes: Determine the number of pending work orders currently generated; If the quantity reaches a preset quantity threshold, then a fine-tuning task for the feature extraction model is initiated based on the second field corresponding to the work order to be reviewed and the first field with the highest similarity to the second field.

[0015] Furthermore, the training process of the three-flow graph neural network includes: Obtain sample pairs, each sample pair including an initial sample semantic vector, an initial sample structure vector, and an initial sample statistical vector corresponding to two sample fields respectively, whether the two sample fields are labels of the same node, and the boundary value of the sample pair; wherein, the boundary value is determined based on the number of sample fields that have a direct correlation with any sample field in the sample pair; Using a three-flow graph neural network to be trained, the initial sample semantic vector, initial sample structure vector, and initial sample statistical vector corresponding to the two sample fields included in the sample pair are updated respectively to obtain the sample semantic vector, sample structure vector, and sample statistical vector corresponding to the two fields respectively; Based on the sample semantic vectors corresponding to the two fields respectively, the semantic distance is determined, and based on the relationship between the semantic distance and the boundary value, and whether the two sample fields are labels of the same node, the semantic loss value is determined. The statistical loss value is determined based on the label, the sample statistical vectors corresponding to the two fields, and the boundary value. The structural loss value is determined based on the sample structure vector corresponding to the two sample fields included in the sample pair and the boundary value, respectively. Based on the semantic loss value, the statistical loss value, and the structural loss value, a total loss value is determined, and the parameters of the three-flow graph neural network are adjusted according to the total loss value.

[0016] In this embodiment, feature vectors of fields in the old and new systems are determined through a feature extraction model, thereby intelligently constructing a mapping relationship between fields in the old and new systems, greatly simplifying the data migration process. It eliminates the need for the person in charge of migrating to the new system to familiarize themselves with the complex table structure of the old system, cleverly shielding potential differences between the two systems. This makes the entire migration process not only efficient but also highly accurate, eliminating the reliance on manual familiarization with the old system architecture in traditional migration methods, reducing migration risks caused by differences in personnel understanding, and providing solid support for a smooth transition of data between different systems. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic diagram of a field mapping process in a heterogeneous system provided in an embodiment of this application; Figure 2 A flowchart of a three-flow graph neural network provided in an embodiment of this application; Figure 3 An overall flowchart provided for embodiments of this application; Figure 4 This is a schematic diagram comparing the embodiments of this application with the prior art; Figure 5 A schematic diagram of a field mapping device for a heterogeneous system provided in an embodiment of this application; Figure 6 This is a schematic diagram of an electronic device structure provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] In the complex process of switching between old and new core systems, data migration is a crucial and self-evident step. Data migration is not simply a matter of moving data; it requires comprehensive consideration. Among the key aspects, the mapping of business fields between the old and new systems is extremely critical.

[0021] Because the old and new systems differ in architecture and functional focus, the definitions and data types of various business fields may also differ. Only by carefully sorting out and establishing accurate mapping relationships between business fields can we ensure that data can be completely, accurately, and effectively migrated from the old system to the new system during the data migration process. This will guarantee that all types of business operations can run smoothly and stably after the new system goes live, and prevent business errors caused by data mismatch or loss.

[0022] Most existing technologies use a three-layer chain of "offline extraction - rule mapping - real-time verification" to handle data mapping tasks between old and new systems. In the offline stage, the two versions of Data Definition Language (DDL), interface descriptions and version change logs are first parsed to generate a unified metadata snapshot at the table-field granularity. Then, with the help of manually maintained thesaurus, regularization templates and service domain tags, differences such as capitalization, camelCase and abbreviations are normalized, and candidate mapping pairs are ranked according to table name similarity, version number and call context.

[0023] During the alignment refinement process, many implementations have begun to introduce semantic embedding technology: concatenating field names, Chinese comments, and even service code snippets into sequences, encoding them through a lightweight pre-trained language model, and then calculating vector distances to improve the automatic hit rate in cross-version naming drift scenarios, replacing pure rule-based scoring. Simultaneously, to address numerical scale differences caused by field type upgrades or changes in units of measurement, the system performs binning statistics on real-time sampled data, tracks histogram, mean, and extreme value changes, and detects distribution drift using a historical threshold library. Upon detecting anomalies, the platform triggers incremental comparisons or manual reviews via message middleware to ensure that accounting and other deviations are not caused by field mismatches during the migration process.

[0024] To unify the traceability chain, existing solutions often integrate database foreign keys, microservice API calls, and scheduling dependencies into a directed graph, and then use random walks or graph embedding algorithms to infer indirectly related fields. After the mapping goes live, the system continuously records the version lineage, uses scheduled offline jobs to statistically analyze mapping accuracy and the frequency of manual intervention, and automatically adjusts rule parameters or reruns the matching process based on threshold logs, forming a rolling iterative migration assurance mechanism.

[0025] However, existing technologies still have the following technical problems: 1. Existing cross-system field mapping methods remain fragmented in practice: metadata extraction, rule matching, semantic encoding, statistical verification, and link analysis belong to different components, loosely connected by offline files or message channels, lacking a unified representation space and end-to-end feedback mechanism. When adding fields to an old system under high concurrency scenarios or when a new system extends the call chain due to service splitting, manual updates to the rule base and the thesaurus often cannot be completed synchronously, leading to a sharp drop in candidate field recall.

[0026] 2. Semantic vectors are mostly generated based on textual clues and lack the ability to perceive numerical scale drift such as the expansion of measurement units and the upgrading of field types. The vector distance is no longer consistent with the original business relationship.

[0027] 3. The structural analysis layer focuses on single-hop dependencies or random walks, lacking sensitivity to multi-hop links that arise after database sharding and table partitioning. It easily overlooks key fields hidden in deep call chains. Monitoring metrics for each modality are scattered across ETL jobs, queue alarms, and offline reports, lacking centralized backtracking. As a result, when mapping mismatches cause fluctuations in accounting, the problem can only be located manually afterward, making it difficult to meet continuous delivery requirements in terms of response time and repair costs.

[0028] In summary, with the old and new systems running in parallel and continuously evolving, how to establish a unified multimodal alignment mechanism for data migration has become a technical problem to be solved.

[0029] Based on this, in order to adapt to various scenarios and improve the accuracy of field mapping, this application provides a field mapping method, apparatus, device and medium for heterogeneous systems, which enables semantic information, value domain statistics and call chain structure to jointly drive field matching in the same vector space, and continuously adaptively update through online feedback, thereby achieving automatic, real-time and high-precision mapping of key fields between new and old systems without interrupting business, thereby improving the accuracy and efficiency of data migration between new and old systems.

[0030] In this embodiment, a feature extraction model is used to extract feature vectors for each field in the system based on the basic information of each field, the field associations corresponding to each business, and the sub-data information corresponding to each field within a set time period. The feature vectors for each field are then obtained. Each feature vector includes at least a semantic vector, a statistical vector, and a structural vector. The system includes a new system and an old system. The feature vectors for each field include a first feature vector corresponding to each first field in the old system and a second feature vector corresponding to each second field in the new system. Based on the second feature vector of each second field and the first feature vector of each first field, a target first field matching that second field is determined. The first data information of the target first field in the old system is mapped to the second data information of the second field corresponding to the target first field in the new system.

[0031] Example 1: Figure 1 This application provides a schematic diagram of a field mapping process for a heterogeneous system, which includes: S101: Using a feature extraction model, based on the basic information of each field in the system, the field association relationship corresponding to each business, and the sub-data information corresponding to each field within a set time period, feature vectors are extracted for each field in the system to obtain the feature vector corresponding to each field; wherein, the feature vector includes at least semantic vector, statistical vector, and structural vector; the system includes a new system and an old system; the feature vector corresponding to each field includes a first feature vector corresponding to each first field in the old system and a second feature vector corresponding to each second field in the new system.

[0032] The field mapping method for heterogeneous systems provided in this application is applied to electronic devices, such as PCs and servers.

[0033] In heterogeneous system data migration scenarios, especially during the replacement of core systems in the financial sector, there are significant differences between the old and new systems in terms of database structure, interface protocols, and business semantics. Traditional field mapping relies on manual rules or simple string matching, which is insufficient to handle complex situations such as field name drift, value domain changes, and call chain reconstruction. For example, the old system might name "Customer Number" CUST_ID, while the new system uses CLIENT_NO; similarly, "Amount" is LOAN_AMT (INT type) in the old system, but DISBURSE_AMOUNT (DECIMAL(20,4)) in the new system. If the name or type alone cannot determine whether the two fields are the same, a richer context must be considered.

[0034] Based on this, this embodiment proposes a field feature vector extraction method based on a feature extraction model, which aims to comprehensively characterize field features from three dimensions: semantics, statistics, and structure, thereby achieving high-precision automatic mapping.

[0035] Specifically, in this embodiment, the electronic device can obtain basic information of each field in the system, the field relationships corresponding to each service, and the sub-data information corresponding to each field within a set time period. Then, using a feature extraction model, based on the basic information of each field in the system, the field relationships corresponding to each service, and the sub-data information corresponding to each field within the set time period, feature vectors are extracted for each field in the system, resulting in a feature vector corresponding to each field in the system. This system can be a new system or an old system. The new system uses the above method to determine the feature vector corresponding to each field, and the old system also uses the above method to determine the feature vector corresponding to each field.

[0036] In the following embodiments, "system" is used to refer to both the new system and the old system.

[0037] In one possible implementation, the electronic device can load the system's schema metadata, which includes, but is not limited to, the table name of each data table contained in the system, and basic information about each field contained in each data table, including but not limited to the English name. Chinese labels Business Notes Microservice identifier and the version number wait.

[0038] The electronic device can also acquire the system's real-time log stream and extract sub-logs within a set time period from the real-time log stream, for example, the set time period is the last 7 days. The electronic device obtains the sub-data information corresponding to each field contained in the sub-log, where the sub-data information corresponding to any field can be each value corresponding to that field in the sub-log.

[0039] In this embodiment of the application, the electronic device can use a feature extraction model to extract feature vectors for each field in the system based on the basic information of each field in the system, the field association relationship corresponding to each business, and the sub-data information corresponding to each field within a set time period, so as to obtain the feature vector corresponding to each field.

[0040] Specifically, for each field in each system, the electronic device can input the basic information of the field, the field association relationship corresponding to each business, and the sub-data information corresponding to the field within a set time period into the feature extraction model. The feature extraction model then extracts the feature vector to obtain the feature vector corresponding to the field.

[0041] The feature vector extraction process includes, but is not limited to: semantic feature extraction for each field to determine the semantic vector of each field; statistical feature extraction for each field to determine the statistical vector of each field; structural feature extraction for each field to determine the structural vector of each field; and finally, the semantic vector, statistical vector, and structural vector are weighted and fused in a progressive order of semantic, statistical, and structural to generate the feature vector of each field. The feature vector of each field not only integrates the field's naming, value range, and call chain information, but also implicitly encodes the business stage and system origin, laying a high-quality input foundation for subsequent graph neural network learning.

[0042] Based on the above method, the electronic device determines the first feature vector of each first field in the old system and the second feature vector of each second field in the new system.

[0043] S102: Based on the second feature vector of each second field and the first feature vector of each first field, determine the target first field that matches the second field corresponding to each second field; map the first data information of the target first field in the old system to the second data information of the second field corresponding to the target first field in the new system.

[0044] In this embodiment of the application, after determining the first feature vector corresponding to each first field in the old system and the second feature vector corresponding to each second field in the new system, the electronic device can perform field matching based on the first feature vector corresponding to each first field and the second feature vector corresponding to each second field.

[0045] Specifically, the electronic device can identify the target first field that best matches each second field based on the similarity between the second feature vector of each second field and the first feature vector of all first fields.

[0046] For example, the electronic device can employ an algorithm for calculating the similarity between vectors to determine the similarity between each second feature vector and each first feature vector. For each second field, the electronic device can determine the maximum similarity corresponding to the second feature vector of that second field, and determine the first field of the first feature vector corresponding to the maximum similarity as the target first field that best matches that second field. This algorithm can be a cosine similarity algorithm, Euclidean distance algorithm, Manhattan distance algorithm, etc., and is not limited here.

[0047] After the electronic device determines the target first field that best matches each second field in the new system in the old system, it can accurately migrate the data from the old system to the new system based on the correspondence between each second field and the corresponding target first field.

[0048] Specifically, for each second field, the electronic device can obtain the first data information of the target first field that matches the second field in the old system, and map the first data information to the second data information of the second field in the new system.

[0049] In this embodiment, the data migration process is greatly simplified by intelligently constructing mapping relationships. The migration manager for the new system does not need to familiarize themselves with the complex table structure of the old system, cleverly shielding potential differences between systems. This makes the entire migration process not only efficient but also highly accurate, eliminating the reliance on manual familiarization with the old system architecture in traditional migration methods. It reduces migration risks caused by differences in personnel understanding and provides solid support for the smooth transition of data between different systems.

[0050] Example 2: To adapt to various scenarios and improve the accuracy of field mapping, based on the above embodiments, in this application embodiment, the feature extraction model includes a semantic feature extraction sub-model, a structural feature extraction sub-model, a statistical feature extraction sub-model, and a three-flow graph neural network; Before extracting the feature vector for each field in the system by using a feature extraction model, based on the basic information of each field in the system, the field relationships corresponding to each business, and the sub-data information corresponding to each field within a set time period, the method further includes: Based on the field association relationships corresponding to each business, determine the shortest path between each field and the set field; The feature extraction model is adopted to extract feature vectors for each field in the system based on the basic information of each field, the field association relationship corresponding to each business, and the sub-data information corresponding to each field within a set time period. The feature vectors corresponding to each field include: Using the semantic extraction sub-model, the semantic vector corresponding to each field is determined based on the basic information of each field in the system; Using the aforementioned structural feature extraction sub-model, the structural vector corresponding to each field is determined based on the shortest path between each field and a set field; Using the aforementioned statistical feature extraction sub-model, the statistical vector corresponding to each field is determined based on the sub-data information corresponding to each field within a set time period; The three-flow graph neural network is used to update the semantic vector, structural vector, and statistical vector corresponding to each field; Based on the weights corresponding to each field, the updated semantic vector, statistical vector, and structural vector of each field are weighted and fused to obtain the feature vector corresponding to each field.

[0051] In heterogeneous system field mapping tasks, relying solely on single-dimensional features often fails to address complex scenarios such as naming ambiguity, structural drift, or abrupt changes in data distribution. Therefore, this application proposes a method that integrates semantic, structural, and statistical features, and performs collaborative optimization through a three-flow graph neural network. This results in a feature vector that not only focuses on the business relationships between fields but also introduces a dynamic weighting mechanism for adaptive fusion of multi-source features, thereby significantly improving the robustness and generalization ability of cross-system field matching.

[0052] Specifically, in this embodiment, the feature extraction model includes a semantic feature extraction sub-model, a structural feature extraction sub-model, a statistical feature extraction sub-model, and a three-flow graph neural network. The semantic feature extraction sub-model can be a Transformer model fine-tuned using domain data, the statistical feature extraction sub-model can be a convolutional neural network, and the structural feature extraction sub-model can be a Bidirectional Long Short-Term Memory (Bi-LSTM) network.

[0053] Electronic devices can determine the shortest path between each field and a designated field based on the field relationships corresponding to each service. For example, electronic devices can construct a heterogeneous graph based on the field relationships corresponding to each service, and determine the shortest path between each field and a designated field based on this heterogeneous graph. The designated field can be one or more, and can be configured by technical personnel.

[0054] The field relationships corresponding to each business function can be stored in the business view. API call chain and database foreign key relationships In the middle. Electronic devices according to this business view. API call chain and database foreign key relationships The relationships between fields corresponding to each business transaction and the order of fields within the business transaction are stored in the database to generate two levels of directed edges, including but not limited to table-field edges. With field - field edge ,in, Indicates the name of table i. Indicates the name of field j. Indicates the name of field u. Indicates the name of field v. Indicate business view The number of hops between field u and field v.

[0055] Among them, business view Indicates the fields required for each business function and the relationships between these fields; API call chain. Indicates the calling relationship between each field and the database foreign key relationship. Indicates different names for the same thing in different data tables.

[0056] In the embodiments of this application, the nodes in the heterogeneous graph include, but are not limited to, field nodes and data table nodes, and the edges in the heterogeneous graph include, but are not limited to, the belonging edge of a field to a certain data table, the sequential edge of a field being continuously called in the same business, and the reference edge formed by a field between different data tables through foreign keys or logical associations.

[0057] Before formally extracting feature vectors, the electronic device can execute Dijkstra's algorithm for any field to calculate the shortest path from that field to a specified field. This shortest path contains only the number of hops and also records the sequence of nodes traversed in between. For example, the shortest path could be field 1 → field 2 → field 3.

[0058] After determining the shortest path for each field, the electronic device can concurrently invoke three dedicated sub-models to extract features from each second field in the new system and each first field in the old system. Specifically, the electronic device invokes the semantic extraction sub-model to determine the semantic vector corresponding to each field based on the basic information of each field in the system; the electronic device invokes the structural feature extraction sub-model to determine the structural vector corresponding to each field based on the shortest path between each field and a set field; and the electronic device invokes the statistical feature extraction sub-model to determine the statistical vector corresponding to each field based on the sub-data information corresponding to each field within a set time period.

[0059] Subsequently, the electronic device invokes a three-flow graph neural network to update the semantic vector, structural vector, and statistical vector corresponding to each field; and according to the weights corresponding to each field, it performs weighted fusion of the updated semantic vector, statistical vector, and structural vector of each field to obtain the feature vector corresponding to each field.

[0060] To adapt to various scenarios and improve the accuracy of field mapping, based on the above embodiments, in this embodiment, the step of using the semantic extraction sub-model to determine the semantic vector corresponding to each field based on the basic information of each field in the system includes: For each field, the semantic extraction sub-model concatenates the basic information corresponding to the field into a text sequence in a preset order, and uses a pre-configured business domain vocabulary to segment the entities in the text sequence to obtain a segmented intermediate sequence; the semantic extraction sub-model performs semantic recognition on the intermediate sequence to obtain the semantic vector corresponding to the field.

[0061] In field mapping tasks across heterogeneous systems, accurate extraction of semantic information is fundamental to achieving high matching rates. However, traditional natural language processing methods often treat field names or annotations as isolated text, ignoring their deeper meaning within specific business contexts. Therefore, in this embodiment, the electronic device acquires basic information corresponding to a field, which includes multiple sub-basic information. Before semantic feature recognition, these multiple sub-basic information can be concatenated into a text sequence according to a preset order. Then, a pre-configured business domain lexicon is used to segment the text sequence to ensure that field semantics remain consistent and comparable during cross-system migration.

[0062] The sub-basic information for each field includes, but is not limited to, the field's English name, Chinese label, business comments, microservice identifier, and version number. It may also include the business module it belongs to, data type, whether it is a primary key, whether it is a sensitive field, historical change records, and descriptions of associated business rules.

[0063] To avoid information fragmentation, the above-mentioned basic information can be structurally concatenated in a preset order to obtain a text sequence. For example, the concatenation order can be: English field name, Chinese label, business comment, microservice identifier, online version number, module, data type, primary key identifier, sensitive identifier, business rule summary, and historical change description, etc. For example, a certain field concatenated to form the following text sequence: CUST_ID; customer number; customer management; VARCHAR(32); is a primary key; non-sensitive; uniquely identifies enterprise customers; cannot be null; renamed from CUST_NO in Q2 2023. The text sequence in this embodiment not only preserves the integrity of the original information, but also provides a stable input structure for subsequent segmentation.

[0064] In this embodiment, a pre-configured business domain lexicon can also be invoked to perform entity recognition and segmentation on the text sequence. This lexicon is not a general vocabulary database, but consists of hundreds of core terms from vertical fields such as finance, manufacturing, and retail. This business domain lexicon is dynamically updated by regularly crawling industry standard documents, internal work order systems, and change logs, and can also use a trie tree structure to achieve efficient matching. During the segmentation process, the longest word can be matched first—when encountering "loan date," it will not be split into "loan" and "date," but will be retained as a whole semantic unit. Special symbols such as numbers, units, and version numbers can also be standardized, such as converting "VARCHAR(32)" to "string_32-bit" and "DECIMAL(18,2)" to "amount_2 decimal places," etc.

[0065] After segmentation, an intermediate sequence consisting of domain entities and standardized symbols is obtained. The length of this intermediate sequence is typically controlled to be within 32 tokens, preserving key semantics while avoiding redundant noise. This intermediate sequence is then fed into a semantic extraction sub-model. This sub-model can be based on the BERT-based architecture, but it is fine-tuned during the pre-training phase using a large sample corpus containing domain entities. The fine-tuning tasks include: field naming prediction, business tag classification, and synonym field discrimination. Through these fine-tuning tasks, the semantic extraction sub-model deeply learns the semantic relationships and contextual dependencies between business terms.

[0066] In addition, the basic information splicing and entity segmentation process in the above embodiments can be completed by electronic devices. After obtaining the intermediate sequence, the electronic devices input the intermediate sequence into the semantic extraction sub-model. The basic information splicing and entity segmentation process can also be completed by the semantic extraction sub-model.

[0067] During the inference phase, the semantic extraction sub-model can perform bidirectional context modeling on the intermediate sequence to obtain the context embedding of each token, and based on the context embedding of each token, obtain a semantic vector. Each entity in the intermediate sequence is a token. In this embodiment, the semantic vector not only encodes the literal meaning of the field, but also implicitly integrates its business role, data constraints, and historical evolution trajectory. For example, although "LOAN_AMT" and "DISBURSE_AMOUNT" have different names, their semantic vectors are highly similar in space because they share core entities such as "loan," "amount," and "loan disbursement." Similarly, although "CREATE_TIME" and "UPDATE_TIME" both contain "time," they are effectively distinguished by the attention mechanism of the semantic extraction sub-model due to their different business actions.

[0068] Specifically, in cross-system data migration scenarios, order fields in different versions often have names that appear different on the surface but are consistent in business logic, such as order-amt and total-amount. Furthermore, the same name can have semantic drift issues, with create-time being used as the order placement time in the old system and then reloaded as the creation time in the new system.

[0069] Based on this, first, the English names of the fields are... Chinese labels Business Notes Microservice identifier and the version number Assemble in order The sequence fully covers the semantic clues and version context of the field.

[0070] Then The input q is fed into the Transformer semantic extraction sub-model, which has been fine-tuned on the ERP-CRM industry corpus (including change logs, work order comments, launch announcements, etc.). This enables the semantic extraction sub-model to learn the contextual dependencies of domain-specific terms and correct common industry abbreviations. To further highlight key business entities, token-level entity gating weights are implemented. Filter tokens related to the loan receipt number, customer ID, and account number, then adjust them to generate a semantic vector for field q. .

[0071] For example, the semantic vector of field q satisfies the following formula:

[0072] in, Indicates the length of the intermediate sequence; Indicates the first in the intermediate sequence One Token; This represents the corresponding context vector; Indicates the entity gating weight; Represents a nonlinear activation function; and This represents the linear projection parameters obtained during training; Indicates the normalization factor; Represents the basic semantic vector.

[0073] Using the above formula, the semantic extraction sub-model can cluster cross-version synonymous fields into similar vector positions while suppressing irrelevant descriptive noise.

[0074] Furthermore, the old system frequently uses "Debit Credit Flag" in the first field, while the new system often uses abbreviations like "Dr Cr Flag" in the second field. Additionally, status terms such as "frozen" and "not yet settled" are mixed into the field comments. If a general semantic model is directly used for semantic feature extraction, these entities are often broken down into scattered sub-words, resulting in the dilution of cognate information. Therefore, this application's embodiments use a business domain lexicon. Inject a tokenizer to atomically segment entities in the text sequence and set tokens for these entities. Position Marker It can not only record the order of entities in field descriptions, but also enable the semantic extraction sub-model to distinguish subtle differences in meaning such as pending loan disbursement and completed loan disbursement. This allows the final semantic vector to maintain a high degree of consistency in interpreting business stages across system scenarios, further reducing the reliance on manual alignment rules.

[0075] To adapt to various scenarios and improve the accuracy of field mapping, based on the above embodiments, in this embodiment, the step of using the structural feature extraction sub-model to determine the structural vector corresponding to each field according to the field association relationship corresponding to each business includes: For each field, the structural feature extraction sub-model constructs a shortest path vector based on the shortest path corresponding to that field, where each dimension of the shortest path vector corresponds to a node in the shortest path; the structural feature extraction sub-model determines a position vector based on the position of each node in the shortest path; the structural feature extraction sub-model extracts structural features by concatenating the position vector with the shortest path vector to obtain a concatenated vector, thus obtaining the structural vector corresponding to that field.

[0076] In complex business systems, fields may have similar semantics, but their positions and functions in the business process can be quite different. For example, "Customer ID" is an input parameter in the customer registration service, a query condition in the risk control service, and an aggregation dimension in the reporting service. Relying solely on semantic or statistical features can easily lead to incorrect matching of fields with the same name for different roles. Based on this, the embodiments of this application can construct and determine the structure vector of a field based on the shortest path of the field, which is used to accurately characterize the structural role of the field in the business topology.

[0077] After obtaining the shortest path, the structural feature extraction sub-model constructs a shortest path vector for each field based on the shortest path corresponding to that field. Each dimension of the shortest path vector corresponds to a node in the shortest path. The structural feature extraction sub-model determines the position vector based on the position of each node in the shortest path. The structural feature extraction sub-model then performs structural feature extraction by concatenating the position vector and the shortest path vector to obtain the concatenated vector, thus obtaining the structural vector corresponding to that field.

[0078] For example, the shortest path vector is represented as ,in This is the field corresponding to the shortest path. Set a specific field. Each node It is given a learnable embedding vector. To preserve path sequence information, the structural feature extraction sub-model is based on each node. At the position in the shortest path, generate a position vector. Next, the structural feature extraction sub-model concatenates the position vector with the shortest path vector to obtain the structural vector corresponding to the field. This structural feature extraction sub-model can simultaneously capture information flows from front to back (from field F to the specified field) and from back to front (back from the specified field to field F), thereby comprehensively understanding the contextual role of the field in the path.

[0079] During the period of mixed operation of old and new systems, the call chain depth of the same business request may increase due to database sharding and service decoupling, resulting in inconsistent dependency hop counts between fields. To ensure that the structure flow can accurately determine the importance of self-loops and long chains across tables in the local table for mapping, this sub-step will use the shortest path Updated to a sequence of tuples including jump distances. Then adjust the jump distance. Mapping to position vectors via sine and cosine functions with shortest path The concatenation is performed to obtain a concatenated vector, which is then fed into the Bi-LSTM structural feature extraction sub-model to simultaneously capture forward and backward dependencies, thus obtaining a structural vector. The final output structure vector It can remain sensitive to structural differences between old and new systems, leaving room for structural decision-making in alignment.

[0080] To adapt to various scenarios and improve the accuracy of field mapping, based on the above embodiments, in this embodiment, the step of using the statistical feature extraction sub-model to determine the statistical vector corresponding to each field based on the sub-data information corresponding to each field within a set time period includes: For each field, the input layer of the statistical feature extraction sub-model determines the initial statistical vector corresponding to the field by counting the number of sub-data information corresponding to each field in each preset interval within the set time period. Multiple convolutional layers of the statistical feature extraction sub-model convolve the initial statistical vector. The pooling layer of the statistical feature extraction sub-model performs global max pooling on the output of each convolutional layer to obtain each intermediate statistical vector. The output layer of the statistical feature extraction sub-model concatenates each intermediate statistical vector into a joint vector and determines the joint vector as the statistical vector corresponding to the field.

[0081] In actual business operations, the value range distribution of fields often changes dynamically with time, region, or promotional activities. Ignoring such statistical characteristics and relying solely on static information can easily lead to field mismatches. Therefore, this application embodiment also extracts the statistical features of fields to capture the time distribution patterns of field values.

[0082] Specifically, for each field, the input layer of the statistical feature extraction sub-model sets the number of sub-data information corresponding to each field in each preset interval within a set time period, and determines the initial statistical vector corresponding to the field; multiple convolutional layers of the statistical feature extraction sub-model convolve the initial statistical vector respectively; the pooling layer of the statistical feature extraction sub-model performs global max pooling on the output of each convolutional layer to obtain each intermediate statistical vector; the output layer of the statistical feature extraction sub-model concatenates each intermediate statistical vector into a joint vector, and determines the joint vector as the statistical vector corresponding to the field.

[0083] In one possible implementation, for each field, a setting duration can be enabled in the real-time synchronization log. The scrolling window extracts the set of sub-data information corresponding to the field. The value range of this field is determined based on the business enumeration table. A closed interval, such as Calculate the initial statistical vector for this field. The initial statistical vector is the normalized discrete frequency vector of the sub-data information in each preset interval.

[0084] The following formula can be used to determine the value corresponding to each preset interval in the initial statistical vector:

[0085] in: Representation field The Individual data information; Indicates the total number of samples; Indicates the first The upper bound of the minimum value of a preset interval; This represents a Boolean indicator function that takes the value 1 if the condition is met, and 0 otherwise. The first statistic in the initial statistical vector The values ​​corresponding to each preset interval.

[0086] In this embodiment, while the initial statistical vector is concise, it lacks the ability to jointly model local patterns and global trends. Therefore, the statistical feature extraction sub-model in this embodiment is a multi-branch convolutional neural network. This sub-model contains three parallel one-dimensional convolutional layers with kernel sizes of 1, 3, and 5, all with a stride of 1 and the same padding method to ensure the output length matches the input. Specifically, the convolutional layer with a kernel size of 1 is used to capture single-point anomalies, such as a sudden large fluctuation in a certain field at a certain moment; the convolutional layer with a kernel size of 3 is used to focus on local fluctuations, such as a continuous increase in amount over three days; and the convolutional layer with a kernel size of 1 is used to learn long-term trends, such as a peak at the end of the month. Each convolutional layer is followed by a ReLU activation function to enhance nonlinear expressive power.

[0087] After each convolutional layer output, the statistical feature extraction sub-model performs global max pooling on each convolutional layer output to obtain each intermediate statistical vector.

[0088] For example, during data migration, the same numerical field, often of type INT in the old system, is expanded to DECIMAL after the upgrade; or the unit of measurement changes from piece to box, causing the value range to expand exponentially. Relying solely on single-scale statistics will make it difficult to simultaneously capture mainstream values ​​and long-tail anomalies. To address this technical challenge, the embodiments of this application address the field... initial statistical vector Implement adaptive binning, partitioning point set It will automatically stretch or compress based on the field type (integer, amount, percentage, etc.). Then, one-dimensional convolutional kernels with kernel sizes of 1, 3, and 5 are set in parallel. The system captures single-point peaks, local fluctuations, and global trends respectively. Then, it performs global max pooling on the results of each convolution and concatenates them to obtain a statistical vector. .

[0089] For example, statistical vectors satisfy the following formula:

[0090] in, Initial statistical vector; One-dimensional convolution kernels with kernel sizes of 1, 3, and 5; Global max pooling operation; Vector concatenation; Joint vector.

[0091] Based on the above multi-scale design, the model is guaranteed to have the ability to detect extreme distributional anomalies such as sudden increases in amount, thereby providing more reliable statistical evidence in the mapping stage.

[0092] Furthermore, in this embodiment, the value range of each field may drift due to version switching, business seasonality, or interface changes. If not marked in advance, this can lead to misjudgment of statistical similarity during subsequent graph alignment. Therefore, in this embodiment, statistical vectors can be... With threshold library For comparison, the threshold library is automatically generated from historically stable distributions and is periodically refreshed to follow business peak and trough cycles. When the distribution offset rate is calculated... When the threshold is exceeded, the electronic device immediately generates an anomaly flag. And write the tag back to the node attribute field. This is a preset threshold.

[0093] This allows for stricter constraints on outlier fields during the third-stream GNN stage, preventing low-confidence information from interfering with the main mapping results.

[0094] To adapt to various scenarios and improve the accuracy of field mapping, based on the above embodiments, in this application embodiment, the basic information of each field, the field association relationship corresponding to each business, and the sub-data information corresponding to each field within a set time period are stored in a heterogeneous graph; The step of using the three-flow graph neural network to update the semantic vector, structural vector, and statistical vector corresponding to each field includes: Add the semantic vector, structural vector, and statistical vector corresponding to each field to the node corresponding to each field in the heterogeneous graph; The updated heterogeneous graph is input into the three-flow graph neural network; the three-flow graph neural network updates the semantic vector of each field in the heterogeneous graph based on whether the two fields corresponding to each edge belong to the same business entity; updates the statistical vector corresponding to each field in the heterogeneous graph based on the difference vector between the statistical vector of each field and the preset global distribution primitive vector; and updates the structure vector of each field based on the hop count corresponding to the structure vector of each field.

[0095] In heterogeneous system field mapping tasks, while semantic, structural, and statistical features are effective individually, they are difficult to capture deep interactions between fields if processed in isolation. For example, two fields may have similar semantics and symmetrical structures, but are actually unrelated due to different business entities; or their statistical distributions may be highly consistent, but they cannot be mapped due to broken call chains. Based on this, this application proposes a feature collaborative update mechanism based on heterogeneous graphs and three-flow graph neural networks. This mechanism explicitly models the relationships between fields through graph structures and dynamically integrates multimodal information during propagation.

[0096] In the embodiments of this application, it is possible As the skeleton, it is split into old system subgraphs based on the field source system. With the new system subgraph Write a unified primary key to all nodes in both graph files. This forms a heterogeneous graph of the old and new systems with metadata dictionary identifiers, providing a unique and consistent graph base for trimodal coding and alignment optimization.

[0097] Specifically, the basic information of each field, the field relationships corresponding to each business, and the sub-data information corresponding to each field within a set time period can be stored in the heterogeneous graph of the corresponding system. In addition, the semantic vector, structural vector, and statistical vector corresponding to each field will be added to the node corresponding to each field in the heterogeneous graph.

[0098] For example, each node in the heterogeneous graph contains basic information about the field corresponding to that node. Path features that are related to this field Sub-data information corresponding to each field within a set time period The above information can be stored in the node attribute field of the node corresponding to this field.

[0099] The electronic device inputs the updated heterogeneous graph into a Tri-stream GNN. This Tri-stream GNN is not a single process, but rather runs three message-passing streams in parallel, each stream focusing on the intra-graph propagation and update of one type of feature.

[0100] Specifically, in cross-system mapping, relying solely on semantic vectors can lead to confusion between order-time and pay-time, while using distribution alone makes it difficult to distinguish between truncation caused by DB field migration and actual business growth. Therefore, this application's embodiments adopt a progressive fusion strategy of semantics, distribution, and path: based on weights... semantic vectors With statistical vectors Weighted concatenation to form an intermediate vector Vectors integrate naming information and range differences into the same dimension. This is then achieved through a linear mapping matrix. intermediate vector With structure vector Further fusion yields the final three-modal enhancement vector. .

[0101] The aforementioned three-modal augmentation vectors not only align the semantic and statistical features of the old and new fields in the same space, but also implicitly carry the link depth and calling direction, providing a unified and high-quality input base for three-flow graph neural networks and multi-view contrastive learning.

[0102] In semantic flow, messages are only transmitted along edges that co-occur within the same business entity. For any edge... If the side If field i and field j at both ends belong to the same business entity, then the semantic vectors of field i and field j in the heterogeneous graph are updated.

[0103] For example, in migration practice, although the old system field ORDER-NO and the new system field APP-NO have the same meaning, the difference in numbering can lead to misalignment of downstream indexes. Therefore, forcibly maintaining complete consistency of node IDs here not only simplifies cross-system edge mapping but also provides a directly alignable index foundation for subsequent cross-flow gating. The three inputs are treated as independent signal channels within the graph to ensure that each type of problem is fully modeled within its dedicated network structure and is not diluted by early fusion.

[0104] Specifically, entity-aware gated graph attention is used in the semantic flow to update semantic vectors and reverse flow. In response to two typical ambiguities in field naming between the old and new versions of the system: abbreviation differences and business stage overload, this application proposes a system migration entity gated attention mechanism to strengthen information exchange between truly homologous fields with distinguishable entities as the link, while reducing interference from irrelevant fields.

[0105] In this embodiment of the application, for each edge Calculate entity gating weights ; if and only if field i and j share the same entity type Otherwise Based on weights Recalculate the attention coefficient, based on the following expression of the idea that amplification only occurs when entities are from the same source:

[0106] in: Representation field For fields Semantic attention weights; This represents a pre-configured temperature coefficient used to control the steepness of the weight distribution; This indicates the entity gating weight, which is 1 when the fields and entities are the same. This represents a trainable query vector; Represents a node , semantic vector; This indicates vector concatenation; Represents a node The semantic neighbor set.

[0107] Through the aggregation of the above formula, the updated semantic vector can bring ORDER-ID and ORDER-NO closer together in terms of representation distance, while separating CREATE-TIME and UPDATE-TIME due to different entities, directly alleviating cross-system naming mismatches.

[0108] In the statistical flow, the three-flow graph neural network updates the statistical vector corresponding to each field in the heterogeneous graph based on the difference vector between the statistical vector of each field and the preset global distribution primitive vector. For each field v, the difference between its c and the global distribution primitive vector is calculated. This difference reflects the degree of deviation of the field relative to the global typical distribution.

[0109] Specifically, in the distribution memory residual update of statistical streams, the old system's database fields mostly used INT, with a numerical upper limit of 2147483647; the new system upgrades to DECIMAL(20,4), which not only expands the positive range but also allows decimal places. This involves simply comparing statistical vectors. This will produce significant differences. In this embodiment, a globally distributed primitive vector m is constructed to learn a common statistical template across systems, and binning residuals are calculated for each field. .

[0110] The updated statistical vector satisfies the following formula:

[0111] in: Representation field Updated statistical vector; This represents a pre-configured trainable mapping matrix used to transform the template and residual to the latent space; Represents the global distribution primitive vector, used to capture the common value range baseline of the old and new systems; This indicates the individualized distribution offset of the nodes; This represents the statistical vector of field i, which is the statistical vector of field i obtained by the three-flow graph neural network.

[0112] By combining global templates with local residuals, statistical flow can both absorb the overall drift caused by changes in binning specifications between the old and new systems and retain the extreme growth of certain fields during promotional seasons, providing a more stable statistical basis for threshold matching.

[0113] In structured flows, the updating of field structure vectors depends on hop count information. Multi-hop relationship aggregation in structured flows is crucial during the transition between old and new systems, where service architectures often undergo gradual splitting from monolithic to vertical domain to microservices, thus expanding what were originally single-hop field dependencies into multi-hop links. Directly applying single-hop aggregation will lead to missing long-chain contexts and increased mapping risks. This application's embodiment records a hop count label for each edge in the structured flow. And in the multi-head attention mechanism of Relational GraphTransformer, for different Configure a separate parameter subspace to ensure that one-hop and four-hop models are modeled differently. The updated structure vector is obtained after multi-hop aggregation. It can fully restore the cross-table context, enabling subsequent comparisons to identify scenarios where fields are split but the business meaning remains the same.

[0114] After one round of three-flow propagation, each field receives an updated semantic vector, structural vector, and statistical vector. The three-flow graph neural network then automatically assigns fusion weights based on the field type. We perform weighted fusion of the updated semantic vector, statistical vector, and structural vector for each field to obtain the feature vector corresponding to each field.

[0115] Cross-stream fusion generates a unified scene-aware vector. Semantic stream excels at handling named synonyms, statistical stream focuses on value range drift, and structural stream characterizes link depth, but their judgment criteria are not always consistent. Simple concatenation would lead to a mixture of strong and weak evidence, making training difficult. Therefore, this application's embodiments design scene gating weights. The system automatically learns weight values ​​based on system source, field hotspot level, and anomaly markers, and then updates the semantic vector. Updated statistical vectors Updated structure vector Perform a weighted summation to obtain the feature vector corresponding to each field. This approach maintains a unique perspective from all three sources while avoiding information conflicts. The resulting vectors automatically increase the semantic components for high-business-value fields and automatically increase the statistical components for numerically sensitive fields.

[0116] After all nodes have been merged, the electronic device will update the semantic vector corresponding to each field. Updated statistical vectors Updated structure vector With feature vectors The data is also written back to the node attribute domain of the node corresponding to the field in the heterogeneous graph. Based on this, subsequent multi-view comparison learning can either perform fine-grained alignment by reading from a single viewpoint, or directly call the unified vector for rapid candidate generation. At the same time, this write-back operation allows operations and maintenance personnel to simultaneously see semantic, statistical, and structural similarity when viewing the field in the visualization interface, providing multi-faceted decision-making basis for manual review.

[0117] Figure 2 The flowchart of the three-flow graph neural network provided in the embodiments of this application is as follows: Figure 2 As shown, the process includes: In the input splitting stage, the old / new system field nodes are processed, and a trimodal vector is extracted and initialized according to semantic, statistical, and structural paths. Semantic stream encoding uses an entity-aware gating attention mechanism to address abbreviation differences and business stage overload issues, outputting semantic stream results. Statistical stream encoding is implemented through a globally distributed primitive plus residual module, capturing type upgrades and seasonal long-tail features, and outputting statistical stream results. Structural stream encoding is based on a multi-hop relational attention mechanism, restoring cross-table context information after database splitting, and outputting structural stream results. The outputs of the three streams enter the cross-stream fusion stage, generating a unified vector through a scene-gating weight mechanism. Finally, the unified vector is embedded and written back, synchronously writing semantic, statistical, and structural information into the graph. This entire process includes the hop distance adjustment paths of semantic → distribution and distribution → structure, as well as information transmission and interaction between the streams.

[0118] To adapt to various scenarios and improve the accuracy of field mapping, based on the above embodiments, the method in this application embodiment further includes: Obtain the preset constraint vector corresponding to each field, which is determined based on the field's anomaly alarm label and / or business label; The preset constraint vector corresponding to each field is concatenated with the corresponding feature vector, and the feature vector is updated using the concatenated vector.

[0119] In real-world business systems, some fields suffer from issues such as inconsistent naming, abnormal value ranges, or frequent changes due to historical reasons, making reliable mapping difficult using only general features. For example, a certain "Remarks" field was once temporarily used to store approval opinions, error codes, or even encryption keys, resulting in an extremely discrete statistical distribution and a semantic vector without clear direction. To address the mapping challenges of such problematic fields, this embodiment introduces a pre-defined constraint vector mechanism, integrating external monitoring signals (such as anomaly alarm tags and business classification tags) into the feature representation to achieve targeted enhancement of high-risk fields.

[0120] First, two types of constraint information are extracted from the operation and maintenance monitoring platform and the business management system: one is anomaly alarm tags, which come from field-level data quality monitoring rules, such as "null value rate > 90%", "enumerated value violation", "value mutation alarm", etc.; the other is business tags, which come from the data governance catalog, such as "customer master data", "financial sensitive fields", "regulatory reporting fields", etc. Each type of tag is encoded as a one-hot vector or an embedded vector.

[0121] In addition, the abnormal alarm label for a field can also be determined based on the sub-data information corresponding to each field within a set time period.

[0122] Specifically, for each field p, the value range of field p registered in the manual rule base can be used. For each field within a set time period, the set of sub-data information is... Perform out-of-bounds detection to obtain the anomaly rate for that field. and the abnormal ratio As an abnormal alarm label.

[0123]

[0124] in: This indicates the minimum value within the specified range of values ​​for field p; This indicates the maximum value within the specified range of values ​​for field p; This represents the s-th sub-data information of field p; This indicates the total number of samples of the sub-data information; This represents the anomaly rate, is used as an anomaly alert label, and is written back to the field metadata.

[0125] Electronic devices can concatenate the preset constraint vector corresponding to each field with the corresponding feature vector, and use the concatenated vector to update the feature vector.

[0126] Example 2: To adapt to various scenarios and improve the accuracy of field mapping, based on the above embodiments, in this embodiment, determining the target first field that matches each second field according to the second feature vector of each second field and the first feature vector of each first field includes: For each second field, the similarity between the second field and each first field is determined based on the second feature vector of the second field and the first feature vector of each first field; the first field whose similarity exceeds the similarity threshold is determined as the target first field that matches the second field.

[0127] In this embodiment, the electronic device can use a unified field vector to calculate the similarity matrix from the old system to the new system, and generate candidate mappings by combining table name consistency, enumeration intersection and data type compatibility dynamic thresholds. It can automatically pass low-risk mappings, generate manual review work orders for high-risk mappings, and write the review results back to the model through a message queue. When capturing field addition, deletion or enumeration change events, it triggers a full-link adaptive process of candidate generation, manual closure and incremental fine-tuning.

[0128] Specifically, for each second field, the similarity between the second field and each first field is determined based on the second feature vector of the second field and the first feature vector of each first field; the first field whose similarity exceeds the similarity threshold is determined as the target first field that matches the second field.

[0129] To adapt to various scenarios and improve the accuracy of field mapping, based on the above embodiments, in this embodiment of the application, if there is no corresponding first field with a similarity exceeding a similarity threshold, the method further includes: Determine the difference between the highest similarity among the similarities between the second field and each of the first fields and the similarity threshold; If the difference is within a preset range, a work order awaiting review is generated, and a manual review notification is triggered.

[0130] While dynamic thresholding significantly improves matching quality, "no candidates" may still occur in scenarios with severe field naming drift or business logic refactoring. Directly abandoning the mapping will lead to data loss; forcing matching may introduce errors. Therefore, this embodiment designs a difference-aware pending work order generation mechanism to maximize automated coverage while ensuring security.

[0131] When the similarity between a second field and all the first fields is lower than the similarity threshold, the electronic device will not immediately report an error. Instead, it will calculate the difference between the highest similarity between the second field and each of the first fields and the similarity threshold.

[0132] Electronic devices are pre-configured with a tolerance range, which is a preset range. If the difference between the highest similarity score of the second field and each of the first fields and the similarity threshold falls within this preset range, it indicates that although the second field does not reach the threshold, it is highly similar to a certain first field and warrants manual review. In this case, the system automatically generates a pending review work order, which is automatically notified to the data governance specialist via a preset method, with a 48-hour timeout reminder. The specialist can confirm, reject, or request more information with one click. This preset method can be a client pop-up, email, SMS, or telephone.

[0133] The content of the work order includes, but is not limited to: the Chinese title of the second field and the first field with the highest similarity to the second field, typical sample values, and call chain path. The operation and maintenance personnel can confirm or reject the work order with one click.

[0134] To adapt to various scenarios and improve the accuracy of field mapping, based on the above embodiments, the method in this application embodiment further includes: Determine the number of pending work orders currently generated; If the quantity reaches a preset quantity threshold, then a fine-tuning task for the feature extraction model is initiated based on the second field corresponding to the work order to be reviewed and the first field with the highest similarity to the second field.

[0135] In engineering practice of heterogeneous system field mapping, while automated matching can cover most common scenarios, it inevitably generates several boundary cases when faced with complex situations such as business logic restructuring, naming convention changes, or the accumulation of historical technical debt, requiring manual intervention for judgment. However, if manual review is only regarded as the end point, without transforming its feedback into improvements in model capabilities, electronic devices will repeatedly make mistakes. Based on this, this application's embodiments construct a closed-loop optimization mechanism driven by pending work orders. By dynamically sensing the review load and triggering model fine-tuning, it achieves an intelligent leap from passive response to proactive evolution.

[0136] Specifically, during operation, the electronic device records the number of work orders that have been generated and are awaiting review. If the number reaches a preset threshold, the device will start a fine-tuning task for the feature extraction model based on the second field corresponding to the work order and the first field that has the highest similarity to the second field.

[0137] The fine-tuning process employs a lightweight incremental learning strategy to balance efficiency and stability. Throughout the fine-tuning process, the underlying parameters of the semantic feature extraction sub-model, structural feature extraction sub-model, and statistical feature extraction sub-model are frozen, as they have converged to a stable state during large-scale pre-training and domain-specific fine-tuning. The high-level interaction modules in the three-flow graph neural network are the ones truly involved in the update, including cross-modal attention weights, graph message passing gating coefficients, and the final feature fusion gate. These components are responsible for determining the relative importance of different features in a specific context, and they are the most susceptible to the influence of new business models or naming conventions.

[0138] The training data used in the fine-tuning process comes directly from the latest batch of approved work orders. For each confirmed field pair, the system constructs it as a positive sample, aiming to narrow their distance in the feature space; for rejected field pairs, they are treated as negative samples, aiming to widen their distance. The loss function uses a weighted contrastive loss, where the weights are determined by the risk level of the field—highly sensitive fields incur higher penalties for misjudgment, thus guiding the model to prioritize the accuracy of mapping key data.

[0139] Specifically, electronic devices use a unified field vector to calculate the similarity matrix from the old system to the new system. Combined with table name consistency, enumeration intersection and data type compatibility dynamic thresholds, candidate mappings are generated. Low-risk mappings are automatically approved, while high-risk mappings generate manual review work orders and the review results are written back to the model through a message queue. When capturing field addition, deletion or enumeration change events, a full-link adaptive process of candidate generation, manual closure and incremental fine-tuning is triggered.

[0140] The above end-to-end adaptive process includes the following steps: 1. Calculate the cosine similarity between the field sets of the old system and the field sets of the new system, generating matrix S. For each old field u, select the first... A new field is created to form the candidate table. ; 2. The candidate pair (u,v) is fed into the business rule engine, and the similarity threshold is dynamically set based on business characteristics such as table name consistency, enumeration intersection, and data type compatibility. .when The time stamp indicates automatic passage; otherwise, proceed to the next risk correction step. Among these, Let be the similarity between any first field and any second field.

[0141] 3. For those falling For high-risk mappings within a given range, the system automatically generates a work order for review on the visualization platform. The interface simultaneously displays the field's Chinese title, typical sample values, and call chain path, allowing operations personnel to confirm or reject the request with a single click.

[0142] 4. Manual operations immediately write back to the Kafka queue and record. The threshold log is updated. Confirmation / rejection flags are simultaneously sent to the comparison sample generator to ensure the model sees the latest facts in the next training iteration.

[0143] 5. Small batch incremental fine-tuning: When the feedback samples accumulate to the batch threshold B, the incremental training task is automatically started; the confirmed samples are used to positively strengthen the in-stream weights, improving the model's sensitivity to new naming conventions; the rejected samples are penalized in the reverse vector distance to quickly correct misjudgment directions.

[0144] The entire incremental process described above adopts a lightweight strategy of freezing the previous layer and only fine-tuning the higher-order gating coefficients. The typical cost is to complete one iteration within 3-5 seconds, which meets the business need to adapt immediately even if the field changes suddenly on the day of launch.

[0145] 6. Continuously monitor the event stream of the metadata center. Once a field addition, field deletion, enumeration update, or version iteration event is captured, immediately trigger the automatic re-run of the entire chain described in 1-5 above.

[0146] When the second phase of the new system is launched and the old system components are decommissioned, the system can complete the adaptation within 5 minutes, thus resolving the long-term hidden danger of old mappings becoming invalid due to metadata drift.

[0147] 7. Daily offline task summary includes three metrics: mapping accuracy, frequency of manual intervention, and backtracking rate of false match warnings. Rolling parameter tuning is performed on rule parameters such as v. High-false-match tables are automatically written into regularized naming rules or have their enumerated intersection weights increased. Through continuous evaluation and self-tuning processes, progressive collaboration among rules, models, and human intervention is achieved.

[0148] Example 3: To adapt to various scenarios and improve the accuracy of field mapping, based on the above embodiments, in this application embodiment, the training process of the three-flow graph neural network includes: Obtain sample pairs, each sample pair including an initial sample semantic vector, an initial sample structure vector, and an initial sample statistical vector corresponding to two sample fields respectively, whether the two sample fields are labels of the same node, and the boundary value of the sample pair; wherein, the boundary value is determined based on the number of sample fields that have a direct correlation with any sample field in the sample pair; Using a three-flow graph neural network to be trained, the initial sample semantic vector, initial sample structure vector, and initial sample statistical vector corresponding to the two sample fields included in the sample pair are updated respectively to obtain the sample semantic vector, sample structure vector, and sample statistical vector corresponding to the two fields respectively; Based on the sample semantic vectors corresponding to the two fields respectively, the semantic distance is determined, and based on the relationship between the semantic distance and the boundary value, and whether the two sample fields are labels of the same node, the semantic loss value is determined. The statistical loss value is determined based on the label, the sample statistical vectors corresponding to the two fields, and the boundary value. The structural loss value is determined based on the sample structure vector corresponding to the two sample fields included in the sample pair and the boundary value, respectively. Based on the semantic loss value, the statistical loss value, and the structural loss value, a total loss value is determined, and the parameters of the three-flow graph neural network are adjusted according to the total loss value.

[0149] The three-flow graph neural network, as the core engine of the field mapping method in this application embodiment, highly depends on whether the training strategy can truly reflect the topological characteristics of the business graph. Traditional graph representation learning methods often assume that all node pairs have equal importance and use a uniform loss function for optimization. This can easily lead to overfitting in high-connection-density regions and underfitting in sparse regions in real-world business graphs. Based on this, this application embodiment proposes a boundary-value-aware multi-task contrastive learning framework. By introducing graph structure priors to dynamically adjust the loss weights, the model can adapt to learning needs under different connection densities.

[0150] In this embodiment, the training data is organized in the form of sample pairs, each pair containing two field nodes and their corresponding initial semantic vector, initial structural vector, and initial statistical vector. These vectors are generated by the aforementioned sub-model without graph propagation, preserving the original feature distribution. Each sample pair also includes a binary label indicating whether the two fields belong to the same entity or perform the same function in business logic. Each sample pair also includes a boundary value, defined as the smaller of the number of neighbors of each field in the original heterogeneous graph. The boundary value essentially characterizes the credibility of the local topology of the sample pair; that is, the more neighbors, the more active the field is in the business process, and the more stable and reliable its association is. Therefore, the supervision signal provided by the sample pair should also be given higher weight.

[0151] During training, the three-flow graph neural network to be trained first performs independent message passing and updates on the three initial vectors of the two fields, outputting their respective enhanced semantic vectors, structural vectors, and statistical vectors. Subsequently, the electronic device calculates the contrastive loss for each type of feature. Taking semantic flow as an example, the model calculates the cosine distance between the sample semantic vectors of the two sample fields, the relationship between the semantic distance and boundary values, and whether the two sample fields belong to the same node, to determine the semantic loss value. Additionally, it determines the statistical loss value based on the label, the sample statistical vectors corresponding to the two fields, and the boundary values; and it determines the structural loss value based on the sample structural vectors and boundary values ​​corresponding to the two sample fields included in the sample pair.

[0152] The electronic device also determines the total loss value based on the semantic loss value, statistical loss value, and structural loss value, and adjusts the parameters of the three-flow graph neural network based on the total loss value.

[0153] In this embodiment of the application, the electronic device stores a sample dataset, which includes multiple sample pairs. Each sample pair includes an initial sample semantic vector, an initial sample structure vector, and an initial sample statistical vector corresponding to two sample fields, whether the two sample fields are labels of the same node, and the boundary value of the sample pair. The boundary value is determined based on the number of sample fields that are directly related to any sample field in the sample pair.

[0154] Specifically, the process of building sample pairs using electronic devices includes: 1. Construction of cross-flow positive examples in the old system subgraph With the new system subgraph On the union of the nodes, traverse the field index node by node. ,Will , and The three pairs of embeddings are sequentially stored in the database and uniformly tagged as Positive-Intra.

[0155] 2. Call the real-time alarm interface of the monitoring platform to identify the set of fields with outlier values ​​or abnormal call chains. Perform a slight disturbance: Gaussian-Noise injection is performed on the statistical stream embedding, introducing... Amplitude distribution drift; Simulate service circuit breaking by randomly pruning a non-critical hop edge in the embedded structured flow.

[0156] The perturbed vector is paired with the original vector and uniformly labeled with Negative-Noise. This controllable noise not only prevents the business context from being completely distorted, but also forces the model to learn robust identification capabilities for anomalous offsets.

[0157] 3. Cross-system positive and negative sample sampling: Read the release log Manual mapping table : like record fields With fields To upgrade the inheritance relationship, or If both are manually mapped to have the same business meaning, then... Labeling as Positive-Cross; If the two fields belong to different business domains and have no inherited records, they are randomly paired and labeled Negative-Cross.

[0158] Ensure that the three-flow graph structure model can still learn to aggregate even when the semantics are consistent but the system origins are different, and maintain separation even when the business is unrelated but the names may be similar.

[0159] 4. Hotspot-driven positive and negative example sampling weighting to count the frequency of each field's calls in the order process. With risk marker (If the field has triggered a risk control alarm) ,otherwise ), calculate the comprehensive hot topic score ,according to Adaptive adjustment of sampling weights .

[0160] The overall hot topic score satisfies the following formula:

[0161] in, Representation field The frequency of calls; Representation field The risk flag, with a value of 0 or 1; This represents the weighted hyperparameter, which controls the ratio of call frequency to the influence of risk factors. Represents the Sigmoid normalization function, guaranteeing ; Representation field The final weight in the sampling pool.

[0162] Electronic devices through Adjusting the sampling probabilities of positive and negative examples allows high-frequency and high-risk fields to account for a larger proportion in the training set, directly improving the migration accuracy and its focus on key business nodes.

[0163] 5. Perspective splitting and business weight writing: Based on the embedding types involved in the paired samples, they are split into semantic perspective sets. Statistical perspective set Structural perspective set At the same time, additional business weights are assigned to each sample. (If the sample source field is) Version tag (Values ​​Old, New, Cross) thus forming a tuple. .

[0164] 6. Sample deduplication consistency check and data packet encapsulation: Will , , After merging, perform the following two steps: Based on key Remove duplicates caused by multiple rounds of sampling; Check the consistency of labels for the same pair of samples under different perspectives. If there is a conflict, the higher-weighted perspective shall be used to make a unified correction.

[0165] The final output is a multi-view comparison sample set. Each record contains a business weight. Version tags This provides a high-quality, business-aware, and multimodal consistent source of training data for the dynamic boundary comparison optimization in step 5 above.

[0166] In this embodiment of the application, the electronic device can also read sample pairs to perform dynamic boundary comparison optimization in the three-flow network. The boundary value is obtained by statistical flow centrality index compression. Piecewise cosine distance loss is used in the semantic flow, residual consistency loss is used in the statistical flow, and multi-hop branch loss is used in the structural flow. Three-flow consistency constraints are applied, and the network parameters are iteratively updated until the verification index is stable.

[0167] Specifically, the process includes: 1. Differentiated boundary writing: Batch scan the sample package to obtain each field. Degree centrality in statistical flow graphs In data migration practice, high degree centrality means that the field is on the main business backbone, and mismatch will directly cause production accidents; while low degree centrality mostly refers to log shadow fields or redundant identifiers, with limited impact. Therefore, in the embodiments of this application, Mapped to the boundary via exponential compression The higher the centrality, The smaller the value, the faster the high-risk fields converge during comparative learning. Boundary values, sample labels, and business weights are written into the metadata to ensure that the semantic flow, statistical flow, and structural flow share a unified standard of slackness, eliminating the problem of inconsistent scales when operating on three fronts.

[0168] 2. Semantic Flow Segmentation Comparison Loss: In the forward stage of semantic flow, the embodiments of this application aim to solve the problem of naming differences and synonym aggregation: CREATE-TIME in the old system is often overloaded as ORDER-PLACED-TIME in the new system. If no boundary constraints are applied, the model will distance the homologous semantics due to the literal differences in description.

[0169] To this end, read the sample tuple Calculate the cosine distance A segmented loss method is proposed, which uses business weights. Strengthen the gradient of high-frequency fields, and utilize A suitable distance threshold is set to obtain the semantic loss value. When the semantic distance of a positive example exceeds a certain threshold... Time amplification error, when the distance to the negative example is less than Apply pressure at the same time to ensure that synonym fields are grouped together and to prevent near-synonym deception.

[0170] The semantic loss value satisfies the following formula:

[0171] in, This is the semantic loss value. The semantic distance between sample field u and sample field v; The label indicating whether sample field u and sample field v belong to the same node; Weighting based on business hotspots; This is a dynamic boundary. Both positive and negative examples trigger penalties when they reach their respective boundaries, ensuring that the semantic flow continuously corrects naming mismatches throughout the entire migration cycle.

[0172] 3. Statistical flow residual-driven loss: For numerical fields, older systems often suffer from truncation, scaling, and carry-over errors due to inconsistent field types or units of measurement. This application's embodiment superimposes local node residuals onto the global distribution primitive vector, proposing an enhanced template consistency function: when the sample is positive, the Euclidean distance between the two residual vectors is directly minimized; if it is negative and the distance is below the boundary, the penalty is doubled to prevent spurious similarity distributions from misleading the mapping, thereby obtaining the statistical loss value.

[0173] The structural loss value satisfies the following formula:

[0174] in, This represents the structural loss value. The label indicating whether sample field u and sample field v belong to the same node; Weighting based on business hotspots; For dynamic boundaries, Let be the sample structure vector for field u. For the sample structure vector of field v, This is a globally distributed primitive vector.

[0175] In the embodiments of this application, both positive and negative examples trigger penalties when they reach their respective limits, so that the semantic flow continuously calibrates naming mismatches throughout the entire migration cycle.

[0176] The first branch of the above equation ensures that homogeneous fields across systems can still align primitives even after type upgrades. The second branch sets a minimum interval for non-homogeneous fields to prevent large number occlusion. The gradient is propagated back to the globally distributed primitive vector and mapping matrix via the residual path. This allows statistical templates to evolve adaptively during migration, rather than being fixed in any historical version.

[0177] 4. The new system's finer-grained service leads to a sharp increase in link depth. If only a single-hop comparison is performed on cross-system fields, they will be misjudged as missing. To correct this problem, the embodiments of this application maintain the same dynamic boundary logic in the structured flow as in steps 2 and 3 above. However, multi-head attention is additionally allowed to independently absorb gradients on branches with different hop counts, prompting the network to automatically fold long chain dependencies back into the nearest neighbor space, thereby reducing the interference of path distortion on the mapping.

[0178] 5. Although the three streams each perform their respective functions, the same field must ultimately form an aggregation anchor point in the three implicit spaces of semantics, statistics, and structure. Based on this, the embodiments of this application extract fields at the end of each training batch. The three-stream vectors and the total difference are calculated. The total difference is then used to calculate the three-stream loss value. That is, embodiments of this application propose applying a consistency constraint: high-exposure fields... boundary Smaller sizes force the three streams of expression to converge as quickly as possible; low-frequency fields retain moderate freedom to avoid overfitting.

[0179] The three-stream loss values ​​satisfy the following formula:

[0180] in, This represents the loss value of the three streams. The total difference between the three streams; Weighting based on business hotspots; It is a dynamic boundary.

[0181] The mechanism explicitly suppresses modality splitting, ensuring that subsequent fusion layers have a stable and unified input foundation when facing high-risk fields.

[0182] 6. , Structure flow corresponding loss and Summation by batches The data is then uniformly fed back into the three-flow graph structure model for adjusting its parameters. Because the dynamic boundary amplifies the error of high-exposure fields, the gradients of these fields are automatically weighted during backpropagation, resulting in finer-grained updates to the gating vector, residual coefficients, and hop count attention. Gradient decay in low-frequency fields prevents noise dominance in the early stages of migration.

[0183] 7. Iterate continuously until the field mapping accuracy, three-stream consistency, and cross-system recall of the validation set are simultaneously stable, triggering early stopping. Freeze the three-stream network weights and derive the comparatively optimized semantic vectors. Statistical vectors Structure vector The data is stored in the node attributes to provide an input vector with higher domain independence and business sensitivity for the next step of cross-stream weighted fusion, while also reserving a gradient entry point for subsequent online incremental fine-tuning.

[0184] In this embodiment, after freezing the three-stream network, cross-stream learnable attention is introduced. Under business prior initialization, semantic, statistical, and structural weights are adaptively allocated to generate a fusion vector. Domain adversarial training is performed through a gradient inversion layer to reduce system domain differences. Under scene label supervision, the top-level feedforward network is fine-tuned to output a unified field vector.

[0185] The process includes: 1. Initialize business-specific weights by reading the semantic vector of each field in the heterogeneous graph. Statistical vectors Structure vector Simultaneously, retrieve the order payment from the subsystem to which the field belongs from the configuration center. Create status amount with function label For any two streams Calculate business similarity prior value range . Provides a business initialization point for cross-stream attention, enabling the order subsystem – amount field to gain higher prior confidence when semantically and statistically fused.

[0186] 2. Cross-stream attention weighted fusion is employed, with three embedding paths implemented in the semantic, statistical, and structural order to construct a two-layer cross-stream learnable attention network. For fields... Let the three-stream input matrix be... The flow weight is calculated by combining business priors.

[0187] The flow weight satisfies the following formula:

[0188] in, Representation field In the flow Embedding; This represents pre-configured trainable parameters; This represents the element-wise activation function; Indicates the business prior temperature coefficient; Represents a stream The business similarity prior corresponds to the calculation result in step 1 above.

[0189] After obtaining the weights, calculate the first-version fusion vector. Vectors have amplified the importance of order amount in semantic and statistical dimensions, while moderately compressing the impact of structure flow on this field.

[0190] 3. Retrieve order stage tags associated with fields from the metadata center for loan disbursement. and risk warning level Based on a learnable vocabulary Mapped to vector , and then with Splicing to generate joint representation Based on this representation, the order placement stage (high risk) and the loan disbursement stage (low risk) can generate separable features in the vector space, enabling downstream users to quickly filter field pairs according to scenarios.

[0191] 4. Feed into a single-layer feedforward network Predicting based on scene labels For monitoring signals, during fine-tuning, only the feedforward layer weights are updated, and the cross-flow attention parameters are frozen to ensure that scene information is explicitly represented in the final representation without diluting business priors. The F1 score of the predicted labels is monitored in the validation set to ensure that scene separability meets the deployment threshold.

[0192] 5. Domain Invariance Adversarial Layer and Final Vector Output: Although S5 has suppressed domain shift at the embedding level, differences in interface protocols still exist between the old and new systems. This sub-step inserts a domain adversarial layer to minimize the discriminative power of the system's classifier while maintaining the scene label prediction accuracy through gradient inversion.

[0193] Assume the system classifier output Scene loss Losses in the fight against the domain Overall objectives:

[0194] in, This represents the cross-current attention and feedforward layer parameters; This represents the parameters of the scene label prediction header; Indicates the system classifier parameters; Represents the cross-entropy scenario loss; This represents the cross-entropy domain discriminant loss; This indicates adversarial weights.

[0195] The optimization employs a gradient inversion layer to solve the min-max problem: Update Improve domain discriminative power and perform reverse gradient updates. Suppressing the domain signal improves domain invariance without sacrificing scene prediction accuracy. After training convergence, all parameters are frozen, and the final unified field vector is derived. It also features deep alignment capabilities, explicit scene identification, and system domain inseparability, making it directly applicable to downstream field matching, data lineage tracing, or knowledge graph completion tasks.

[0196] Figure 3 The overall flowchart provided for the embodiments of this application is as follows. Figure 3 As shown, the process includes: S1: Perform field extraction and knowledge graph construction, including English names, Chinese tags, business comments, service registry, online version numbers and distribution histogram samples, and generate a table-field-field heterogeneous knowledge graph.

[0197] Specifically, the electronic device loads the field schema of the old system and the new system, parses the table-field hierarchy in lifecycle order, extracts field value samples based on the real-time log window, calculates discrete frequency vectors and performs out-of-bounds detection, and writes information such as field English name, Chinese label, business comments, microservice identifier and online version number, along with table-field edges and field-field edges, into a heterogeneous field graph with a unified primary key.

[0198] S2: Perform trimodal feature encoding and fusion on the text, statistical and structural information output by S1. The text stream is processed by lightweight Transformer + ontology entity, the statistical stream is processed by multi-scale 1D convolution + adaptive binning, and the structural stream is processed by path-level Bi-LSTM + Hop-aware encoding, realizing the layer-by-layer fusion of text → distribution → path.

[0199] Specifically, the electronic device generates semantic vectors for each field node in the graph using a Transformer fine-tuned from the ERP-CRM industry corpus. It then generates distribution vectors using adaptive binning and one-dimensional convolutions with kernel sizes of 1, 3, and 5. Based on the shortest jump distance sequence, it obtains path vectors through Bi-LSTM and weights and fuses them in the order of semantics → distribution → path to obtain a trimodal augmented vector, which is then written into the node attributes.

[0200] S3: Innovative encoding is implemented in a three-stream GNN. The semantic stream uses Entity-Gated GAT, the statistical stream introduces a distributed memory residual module, and the structural stream adopts a Relational Graph Transformer. Scene-aware information is also embedded through a cross-stream fusion layer.

[0201] Specifically, electronic devices use a three-way graph neural network to encode node features in parallel: semantic flow, statistical flow, and structural flow. The semantic flow uses an entity-gated graph attention mechanism, the statistical flow uses a memory structure with globally distributed primitives and local residuals, and the structural flow uses multi-head relational attention with hop count labels. After in-stream comparison training, the three outputs are fused by scene-gated weights, and the semantic embedding, statistical embedding, structural embedding, and fused embedding are written back to the node.

[0202] S4: Construct multi-perspective comparison samples, including positive examples from the same source but different flows, negative examples with abnormal disturbances, manually labeled positive and negative examples, and dynamic sampling of hot fields, forming three sets of comparison training sets.

[0203] Specifically, electronic devices construct comparative learning samples based on write-back multi-view embedding, including positive examples from the same node, negative examples affected by Gaussian noise and link cut-off, and positive and negative examples from across systems. Hotspot weights are calculated by combining field call frequency and risk labeling, and view splitting and consistency verification are completed.

[0204] S5: Perform multi-view contrast optimization, utilizing exposure number adaptive boundary, contrast loss to drive GNN adjustment, and cross-view consistency constraint to output the optimized multi-stream embedding.

[0205] Specifically, the electronic device reads sample pairs and performs dynamic boundary comparison optimization in the three-flow network. The boundary values ​​are obtained by compression of the statistical flow centrality index. Piecewise cosine distance loss is used in the semantic flow, residual consistency loss is used in the statistical flow, and multi-hop branch loss is used in the structural flow. Three-flow consistency constraints are applied, and the network parameters are iteratively updated until the verification index is stable.

[0206] S6: Perform cross-stream weighted fusion and label injection on the optimized embedding. Through cross-view attention dynamic weighting, scene label injection fine-tuning, and adversarial training domain invariant fine-tuning, output a unified vector with business labels.

[0207] Specifically, after freezing the three-stream network in electronic devices, cross-stream learnable attention is introduced. Under business prior initialization, semantic, statistical, and structural weights are adaptively allocated to generate fusion vectors. Domain adversarial training is performed through gradient inversion layers to reduce system domain differences. Under scene label supervision, the top-level feedforward network is fine-tuned to output a unified field vector.

[0208] S7: Completes field alignment and online adaptation, combines similarity calculation with rule engine initial screening and visual online correction feedback, supports online fine-tuning and adaptation of incremental data, and forms a closed-loop iterative mechanism through sample generation feedback and incremental log synchronization.

[0209] Specifically, electronic devices use a unified field vector to calculate the similarity matrix from the old system to the new system. Combined with table name consistency, enumeration intersection and data type compatibility dynamic thresholds, candidate mappings are generated. Low-risk mappings are automatically approved, while high-risk mappings generate manual review work orders and the review results are written back to the model through a message queue. When capturing field addition, deletion or enumeration change events, a full-link adaptive process of candidate generation, manual closure and incremental fine-tuning is triggered.

[0210] Figure 4 This is a schematic diagram comparing the embodiments of this application with the prior art, as shown below. Figure 4As shown, the left side represents the existing technical process, which starts with a rule base, a thesaurus, and manual maintenance. It proceeds through unimodal text vector cosine distance matching, then random walk structural similarity calculation, and finally completes offline verification and manual review. The right side represents the process of this application's embodiment, using trimodal original features (semantics, distribution, and path) as input. It employs a multimodal approach to replace the unimodal approach, performing three-stream GNN parallel encoding, covering entity gating, primitive residuals, and multi-hop relation processing. Subsequently, dynamic boundary contrastive learning is introduced to replace the original fixed threshold method, and a cross-stream attention mechanism is combined to achieve a unified vector output. The entire process replaces offline review with real-time closed-loop processing, supporting online adaptation and real-time incremental fine-tuning. Multiple connection points exist between the two, indicating that the new scheme improves and optimizes the traditional method in several stages.

[0211] Compared with the prior art, the embodiments of this application have the following advantages: 1. The technical solution proposed in this application first uses the "consistent field dictionary between the old system and the new system" as the anchor point to parse the two schemas into a table-field hierarchy according to their lifecycle order, and constructs a heterogeneous graph containing table-field edges and field-field edges with fields as nodes. All nodes are uniformly written with three types of original attributes: semantic vectors generated by the domain fine-tuning Transformer, distribution vectors obtained through adaptive binning and multi-scale convolution, and path vectors encoding the shortest jump distance sequence. By embedding unique primary keys at the node level, the old and new subgraphs share the address space during the initialization phase, so that subsequent propagation and comparison no longer depend on external intermediate files or multiple sets of indexes, fundamentally avoiding the problems of feature dispersion and index misalignment in existing solutions, and reserving a unified entry point for real-time incremental updates.

[0212] 2. In the graph learning phase, this application proposes and improves a three-stream parallel graph neural network architecture: the semantic stream introduces entity-gated attention, amplifying information propagation only when fields share similar key entities (IOU number, customer ID, etc.); the statistical stream uses globally distributed primitives + local residual models to capture overall drift during field type upgrades while preserving the long tail; the structural stream uses multi-hop relationships as self-attention branches, modeling one-hop direct connections and multi-hop split links separately within the same network. The comparative training within the three streams is controlled by dynamic boundaries, with boundary size compressed in real-time by degree centrality and hotspot weights, prompting rapid convergence of core business fields while maintaining moderate freedom for shadow fields. After multiple rounds of optimization, the three latent spaces form low-difference anchor points at common-origin fields, laying a unified geometric foundation for accurate cross-system alignment.

[0213] 3. To further merge the three latent spaces, this solution constructs a cross-stream learnable attention network. It initializes semantic, statistical, and structural weights with business verification, then dynamically calculates the stream-level attention distribution for each field and outputs a fusion vector. Subsequently, a domain adversarial layer suppresses "old system / new system" domain signals, ensuring the fusion representation retains only business-related differences. Scene labels are attached to the top-level feedforward network in a frozen manner for lightweight fine-tuning, guaranteeing that the vectors can be directly used for scene retrieval without interfering with the underlying weights. In the online phase, the similarity between old and new field vectors is calculated in real time and candidates are determined based on business rules. Manual review and feedback are written back to the model via a Kafka pipeline; adaptation is completed within seconds by simply unfreezing the higher-order gating coefficients. Through these key designs, semantic, numerical, and structural changes are simultaneously perceived within the same vector space, supporting automatic and high-precision field mapping during data migration between old and new systems.

[0214] Example 4: Based on the same inventive concept, embodiments of this application provide a field mapping device for heterogeneous systems. Figure 5 A schematic diagram of a field mapping device for a heterogeneous system provided in this application embodiment is shown below. Figure 5 The device includes: The feature extraction module 501 is used to extract feature vectors for each field in the system by employing a feature extraction model, based on the basic information of each field in the system, the field association relationship corresponding to each business, and the sub-data information corresponding to each field within a set time period, to obtain the feature vector corresponding to each field; wherein, the feature vector includes at least a semantic vector, a statistical vector, and a structural vector; the system includes a new system and an old system; the feature vector corresponding to each field includes a first feature vector corresponding to each first field in the old system, and a second feature vector corresponding to each second field in the new system; The mapping module 502 is used to determine the target first field that matches each second field based on the second feature vector of each second field and the first feature vector of each first field; and to map the first data information of the target first field in the old system to the second data information of the second field corresponding to the target first field in the new system.

[0215] In one possible implementation, the feature extraction model includes a semantic feature extraction sub-model, a structural feature extraction sub-model, a statistical feature extraction sub-model, and a three-flow graph neural network; The feature extraction module 501 is specifically used to: determine the shortest path between each field and a set field based on the field association relationship corresponding to each business; determine the semantic vector corresponding to each field based on the basic information of each field in the system using the semantic extraction sub-model; determine the structure vector corresponding to each field based on the shortest path between each field and the set field using the structural feature extraction sub-model; determine the statistical vector corresponding to each field based on the sub-data information corresponding to each field within a set time period using the statistical feature extraction sub-model; update the semantic vector, structure vector, and statistical vector corresponding to each field using the three-flow graph neural network; and perform weighted fusion of the updated semantic vector, statistical vector, and structure vector of each field according to the weight corresponding to each field to obtain the feature vector corresponding to each field.

[0216] In one possible implementation, the feature extraction module 501 is specifically used for each field, whereby the semantic extraction sub-model concatenates the basic information corresponding to the field into a text sequence in a preset order, uses a pre-configured business domain vocabulary to segment the entities in the text sequence, and obtains a segmented intermediate sequence; the semantic extraction sub-model performs semantic recognition on the intermediate sequence to obtain the semantic vector corresponding to the field.

[0217] In one possible implementation, the feature extraction module 501 is specifically used for each field, whereby the structural feature extraction sub-model constructs a shortest path vector based on the shortest path corresponding to that field, where each dimension of the shortest path vector corresponds to a node in the shortest path; the structural feature extraction sub-model determines a position vector based on the position of each node in the shortest path; and the structural feature extraction sub-model performs structural feature extraction by concatenating the position vector with the shortest path vector to obtain a concatenated vector, thereby obtaining the structural vector corresponding to that field.

[0218] In one possible implementation, the feature extraction module 501 is specifically used to determine the initial statistical vector corresponding to each field by considering the number of sub-data information corresponding to each field in each preset interval within the set time period of the input layer of the statistical feature extraction sub-model for each field; the multiple convolutional layers of the statistical feature extraction sub-model convolve the initial statistical vector respectively; the pooling layer of the statistical feature extraction sub-model performs global max pooling on the output of each convolutional layer to obtain each intermediate statistical vector; the output layer of the statistical feature extraction sub-model concatenates each intermediate statistical vector into a joint vector and determines the joint vector as the statistical vector corresponding to the field.

[0219] In one possible implementation, the basic information of each field, the field association relationship corresponding to each business, and the sub-data information corresponding to each field within a set time period are stored in the heterogeneous graph; The feature extraction module 501 is specifically used to add the semantic vector, structural vector, and statistical vector corresponding to each field to the node corresponding to each field in the heterogeneous graph; input the updated heterogeneous graph into the three-flow graph neural network; the three-flow graph neural network updates the semantic vector of each field in the heterogeneous graph according to whether the two fields corresponding to each edge belong to the same business entity; updates the statistical vector corresponding to each field in the heterogeneous graph according to the difference vector between the statistical vector of each field and the preset global distribution primitive vector; and updates the structural vector of each field according to the hop count corresponding to the structural vector of each field.

[0220] In one possible implementation, the mapping module 502 is further configured to obtain a preset constraint vector corresponding to each field, the preset constraint vector being determined based on the field's abnormal alarm label and / or business label; concatenate the preset constraint vector corresponding to each field with the corresponding feature vector, and update the feature vector using the concatenated vector.

[0221] In one possible implementation, the mapping module 502 is specifically used to determine the similarity between each second field and each first field based on the second feature vector of the second field and the first feature vector of each first field; and to determine the first field whose similarity exceeds the similarity threshold as the target first field that matches the second field.

[0222] In one possible implementation, the mapping module 502 is further configured to, if there is no corresponding first field whose similarity exceeds the similarity threshold, determine the difference between the highest similarity among the similarities of the second field and each of the first fields and the similarity threshold; if the difference is within a preset range, generate a work order to be reviewed and trigger a manual review notification.

[0223] In one possible implementation, the mapping module 502 is further configured to determine the number of currently generated work orders awaiting review; if the number reaches a preset threshold, then a fine-tuning task for the feature extraction model is initiated based on the second field corresponding to the work order awaiting review and the first field with the highest similarity to the second field.

[0224] In one possible implementation, the device further includes: Training module 503 is used to acquire sample pairs, each sample pair including an initial sample semantic vector, an initial sample structure vector, and an initial sample statistical vector corresponding to two sample fields, whether the two sample fields are labels of the same node, and the boundary value of the sample pair; wherein, the boundary value is determined based on the number of sample fields that are directly related to any sample field in the sample pair; using the three-flow graph neural network to be trained, the initial sample semantic vector, initial sample structure vector, and initial sample statistical vector corresponding to the two sample fields included in the sample pair are updated respectively to obtain the sample semantic vectors corresponding to the two fields respectively. The three-flow graph neural network is constructed using the following methods: 1) Sample structure vector and sample statistical vector; 2) Semantic distance is determined based on the sample semantic vectors corresponding to the two fields; 3) Semantic loss is determined based on the relationship between the semantic distance and the boundary value, and whether the two sample fields belong to the same node; 4) Statistical loss is determined based on the label, the sample statistical vectors corresponding to the two fields, and the boundary value; 5) Structural loss is determined based on the sample structure vectors corresponding to the two sample fields included in the sample pair and the boundary value; 6) Total loss is determined based on the semantic loss, the statistical loss, and the structural loss, and the parameters of the three-flow graph neural network are adjusted based on the total loss.

[0225] Example 5: Based on the same inventive concept, embodiments of this application provide an electronic device that can implement the steps of the field mapping method for heterogeneous systems discussed above. Figure 6 This application provides a schematic diagram of an electronic device structure, such as... Figure 6 As shown, it includes: processor 601, communication interface 602, memory 603 and communication bus 604, wherein processor 601, communication interface 602 and memory 603 communicate with each other through communication bus 604. The memory 603 stores a computer program, which, when executed by the processor 601, causes the processor 601 to perform the following steps: A feature extraction model is used to extract feature vectors for each field in the system based on the basic information of each field, the field relationships corresponding to each business, and the sub-data information corresponding to each field within a set time period. The feature vectors for each field are then obtained. These feature vectors include at least semantic vectors, statistical vectors, and structural vectors. The system includes a new system and an old system. The feature vectors for each field include a first feature vector corresponding to each first field in the old system and a second feature vector corresponding to each second field in the new system. Based on the second feature vector of each second field and the first feature vector of each first field, determine the target first field that matches the second field; map the first data information of the target first field in the old system to the second data information of the second field corresponding to the target first field in the new system.

[0226] In one possible implementation, the feature extraction model includes a semantic feature extraction sub-model, a structural feature extraction sub-model, a statistical feature extraction sub-model, and a three-flow graph neural network; Before extracting the feature vector for each field in the system by using a feature extraction model, based on the basic information of each field in the system, the field relationships corresponding to each business, and the sub-data information corresponding to each field within a set time period, the method further includes: Based on the field association relationships corresponding to each business, determine the shortest path between each field and the set field; The feature extraction model is adopted to extract feature vectors for each field in the system based on the basic information of each field, the field association relationship corresponding to each business, and the sub-data information corresponding to each field within a set time period. The feature vectors corresponding to each field include: Using the semantic extraction sub-model, the semantic vector corresponding to each field is determined based on the basic information of each field in the system; Using the aforementioned structural feature extraction sub-model, the structural vector corresponding to each field is determined based on the shortest path between each field and a set field; Using the aforementioned statistical feature extraction sub-model, the statistical vector corresponding to each field is determined based on the sub-data information corresponding to each field within a set time period; The three-flow graph neural network is used to update the semantic vector, structural vector, and statistical vector corresponding to each field; Based on the weights corresponding to each field, the updated semantic vector, statistical vector, and structural vector of each field are weighted and fused to obtain the feature vector corresponding to each field.

[0227] In one possible implementation, the step of using the semantic extraction sub-model to determine the semantic vector corresponding to each field based on the basic information of each field in the system includes: For each field, the semantic extraction sub-model concatenates the basic information corresponding to the field into a text sequence in a preset order, and uses a pre-configured business domain vocabulary to segment the entities in the text sequence to obtain a segmented intermediate sequence; the semantic extraction sub-model performs semantic recognition on the intermediate sequence to obtain the semantic vector corresponding to the field.

[0228] In one possible implementation, the step of using the structural feature extraction sub-model to determine the structural vector corresponding to each field based on the field association relationship corresponding to each business includes: For each field, the structural feature extraction sub-model constructs a shortest path vector based on the shortest path corresponding to that field, where each dimension of the shortest path vector corresponds to a node in the shortest path; the structural feature extraction sub-model determines a position vector based on the position of each node in the shortest path; the structural feature extraction sub-model extracts structural features by concatenating the position vector with the shortest path vector to obtain a concatenated vector, thus obtaining the structural vector corresponding to that field.

[0229] In one possible implementation, the step of using the statistical feature extraction sub-model to determine the statistical vector corresponding to each field based on the sub-data information corresponding to each field within a set time period includes: For each field, the input layer of the statistical feature extraction sub-model determines the initial statistical vector corresponding to the field by counting the number of sub-data information corresponding to each field in each preset interval within the set time period. Multiple convolutional layers of the statistical feature extraction sub-model convolve the initial statistical vector. The pooling layer of the statistical feature extraction sub-model performs global max pooling on the output of each convolutional layer to obtain each intermediate statistical vector. The output layer of the statistical feature extraction sub-model concatenates each intermediate statistical vector into a joint vector and determines the joint vector as the statistical vector corresponding to the field.

[0230] In one possible implementation, the basic information of each field, the field association relationship corresponding to each business, and the sub-data information corresponding to each field within a set time period are stored in the heterogeneous graph; The step of using the three-flow graph neural network to update the semantic vector, structural vector, and statistical vector corresponding to each field includes: Add the semantic vector, structural vector, and statistical vector corresponding to each field to the node corresponding to each field in the heterogeneous graph; The updated heterogeneous graph is input into the three-flow graph neural network; the three-flow graph neural network updates the semantic vector of each field in the heterogeneous graph based on whether the two fields corresponding to each edge belong to the same business entity; updates the statistical vector corresponding to each field in the heterogeneous graph based on the difference vector between the statistical vector of each field and the preset global distribution primitive vector; and updates the structure vector of each field based on the hop count corresponding to the structure vector of each field.

[0231] In one possible implementation, the method further includes: Obtain the preset constraint vector corresponding to each field, which is determined based on the field's anomaly alarm label and / or business label; The preset constraint vector corresponding to each field is concatenated with the corresponding feature vector, and the feature vector is updated using the concatenated vector.

[0232] In one possible implementation, determining the target first field that matches each second field based on the second feature vector of each second field and the first feature vector of each first field includes: For each second field, the similarity between the second field and each first field is determined based on the second feature vector of the second field and the first feature vector of each first field; the first field whose similarity exceeds the similarity threshold is determined as the target first field that matches the second field.

[0233] In one possible implementation, if there is no corresponding first field with a similarity exceeding a similarity threshold, the method further includes: Determine the difference between the highest similarity among the similarities between the second field and each of the first fields and the similarity threshold; If the difference is within a preset range, a work order awaiting review is generated, and a manual review notification is triggered.

[0234] In one possible implementation, the method further includes: Determine the number of pending work orders currently generated; If the quantity reaches a preset quantity threshold, then a fine-tuning task for the feature extraction model is initiated based on the second field corresponding to the work order to be reviewed and the first field with the highest similarity to the second field.

[0235] In one possible implementation, the training process of the three-flow graph neural network includes: Obtain sample pairs, each sample pair including an initial sample semantic vector, an initial sample structure vector, and an initial sample statistical vector corresponding to two sample fields respectively, whether the two sample fields are labels of the same node, and the boundary value of the sample pair; wherein, the boundary value is determined based on the number of sample fields that have a direct correlation with any sample field in the sample pair; Using a three-flow graph neural network to be trained, the initial sample semantic vector, initial sample structure vector, and initial sample statistical vector corresponding to the two sample fields included in the sample pair are updated respectively to obtain the sample semantic vector, sample structure vector, and sample statistical vector corresponding to the two fields respectively; Based on the sample semantic vectors corresponding to the two fields respectively, the semantic distance is determined, and based on the relationship between the semantic distance and the boundary value, and whether the two sample fields are labels of the same node, the semantic loss value is determined. The statistical loss value is determined based on the label, the sample statistical vectors corresponding to the two fields, and the boundary value. The structural loss value is determined based on the sample structure vector corresponding to the two sample fields included in the sample pair and the boundary value, respectively. Based on the semantic loss value, the statistical loss value, and the structural loss value, a total loss value is determined, and the parameters of the three-flow graph neural network are adjusted according to the total loss value.

[0236] Since the principle of the above-mentioned electronic device in solving the problem is similar to the field mapping method of heterogeneous systems, the implementation of the above-mentioned electronic device can be found in the embodiments of the method, and repeated details will not be repeated.

[0237] The communication bus mentioned in the aforementioned electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used in the figure, but this does not indicate that there is only one bus or one type of bus. Communication interface 602 is used for communication between the aforementioned electronic device and other devices. The memory can include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0238] The processors mentioned above can be general-purpose processors, including central processing units, network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0239] Example 6: Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing a computer program executable by a processor. When the program runs on the processor, it causes the processor to perform the following steps: A feature extraction model is used to extract feature vectors for each field in the system based on the basic information of each field, the field relationships corresponding to each business, and the sub-data information corresponding to each field within a set time period. The feature vectors for each field are then obtained. These feature vectors include at least semantic vectors, statistical vectors, and structural vectors. The system includes a new system and an old system. The feature vectors for each field include a first feature vector corresponding to each first field in the old system and a second feature vector corresponding to each second field in the new system. Based on the second feature vector of each second field and the first feature vector of each first field, determine the target first field that matches the second field; map the first data information of the target first field in the old system to the second data information of the second field corresponding to the target first field in the new system.

[0240] In one possible implementation, the feature extraction model includes a semantic feature extraction sub-model, a structural feature extraction sub-model, a statistical feature extraction sub-model, and a three-flow graph neural network; Before extracting the feature vector for each field in the system by using a feature extraction model, based on the basic information of each field in the system, the field relationships corresponding to each business, and the sub-data information corresponding to each field within a set time period, the method further includes: Based on the field association relationships corresponding to each business, determine the shortest path between each field and the set field; The feature extraction model is adopted to extract feature vectors for each field in the system based on the basic information of each field, the field association relationship corresponding to each business, and the sub-data information corresponding to each field within a set time period. The feature vectors corresponding to each field include: Using the semantic extraction sub-model, the semantic vector corresponding to each field is determined based on the basic information of each field in the system; Using the aforementioned structural feature extraction sub-model, the structural vector corresponding to each field is determined based on the shortest path between each field and a set field; Using the aforementioned statistical feature extraction sub-model, the statistical vector corresponding to each field is determined based on the sub-data information corresponding to each field within a set time period; The three-flow graph neural network is used to update the semantic vector, structural vector, and statistical vector corresponding to each field; Based on the weights corresponding to each field, the updated semantic vector, statistical vector, and structural vector of each field are weighted and fused to obtain the feature vector corresponding to each field.

[0241] In one possible implementation, the step of using the semantic extraction sub-model to determine the semantic vector corresponding to each field based on the basic information of each field in the system includes: For each field, the semantic extraction sub-model concatenates the basic information corresponding to the field into a text sequence in a preset order, and uses a pre-configured business domain vocabulary to segment the entities in the text sequence to obtain a segmented intermediate sequence; the semantic extraction sub-model performs semantic recognition on the intermediate sequence to obtain the semantic vector corresponding to the field.

[0242] In one possible implementation, the step of using the structural feature extraction sub-model to determine the structural vector corresponding to each field based on the field association relationship corresponding to each business includes: For each field, the structural feature extraction sub-model constructs a shortest path vector based on the shortest path corresponding to that field, where each dimension of the shortest path vector corresponds to a node in the shortest path; the structural feature extraction sub-model determines a position vector based on the position of each node in the shortest path; the structural feature extraction sub-model extracts structural features by concatenating the position vector with the shortest path vector to obtain a concatenated vector, thus obtaining the structural vector corresponding to that field.

[0243] In one possible implementation, the step of using the statistical feature extraction sub-model to determine the statistical vector corresponding to each field based on the sub-data information corresponding to each field within a set time period includes: For each field, the input layer of the statistical feature extraction sub-model determines the initial statistical vector corresponding to the field by counting the number of sub-data information corresponding to each field in each preset interval within the set time period. Multiple convolutional layers of the statistical feature extraction sub-model convolve the initial statistical vector. The pooling layer of the statistical feature extraction sub-model performs global max pooling on the output of each convolutional layer to obtain each intermediate statistical vector. The output layer of the statistical feature extraction sub-model concatenates each intermediate statistical vector into a joint vector and determines the joint vector as the statistical vector corresponding to the field.

[0244] In one possible implementation, the basic information of each field, the field association relationship corresponding to each business, and the sub-data information corresponding to each field within a set time period are stored in the heterogeneous graph; The step of using the three-flow graph neural network to update the semantic vector, structural vector, and statistical vector corresponding to each field includes: Add the semantic vector, structural vector, and statistical vector corresponding to each field to the node corresponding to each field in the heterogeneous graph; The updated heterogeneous graph is input into the three-flow graph neural network; the three-flow graph neural network updates the semantic vector of each field in the heterogeneous graph based on whether the two fields corresponding to each edge belong to the same business entity; updates the statistical vector corresponding to each field in the heterogeneous graph based on the difference vector between the statistical vector of each field and the preset global distribution primitive vector; and updates the structure vector of each field based on the hop count corresponding to the structure vector of each field.

[0245] In one possible implementation, the method further includes: Obtain the preset constraint vector corresponding to each field, which is determined based on the field's anomaly alarm label and / or business label; The preset constraint vector corresponding to each field is concatenated with the corresponding feature vector, and the feature vector is updated using the concatenated vector.

[0246] In one possible implementation, determining the target first field that matches each second field based on the second feature vector of each second field and the first feature vector of each first field includes: For each second field, the similarity between the second field and each first field is determined based on the second feature vector of the second field and the first feature vector of each first field; the first field whose similarity exceeds the similarity threshold is determined as the target first field that matches the second field.

[0247] In one possible implementation, if there is no corresponding first field with a similarity exceeding a similarity threshold, the method further includes: Determine the difference between the highest similarity among the similarities between the second field and each of the first fields and the similarity threshold; If the difference is within a preset range, a work order awaiting review is generated, and a manual review notification is triggered.

[0248] In one possible implementation, the method further includes: Determine the number of pending work orders currently generated; If the quantity reaches a preset quantity threshold, then a fine-tuning task for the feature extraction model is initiated based on the second field corresponding to the work order to be reviewed and the first field with the highest similarity to the second field.

[0249] In one possible implementation, the training process of the three-flow graph neural network includes: Obtain sample pairs, each sample pair including an initial sample semantic vector, an initial sample structure vector, and an initial sample statistical vector corresponding to two sample fields respectively, whether the two sample fields are labels of the same node, and the boundary value of the sample pair; wherein, the boundary value is determined based on the number of sample fields that have a direct correlation with any sample field in the sample pair; Using a three-flow graph neural network to be trained, the initial sample semantic vector, initial sample structure vector, and initial sample statistical vector corresponding to the two sample fields included in the sample pair are updated respectively to obtain the sample semantic vector, sample structure vector, and sample statistical vector corresponding to the two fields respectively; Based on the sample semantic vectors corresponding to the two fields respectively, the semantic distance is determined, and based on the relationship between the semantic distance and the boundary value, and whether the two sample fields are labels of the same node, the semantic loss value is determined. The statistical loss value is determined based on the label, the sample statistical vectors corresponding to the two fields, and the boundary value. The structural loss value is determined based on the sample structure vector corresponding to the two sample fields included in the sample pair and the boundary value, respectively. Based on the semantic loss value, the statistical loss value, and the structural loss value, a total loss value is determined, and the parameters of the three-flow graph neural network are adjusted according to the total loss value.

[0250] Since the principle of the computer-readable storage medium in solving the problem is similar to the field mapping method of heterogeneous systems, the implementation of the computer-readable storage medium can be found in the implementation of the method, and the repeated parts will not be described again.

[0251] Example 7: Based on the same inventive concept, this application also provides a computer program product, which includes computer program code. When the computer program code is run on a computer, it causes the computer to execute any of the field mapping methods for heterogeneous systems discussed above. Since the principle by which the above-described computer program product solves the problem is similar to that of the field mapping method for heterogeneous systems, the implementation of the above-described computer program product can refer to the implementation of the method, and repeated details will not be elaborated further.

[0252] In this embodiment, feature vectors of fields in the old and new systems are determined through a feature extraction model, thereby intelligently constructing a mapping relationship between fields in the old and new systems, greatly simplifying the data migration process. It eliminates the need for the person in charge of migrating to the new system to familiarize themselves with the complex table structure of the old system, cleverly shielding potential differences between the two systems. This makes the entire migration process not only efficient but also highly accurate, eliminating the reliance on manual familiarization with the old system architecture in traditional migration methods, reducing migration risks caused by differences in personnel understanding, and providing solid support for a smooth transition of data between different systems.

[0253] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0254] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0255] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0256] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0257] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A field mapping method for heterogeneous systems, characterized in that, The method includes: A feature extraction model is used to extract feature vectors for each field in the system based on the basic information of each field, the field relationships corresponding to each business, and the sub-data information corresponding to each field within a set time period. The feature vectors for each field are then obtained. These feature vectors include at least semantic vectors, statistical vectors, and structural vectors. The system includes a new system and an old system. The feature vectors for each field include a first feature vector corresponding to each first field in the old system and a second feature vector corresponding to each second field in the new system. Based on the second feature vector of each second field and the first feature vector of each first field, determine the target first field that matches the second field; map the first data information of the target first field in the old system to the second data information of the second field corresponding to the target first field in the new system.

2. The method according to claim 1, characterized in that, The feature extraction model includes a semantic feature extraction sub-model, a structural feature extraction sub-model, a statistical feature extraction sub-model, and a three-flow graph neural network; Before extracting the feature vector for each field in the system by using a feature extraction model, based on the basic information of each field in the system, the field relationships corresponding to each business, and the sub-data information corresponding to each field within a set time period, the method further includes: Based on the field association relationships corresponding to each business, determine the shortest path between each field and the set field; The feature extraction model is adopted to extract feature vectors for each field in the system based on the basic information of each field, the field association relationship corresponding to each business, and the sub-data information corresponding to each field within a set time period. The feature vectors corresponding to each field include: Using the semantic extraction sub-model, the semantic vector corresponding to each field is determined based on the basic information of each field in the system; Using the aforementioned structural feature extraction sub-model, the structural vector corresponding to each field is determined based on the shortest path between each field and a set field; Using the aforementioned statistical feature extraction sub-model, the statistical vector corresponding to each field is determined based on the sub-data information corresponding to each field within a set time period; The three-flow graph neural network is used to update the semantic vector, structural vector, and statistical vector corresponding to each field; Based on the weights corresponding to each field, the updated semantic vector, statistical vector, and structural vector of each field are weighted and fused to obtain the feature vector corresponding to each field.

3. The method according to claim 2, characterized in that, The semantic extraction sub-model, based on the basic information of each field in the system, determines the semantic vector corresponding to each field, including: For each field, the semantic extraction sub-model concatenates the basic information corresponding to the field into a text sequence in a preset order, and uses a pre-configured business domain vocabulary to segment the entities in the text sequence to obtain a segmented intermediate sequence; the semantic extraction sub-model performs semantic recognition on the intermediate sequence to obtain the semantic vector corresponding to the field.

4. The method according to claim 2, characterized in that, The step of using the structural feature extraction sub-model to determine the structural vector corresponding to each field based on the field association relationship corresponding to each business includes: For each field, the structural feature extraction sub-model constructs a shortest path vector based on the shortest path corresponding to that field, where each dimension of the shortest path vector corresponds to a node in the shortest path; the structural feature extraction sub-model determines a position vector based on the position of each node in the shortest path; the structural feature extraction sub-model extracts structural features by concatenating the position vector with the shortest path vector to obtain a concatenated vector, thus obtaining the structural vector corresponding to that field.

5. The method according to claim 2, characterized in that, The step of using the statistical feature extraction sub-model to determine the statistical vector corresponding to each field based on the sub-data information corresponding to each field within a set time period includes: For each field, the input layer of the statistical feature extraction sub-model determines the initial statistical vector corresponding to the field by counting the number of sub-data information corresponding to each field in each preset interval within the set time period. Multiple convolutional layers of the statistical feature extraction sub-model convolve the initial statistical vector. The pooling layer of the statistical feature extraction sub-model performs global max pooling on the output of each convolutional layer to obtain each intermediate statistical vector. The output layer of the statistical feature extraction sub-model concatenates each intermediate statistical vector into a joint vector and determines the joint vector as the statistical vector corresponding to the field.

6. The method according to any one of claims 2-5, characterized in that, The basic information of each field, the field association relationship corresponding to each business, and the sub-data information corresponding to each field within a set time period are stored in the heterogeneous graph; The step of using the three-flow graph neural network to update the semantic vector, structural vector, and statistical vector corresponding to each field includes: Add the semantic vector, structural vector, and statistical vector corresponding to each field to the node corresponding to each field in the heterogeneous graph; The updated heterogeneous graph is input into the three-flow graph neural network; the three-flow graph neural network updates the semantic vector of each field in the heterogeneous graph based on whether the two fields corresponding to each edge belong to the same business entity; updates the statistical vector corresponding to each field in the heterogeneous graph based on the difference vector between the statistical vector of each field and the preset global distribution primitive vector; and updates the structure vector of each field based on the hop count corresponding to the structure vector of each field.

7. The method according to claim 2, characterized in that, The method further includes: Obtain the preset constraint vector corresponding to each field, which is determined based on the field's anomaly alarm label and / or business label; The preset constraint vector corresponding to each field is concatenated with the corresponding feature vector, and the feature vector is updated using the concatenated vector.

8. The method according to claim 1, characterized in that, The step of determining the target first field that matches each second field based on the second feature vector of each second field and the first feature vector of each first field includes: For each second field, the similarity between the second field and each first field is determined based on the second feature vector of the second field and the first feature vector of each first field; the first field whose similarity exceeds the similarity threshold is determined as the target first field that matches the second field.

9. The method according to claim 8, characterized in that, If there is no corresponding first field with a similarity exceeding the similarity threshold, the method further includes: Determine the difference between the highest similarity among the similarities between the second field and each of the first fields and the similarity threshold; If the difference is within a preset range, a work order awaiting review is generated, and a manual review notification is triggered.

10. The method according to claim 9, characterized in that, The method further includes: Determine the number of pending work orders currently generated; If the quantity reaches a preset quantity threshold, then a fine-tuning task for the feature extraction model is initiated based on the second field corresponding to the work order to be reviewed and the first field with the highest similarity to the second field.

11. The method according to claim 1, characterized in that, The training process of the three-flow graph neural network includes: Obtain sample pairs, each sample pair including an initial sample semantic vector, an initial sample structure vector, and an initial sample statistical vector corresponding to two sample fields respectively, whether the two sample fields are labels of the same node, and the boundary value of the sample pair; wherein, the boundary value is determined based on the number of sample fields that have a direct correlation with any sample field in the sample pair; Using a three-flow graph neural network to be trained, the initial sample semantic vector, initial sample structure vector, and initial sample statistical vector corresponding to the two sample fields included in the sample pair are updated respectively to obtain the sample semantic vector, sample structure vector, and sample statistical vector corresponding to the two fields respectively; Based on the sample semantic vectors corresponding to the two fields respectively, the semantic distance is determined, and based on the relationship between the semantic distance and the boundary value, and whether the two sample fields are labels of the same node, the semantic loss value is determined. The statistical loss value is determined based on the label, the sample statistical vectors corresponding to the two fields, and the boundary value. The structural loss value is determined based on the sample structure vector corresponding to the two sample fields included in the sample pair and the boundary value, respectively. Based on the semantic loss value, the statistical loss value, and the structural loss value, a total loss value is determined, and the parameters of the three-flow graph neural network are adjusted according to the total loss value.

Images