AI-combined code-level data blood relationship tracking method and system
By combining AI technology for code-level data lineage tracing, the problem of insufficient data lineage identification in multi-threaded and asynchronous environments by traditional methods is solved, achieving high-precision data lineage tracing and visualization, and improving the analytical capabilities of software systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-06
- Publication Date
- 2026-04-03
AI Technical Summary
Existing technologies lack the ability to semantically recognize code-level data lineage. Traditional tools struggle to track data flow in multi-threaded and asynchronous environments, resulting in gaps in the data lineage chain and failing to effectively support security auditing, compliance tracking, and business impact assessment.
By combining AI technology for code parsing, call graph construction, and dependency tracing, abstract syntax trees are extracted through syntax analysis to construct a global call graph. A weighted linear dependency model and adaptive classification learning are used for semantic recognition to generate data lineage links and present them in a visual format.
It enables accurate identification of data lineage in complex software systems, improves the accuracy and visualization of data flow analysis, supports security auditing and business impact assessment, and enhances the level of automation.
Smart Images
Figure CN121787416A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of software engineering and data management technology, specifically to a code-level data lineage tracing method and system that combines AI. Background Technology
[0002] As software systems expand their functionality and code structures become increasingly complex, data is frequently transferred between different modules, methods, and threads. Traditional log analysis, breakpoint debugging, and monitoring methods often only reflect partial runtime conditions and struggle to reconstruct the complete data path within the system. Existing static analysis methods can identify variable dependencies through abstract syntax trees, while dynamic analysis can trace runtime data flow through instrumentation or logging; these are widely used in debugging and performance analysis. However, these methods are typically limited to single threads or single modules and cannot fully cover data flows across threads, asynchronous calls, or complex business logic.
[0003] The core flaw in existing technologies lies in the lack of semantic recognition capabilities for code-level data lineage. Traditional tools can only display variable assignment and method call relationships, but cannot understand the true meaning of variables in business logic, nor can they automatically extract development intent from the source code. Especially in multi-threaded and asynchronous environments, the contextual relationships of variables are often fragmented; although data lineage links exist, their semantics are hollow, requiring developers to manually infer them. The lack of intelligent semantic recognition makes data flow analysis results difficult to use for high-level applications such as security auditing, compliance tracking, or business impact assessment, severely limiting the value and automation level of lineage analysis in practical engineering. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a code-level data lineage tracing method and system that combines AI. The technical problem this invention aims to solve is: how to use AI combined with code parsing, call graph construction, and dependency tracing to address the issues of low accuracy and poor visualization in traditional data lineage tracing methods.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a code-level data lineage tracing method combining AI, comprising:
[0006] S1. Perform structural parsing on the source code of the target software system to form a set of abstract syntax trees, wherein the structural parsing is performed using a syntax analysis extraction method;
[0007] S2. Perform call relationship analysis on the abstract syntax tree set to form a global call graph, wherein the call relationship analysis is performed using a graph construction algorithm;
[0008] S3. Perform dependency tracing on the variable nodes in the global call graph to form a data lineage chain. The dependency tracing process includes parameter passing, attribute assignment, and return value association analysis.
[0009] S4. Perform AI semantic recognition processing on the data lineage link to form a semantically enhanced result. The AI semantic recognition processing adopts an adaptive classification learning method.
[0010] S5. The semantic enhancement results are visualized to form a data lineage graph. The visualization process uses a dynamic drawing mechanism based on graph structure mapping.
[0011] Preferably, the syntax analysis extraction method includes lexical segmentation, syntax rule matching, and semantic node annotation, and the structural parsing processing steps are as follows:
[0012] S11. Perform the lexical segmentation on the source code to form a token sequence;
[0013] S12. The notation sequence is matched with the syntax rules to form a syntax unit tree;
[0014] S13. The semantic nodes of the syntax unit tree are labeled to form the abstract syntax tree set.
[0015] Preferably, the graph construction algorithm includes node extraction processing and edge generation processing. The node extraction processing identifies and extracts method definition statements from the abstract syntax tree set to form a method node set, and the edge generation processing extracts and calls the method node set to form the global call graph.
[0016] Preferably, the dependency tracking process employs a weighted linear dependency model formula, the calculation formula of which is:
[0017] .
[0018] in, Representing variables With variables The overall dependence strength between them is dimensionless. This represents the parameter transmission influence factor, which is dimensionless. This represents the attribute assignment influence factor, which is dimensionless. This indicates the return value is associated with a dimensionless factor. The weighting coefficients passed to the parameters are dimensionless. The weighting coefficients assigned to the attribute are dimensionless. The weighting coefficients associated with the return value are dimensionless and satisfy the following conditions: .
[0019] Preferably, the adaptive classification learning method includes feature extraction processing, sample training processing, and category update processing. The feature extraction processing includes semantic keyword extraction and analysis of variable names, method names, and annotation text to form a semantic feature set. The sample training processing performs multiple rounds of training on the semantic feature set to generate an initial classification set. The category update processing identifies and adjusts the initial classification set to generate the semantically enhanced result.
[0020] Preferably, the dynamic rendering mechanism includes frame refresh processing, incremental update processing, and node stabilization control processing. The frame refresh processing adjusts the rendering frequency according to the node change rate, the incremental update processing performs local rendering of the changed areas, and the node stabilization control processing maintains the continuity and display stability of the data lineage map through positional smoothing constraints.
[0021] Preferably, the visualization processing includes node layout calculation processing, path rendering processing, and interactive response processing.
[0022] A code-level data lineage tracing system that combines AI includes:
[0023] The source code parsing module parses and processes the source code of the target application system to form a preliminary set of data nodes. The source code includes Java language files and Python language files.
[0024] A call chain construction module is used to construct a global call graph by performing cross-module call construction on the initial data node set. The cross-module call construction includes call identification, cross-module path connection and context dependency tracking.
[0025] The data lineage analysis module performs dependency analysis on variables and object attributes in the global call graph to form a lineage chain. The dependency analysis includes the integration of parameter passing, return value assignment, and object attribute assignment.
[0026] An asynchronous link reconstruction module performs reconstruction and merging processing on the lineage link to generate semantically enhanced results. The reconstruction and merging processing includes asynchronous node identification, context identifier merging, and inter-thread path alignment.
[0027] The AI semantic recognition module performs semantic recognition and business meaning annotation on the semantic enhancement results to generate a lineage graph. The AI semantic recognition module adopts an adaptive classification learning method.
[0028] This invention provides a code-level data lineage tracing method and system that combines AI. It has the following beneficial effects:
[0029] This AI-integrated code-level data lineage tracing method and system achieves code-level data lineage tracing by performing structural analysis, call relationship analysis, variable dependency tracking, and AI semantic recognition on the source code of the target software system. Through syntax analysis, graph construction algorithms, weighted linear dependency models, and adaptive classification learning methods, it extracts and tracks the data flow paths within the software system.
[0030] Employing graph construction algorithms, weighted linear dependency models, and AI semantic recognition, this system accurately identifies data lineage relationships and generates lineage chains. Through dynamic rendering mechanisms, semantic enhancement processing, and interactive response functions, it optimizes the visualization of data lineages, improving the ability to understand and analyze complex data flows and business logic. Attached Figure Description
[0031] Figure 1 It is a flowchart of the core method for realizing an invention;
[0032] Figure 2 This is a schematic diagram of a system module structure for implementing an invention;
[0033] Figure 3 It is a flowchart for dependency tracking in the implementation of an invention. Detailed Implementation
[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0035] Example 1
[0036] like Figure 1-3 As shown, this embodiment of the invention provides a code-level data lineage tracing method combined with AI, including: S1. Performing structural parsing processing on the source code of the target software system to form an abstract syntax tree set. The structural parsing processing employs a syntax analysis extraction method. The syntax analysis extraction method includes lexical segmentation, syntax rule matching, and semantic node annotation. The steps of the structural parsing processing are as follows:
[0037] S11. Perform lexical segmentation on the source code to form a token sequence.
[0038] S12. Perform syntax rule matching on the token sequence to form a syntax unit tree.
[0039] S13. Semantic node annotations are performed on the syntax unit trees to form an abstract syntax tree set.
[0040] S2. Call relationship analysis is performed on the abstract syntax tree set to form a global call graph. This analysis employs a graph construction algorithm. The algorithm includes node extraction and edge generation. Node extraction identifies and extracts method definition statements from the abstract syntax tree set to form a set of method nodes. Edge generation extracts and calls from the method node set to form the global call graph.
[0041] S3. Dependency tracing is performed on the variable nodes in the global call graph to form data lineage. Dependency tracing includes parameter passing, attribute assignment, and return value association analysis. The dependency tracing uses a weighted linear dependency model formula, the calculation formula of which is:
[0042] .
[0043] in, Representing variables With variables The overall dependence strength between them is dimensionless. This represents the parameter transmission influence factor, which is dimensionless. This represents the attribute assignment influence factor, which is dimensionless. This indicates the return value is associated with a dimensionless factor. The weighting coefficients passed to the parameters are dimensionless. The weighting coefficients assigned to attributes are dimensionless. The weighting coefficients associated with the return value are dimensionless and satisfy the following conditions: .
[0044] S4. AI semantic recognition processing is applied to the data lineage to generate semantically enhanced results. This AI semantic recognition processing employs an adaptive classification learning method. The adaptive classification learning method includes feature extraction, sample training, and category update. Feature extraction involves semantic keyword extraction and analysis of variable names, method names, and annotation text to form a semantic feature set. Sample training involves multiple rounds of training on the semantic feature set to generate an initial classification set. Category update involves recognizing and adjusting the initial classification set to generate the semantically enhanced results.
[0045] S5. The semantic enhancement results are visualized to form a data lineage graph. This visualization process employs a dynamic rendering mechanism based on graph structure mapping. The dynamic rendering mechanism includes frame refresh processing, incremental update processing, and node stabilization control processing. Frame refresh processing adjusts the rendering frequency according to the node change rate; incremental update processing performs local rendering of changed areas; and node stabilization control processing maintains the coherence and display stability of the data lineage graph through positional smoothing constraints. The visualization process also includes node layout calculation processing, path rendering processing, and interactive response processing.
[0046] A code-level data lineage tracing system that combines AI includes:
[0047] The source code parsing module parses and processes the source code of the target application system to form a preliminary set of data nodes. The source code includes Java language files and Python language files.
[0048] The call chain construction module performs cross-module call construction on the initial data node set to generate a global call graph. Cross-module call construction includes call identification, cross-module path connection, and context dependency tracking.
[0049] The data lineage analysis module performs dependency analysis on variables and object attributes in the global call graph to form a lineage chain. The dependency analysis includes the integration of parameter passing, return value assignment, and object attribute assignment.
[0050] The asynchronous link reconstruction module reconstructs and merges lineage links to generate semantically enhanced results. The reconstruction and merging process includes asynchronous node identification, context identifier merging, and inter-thread path alignment.
[0051] The AI semantic recognition module performs semantic recognition and business meaning annotation on the semantic enhancement results to generate a lineage graph. The AI semantic recognition module adopts an adaptive classification learning method.
[0052] This AI-powered code-level data lineage tracing method accurately tracks the interactions and data flow between modules in a software system through multiple steps, including call relationship analysis, data lineage chain formation, AI semantic recognition, visualization, and asynchronous link reconstruction. Call relationship analysis constructs a global call graph to comprehensively identify call relationships between modules. Data lineage chain analysis accurately analyzes dependencies such as parameter passing and attribute assignment to help trace data flow and its mutual influence. AI semantic recognition improves the accuracy and operability of data lineage tracing. Visualization clearly and stably displays the data lineage graph through a dynamic drawing mechanism and provides interactive functions to enhance the user experience. The asynchronous link reconstruction module solves the challenge of lineage chain identification in asynchronous operations and multi-threaded environments, ensuring accurate data flow tracing even in complex execution environments.
[0053] Example 2
[0054] This embodiment is based on a code-level data lineage tracing method combined with AI. Through a weighted linear dependency model, it quantifies the dependency relationship between deposit amount, balance after deducting fees, and loan amount in a banking system. The specific implementation is as follows:
[0055] In a banking system, after a user deposits money, a handling fee is deducted to calculate the loan amount. This example analyzes and tracks the dependencies between changes in user deposits, handling fee deductions, balance, and loan amount.
[0056] 1. Define data flow and related methods
[0057] Method A: Calculate the total deposit amount; Method B: Deduct the handling fee and return the remaining balance; Method C: Calculate the loan amount based on the remaining balance.
[0058] The following are the specific procedures and data:
[0059] User deposit: The deposit amount is 5,000 yuan.
[0060] After deducting the handling fee: The handling fee rate is 2%, so the handling fee is 5000 × 2% = 100 yuan. The remaining balance after deducting the handling fee is 5000 - 100 = 4900 yuan.
[0061] Loan amount calculation: The loan amount is 50% of the balance after deducting handling fees, that is, 4900 × 50% = 2450 yuan.
[0062] 2. Dependency tracing analysis
[0063] Based on the data flow, the inputs, outputs, and dependencies of each method are as follows:
[0064] Method A: Input is the user's deposit amount of 5000 yuan, output is the total deposit amount of 5000 yuan.
[0065] Method B: The input is the output of Method A, with a total deposit of 5000 yuan and the output is the balance of 4900 yuan after deducting handling fees.
[0066] Method C: The input is the output of Method B, the balance after deducting handling fees is 4900 yuan, and the output is the loan amount of 2450 yuan.
[0067] 3. Calculation of Dependency Strength
[0068] Dependency tracking uses a weighted linear dependency model, the calculation formula of which is as follows:
[0069] .
[0070] in, Representing variables With variables The overall dependence strength between them is dimensionless. This represents the parameter transmission influence factor, which is dimensionless. This represents the attribute assignment influence factor, which is dimensionless. This indicates the return value is associated with a dimensionless factor. The weighting coefficients passed to the parameters are dimensionless. The weighting coefficients assigned to attributes are dimensionless. The weighting coefficients associated with the return value are dimensionless and satisfy the following conditions: .
[0071] Weighting coefficients: The transfer of deposit amount is a key factor affecting subsequent calculations, indicating that parameter transfer accounts for 40% of the entire dependency relationship. The assignment of balance has a smaller impact on loan amount, usually involving fewer complex assignment operations, indicating that attribute assignment accounts for 30%. The return value of the calculation relationship between balance and loan amount directly affects other methods, indicating that the impact of return value association accounts for 30%.
[0072] According to the analysis, the weighting coefficients of the parameter transfer Weighting coefficients for attribute assignment The weight coefficient associated with the return value .
[0073] Calculate the dependence strength of method A on method B:
[0074] The output of method A affects the balance in method B after deducting transaction fees. The degree of dependence of method A on method B reflects the impact of deposit amount on the balance.
[0075] Parameter transfer influencing factors Method B depends entirely on the output deposit amount of Method A, therefore .
[0076] Attribute assignment influence factor The balance in method B is directly assigned from the output of method A, therefore .
[0077] Return value associated impact factor : There is no direct impact from the return value, therefore .
[0078] Dependency strength calculation:
[0079]
[0080] Calculate the dependence strength of method B on method C:
[0081] The output of method B affects the loan amount in method C, and the degree of dependence of method B on method C reflects the impact of the balance on the loan amount.
[0082] Parameter transfer influencing factors Method C depends entirely on the output balance of method B, therefore .
[0083] Attribute assignment influence factor In method C, the balance is passed through property assignment, but not directly assigned. .
[0084] Return value associated impact factor Method C depends entirely on the balance returned by method B, therefore .
[0085] Dependency strength calculation:
[0086]
[0087] Based on the above calculations, the following dependency strengths are obtained:
[0088] The dependence strength of method A on method B is The dependency strength of method B on method C is .
[0089] 4. Results visualization
[0090] Visualization tools can be used to visualize data flow and dependencies. Graphical representations of dependencies between variables provide an intuitive understanding of the path and intensity of data flow.
[0091] By tracing the dependencies between data nodes in the banking system, it was found that deposit amount has the greatest impact on subsequent calculations, with a dependency strength of 0.4, indicating the crucial role of deposit amount in the entire data flow. Return value association and attribute assignment have lower dependency strengths, at 0.3 respectively, indicating that return value association and attribute assignment have some influence on the calculation results, but are not as important as parameter passing. The above analysis methods help improve data transparency and provide a basis for system optimization and decision support.
[0092] Example 3
[0093] This embodiment is based on a code-level data lineage tracing method combined with AI. By integrating an AI-powered code-level data lineage tracing system, it enables automatic tracking, analysis, and visualization of data flows within applications, helping developers efficiently manage and understand data dependencies in their code. The specific implementation is as follows:
[0094] A financial application processes user transaction data, involving data computation and output. This example will demonstrate how to achieve data flow tracking and visualization through a code-level data lineage tracing system that incorporates AI.
[0095] 1. Scene Description
[0096] There are two main modules in financial applications:
[0097] Data processing module A: Responsible for receiving the transaction amount and tax rate input by the user, and calculating the total transaction amount.
[0098] Output module B: Responsible for displaying the calculation results.
[0099] The specific business process is as follows:
[0100] The user enters a transaction amount of 1,000 yuan on the interface, with a tax rate of 15%.
[0101] After receiving the transaction amount and tax rate as input parameters, data processing module A performs the calculation using the following formula:
[0102]
[0103]
[0104] The calculated result of 1150 yuan is passed to output module B, which is responsible for displaying the result to the user.
[0105] 2. Source code parsing and abstract syntax tree generation
[0106] The system parses the application's source code and generates an abstract syntax tree (AST). An AST is a tree-like representation of the source code structure, helping the system understand the code's structure and syntax.
[0107] Lexical segmentation: The system breaks down the source code into symbols such as method names, variable names, and operators.
[0108] Syntax rule matching: The system matches the tokens with syntax rules to identify various functional modules, methods, variables and other elements.
[0109] Semantic node annotation: The system adds semantic information to nodes such as methods and variables to help understand the specific role of each node.
[0110] 3. Call relationship analysis and global call graph generation
[0111] The system analyzes the method call relationships between modules and constructs a global call graph to help the system understand how modules and methods depend on and call each other.
[0112] Node extraction: The system identifies key nodes such as data processing methods and data output methods.
[0113] Simultaneous generation: The system recognizes that the data processing method will call the total amount calculation method, and the data processing method itself will call the output data method to display the calculation results.
[0114] Ultimately, the system generates a global call graph: process data - calculate total amount - output data.
[0115] 4. Variable dependency tracking and data lineage construction
[0116] After obtaining the global call graph, the system traces the variable dependencies in the code and constructs the data lineage.
[0117] Parameter passing: In the data processing method, the transaction amount and tax rate are passed as parameters to the method that calculates the total amount.
[0118] Assignment Relationship: In the method for calculating the total amount, the input transaction amount and tax rate are used to calculate the total amount, and the calculation result is assigned to the result variable.
[0119] Return value association: The result variable is returned to the data processing method and passed to the output data method.
[0120] 5. AI Semantic Augmentation and Adaptive Classification Learning
[0121] To optimize the understanding of data flow, the system introduces AI technology for semantic enhancement. Through AI, the system understands the business meaning of elements such as variable names, method names, and comments.
[0122] Feature extraction: The system extracts information such as variable names and method names from the code and constructs a semantic set based on these features. The system identifies that the data processing method is used to calculate transaction amounts, while the data output method is used to display results.
[0123] Sample training: The system learns the actual business functions of each method and variable in the application by training with historical code samples.
[0124] Category Updates: The system continuously updates its classification model to adapt to new code or business requirements, enhancing its ability to understand and track data flows.
[0125] After AI semantic enhancement, the semantic information of the data lineage is made clearer:
[0126] 6. Visualization of the data lineage chart
[0127] The system presents the data lineage to users in a graphical format. These visual diagrams help developers, operations personnel, and business analysts quickly understand the data flow path.
[0128] Node layout calculation: Each method or function is treated as a node and arranged in the order of execution.
[0129] Path rendering: A data flow path is formed from the data processing node to the total amount calculation node, and then to the output data node.
[0130] Interactive response: When a user clicks on a node, they can view the node's input, output, and related dependency information.
[0131] 7. Asynchronous Link Reconstruction and Optimization
[0132] When a program involves multithreading or asynchronous calls, the system will handle the following scenarios:
[0133] Asynchronous node identification: One method is to identify asynchronous operations when they are executed asynchronously.
[0134] Context identifier merging: The system will merge the context information of different threads or asynchronous operations to ensure that the data lineage is not lost.
[0135] Inter-thread path alignment: For asynchronous or multi-threaded calls, the system aligns the data flow paths between threads to ensure the integrity of the data lineage.
[0136] By implementing an AI-powered code-level data lineage tracing system, the system accurately tracks and analyzes data flow paths within programs, identifies method call relationships and variable dependencies, and constructs a clear data lineage. The system combines abstract syntax tree parsing, call relationship analysis, variable dependency tracing, and AI semantic enhancement technologies to improve the understanding and analysis of complex business logic and data flows. Through visualization, developers can intuitively grasp the entire data flow process, optimizing code understanding, debugging, and maintenance efficiency, and improving overall development quality and work efficiency.
[0137] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A code-level data lineage tracing method combining AI, characterized in that, include: S1. Perform structural parsing on the source code of the target software system to form a set of abstract syntax trees, wherein the structural parsing is performed using a syntax analysis extraction method; S2. Perform call relationship analysis on the abstract syntax tree set to form a global call graph, wherein the call relationship analysis is performed using a graph construction algorithm; S3. Perform dependency tracing on the variable nodes in the global call graph to form a data lineage chain. The dependency tracing process includes parameter passing, attribute assignment, and return value association analysis. S4. Perform AI semantic recognition processing on the data lineage link to form a semantically enhanced result. The AI semantic recognition processing adopts an adaptive classification learning method. S5. The semantic enhancement results are visualized to form a data lineage graph. The visualization process uses a dynamic drawing mechanism based on graph structure mapping.
2. The code-level data lineage tracing method combining AI according to claim 1, characterized in that: The syntax analysis and extraction method includes lexical segmentation, syntax rule matching, and semantic node annotation. The steps of the structure parsing process are as follows: S11. Perform the lexical segmentation on the source code to form a token sequence; S12. The notation sequence is matched with the syntax rules to form a syntax unit tree; S13. The semantic nodes of the syntax unit tree are labeled to form the abstract syntax tree set.
3. The code-level data lineage tracing method combining AI according to claim 1, characterized in that: The graph construction algorithm includes node extraction processing and edge generation processing. The node extraction processing identifies and extracts the method definition statements of the abstract syntax tree set to form a method node set. The edge generation processing extracts and calls the method node set to form the global call graph.
4. The code-level data lineage tracing method combining AI according to claim 1, characterized in that: The dependency tracking process employs a weighted linear dependency model, the calculation formula of which is as follows: ， in, Representing variables With variables The overall dependence strength between them This indicates the parameter passing influence factor. Indicates the influence factor of attribute assignment. Indicates the return value correlation factor. The weighting coefficients passed to the parameters. The weighting coefficients assigned to the attribute. The weight coefficients associated with the return value satisfy the following conditions: .
5. The code-level data lineage tracing method combining AI according to claim 1, characterized in that: The adaptive classification learning method includes feature extraction, sample training, and category update. The feature extraction includes semantic keyword extraction and analysis of variable names, method names, and annotation text to form a semantic feature set. The sample training involves multiple rounds of training on the semantic feature set to generate an initial classification set. The category update involves identification and adjustment of the initial classification set to generate the semantically enhanced result.
6. The code-level data lineage tracing method combining AI according to claim 1, characterized in that: The dynamic rendering mechanism includes frame refresh processing, incremental update processing, and node stabilization control processing. The frame refresh processing adjusts the rendering frequency according to the node change rate. The incremental update processing performs local rendering of the changed areas. The node stabilization control processing maintains the continuity and display stability of the data lineage map through positional smoothing constraints.
7. The code-level data lineage tracing method combining AI according to claim 1, characterized in that: The visualization processing includes node layout calculation, path rendering, and interactive response processing.
8. A code-level data lineage tracing system combining AI, characterized in that, include: The source code parsing module parses and processes the source code of the target application system to form a preliminary set of data nodes. The source code includes Java language files and Python language files. A call chain construction module is used to construct a global call graph by performing cross-module call construction on the initial data node set. The cross-module call construction includes call identification, cross-module path connection and context dependency tracking. The data lineage analysis module performs dependency analysis on variables and object attributes in the global call graph to form a lineage chain. The dependency analysis includes the integration of parameter passing, return value assignment, and object attribute assignment. An asynchronous link reconstruction module performs reconstruction and merging processing on the lineage link to generate semantically enhanced results. The reconstruction and merging processing includes asynchronous node identification, context identifier merging, and inter-thread path alignment. The AI semantic recognition module performs semantic recognition and business meaning annotation on the semantic enhancement results to generate a lineage graph. The AI semantic recognition module adopts an adaptive classification learning method.