A Data Quality Governance Method and System Based on AI Agent

Abstract

The application discloses an AI Agent-based data quality management method and system, belongs to the technical field of artificial intelligence and data management, constructs an initial data semantic distribution graph, extracts cross-dimension correlation characteristic factors, constructs a multi-scale quality abnormality sensitive factor matrix, and embeds time sequence evolution weights to form a dynamic characteristic evolution track; an AI Agent is used to perceive modeling on the evolution track, generate a multi-level quality risk heat map, extract a bias dense area and construct an abnormality propagation path set; in combination with upstream and downstream data link and task flow information, potential influence factor weights are calculated, a cause-effect traceability graph is constructed, and a rectification strategy label is injected into a key field node; the AI Agent independently selects a suitable strategy combination according to the rectification strategy label, and carries out online intervention on a target data segment; the application realizes closed-loop management of data quality problems from perception, judgment to intervention and feedback, and has the advantages of self-adaptation, strong interpretability and the like.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data governance technology, specifically to a data quality governance method and system based on AI Agent. Background Technology

[0002] With the increasing prevalence of big data applications, data quality has become a key bottleneck restricting the accuracy and stability of artificial intelligence models. Especially in the fusion and analysis of multi-source heterogeneous data, common problems such as missing data, redundancy, conflicts, falsification, and delayed updates can significantly amplify the prediction errors of AI models. Traditional data quality governance methods often rely on rule matching, manual review, or offline statistics, which cannot meet the multidimensional requirements for real-time data quality, accuracy, and completeness in high-frequency, dynamically updated scenarios.

[0003] Currently, in fields such as smart cities, industrial IoT, and financial risk control, as AI agents increasingly assume the tasks of automated decision-making and control, their reliance on data quality is rapidly increasing. However, existing methods cannot achieve deep governance of fluid, multi-layered, and semantically complex data, leading to AI agents making incorrect judgments based on unreliable data and creating security risks. For example, in intelligent transportation systems, if abnormal sensor data or update delays cause an AI agent to misjudge road conditions and make incorrect dispatching decisions, it could result in serious accidents. Furthermore, most existing data quality assessments remain at the level of single-dimensional indicators, lacking the ability to trace the structure and causality of quality anomalies, and failing to provide an explainable quality feedback mechanism for downstream AI agents. Summary of the Invention

[0004] The purpose of this invention is to provide a data quality governance method and system based on AI Agent to address the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a data quality governance method based on AI Agent, comprising:

[0006] Step S1: Obtain structured or semi-structured data streams, and use a preset data semantic encoder to embed semantic vectors into each data segment to construct an initial data semantic distribution map;

[0007] Step S2: Based on the initial data semantic distribution map, extract cross-dimensional correlation feature factors, construct a multi-scale quality anomaly sensitivity factor matrix, and embed temporal evolution weights in the feature space to form a dynamic feature evolution trajectory;

[0008] Step S3: Use an AI Agent to generate a multi-level quality risk heat map from the dynamic feature evolution trajectory, and construct an anomaly propagation path set based on the areas with dense deviations;

[0009] Step S4: Based on the set of abnormal propagation paths, combined with upstream and downstream data links and task flow graphs, calculate the weights of potential influencing factors, construct a causal tracing graph, and inject correction strategy labels into key nodes in the path.

[0010] Step S5: Based on the causal origination map and the correction strategy label, the AI ​​Agent autonomously selects an appropriate combination of governance strategies to intervene online in the target data segment;

[0011] Step S6: Generate trusted labels for the processed data fragments and feed them back to the AI ​​Agent model input interface to enable the AI ​​Agent model to prioritize and dynamically update high-quality data.

[0012] Preferably, the construction of the initial data semantic distribution map includes:

[0013] By connecting to multiple heterogeneous data source nodes, including but not limited to IoT sensors, API interfaces, databases or log files, structured and semi-structured data can be collected in real time.

[0014] Data preprocessing is performed on the collected data fragments of various types to generate a data slice sequence with a unified input format;

[0015] The data slice sequence is input into a preset data semantic encoder, which adopts a multi-layer bidirectional Transformer structure and integrates a context attention mechanism to model the implicit semantic relationships in the data.

[0016] Using the semantic vector embedding results output by the data semantic encoder, a semantic adjacency graph is constructed in the vector space based on semantic similarity, and an initial data semantic distribution map is generated by combining it with time series labels.

[0017] Preferably, the step of extracting cross-dimensional correlation feature factors and constructing a multi-scale quality anomaly sensitivity factor matrix includes:

[0018] In the initial data semantic distribution map, principal component analysis is performed on the semantic vectors of all nodes. The top K principal component vectors with a cumulative explained variance of more than 95% are selected and used as low-dimensional feature basis.

[0019] On the low-dimensional feature basis, the mutual information values ​​between different field types are calculated respectively, and a cross-dimensional correlation matrix between fields is constructed through standardization.

[0020] For each pair of fields in the correlation matrix, the change sensitivity under different time windows is calculated. The change sensitivity is the ratio of the standard deviation of the feature perturbation within the moving average window to the baseline difference, thus obtaining a set of sensitive factors under different time scales.

[0021] The sensitive factor set is organized into a three-dimensional tensor according to field dimension, time scale, and perturbation intensity to construct a multi-scale quality anomaly sensitive factor matrix. Then, an evolution weight is assigned to each sensitive factor using a sliding time window mechanism, and finally, the corresponding dynamic feature evolution trajectory sequence is generated.

[0022] Preferably, the step of generating a multi-level quality risk heatmap using an AI Agent based on the dynamic feature evolution trajectory includes:

[0023] The dynamic feature evolution trajectory sequence is input into the feature deviation perception model built on graph neural network, and the evolution trend gradient and mutation intensity of each field pair in the continuous time window are calculated to generate an initial two-dimensional heat map with time dimension as the horizontal axis, field combination as the vertical axis, and deviation intensity as heat.

[0024] An adaptive threshold segmentation algorithm is applied to the initial two-dimensional heat map to extract deviation hotspot areas. The threshold is dynamically calculated based on the deviation intensity of the 95th percentile according to the overall heat distribution, and multiple independent sets of high-risk areas are obtained.

[0025] In each high-risk area, the historical change path of the corresponding field combination in the evolution trajectory is tracked. A directed path graph containing the starting field, the ending field, and the intermediate field chain is constructed by combining the field correlation matrix and the time sequence. Each hop in the path represents a potential anomalous transmission.

[0026] All constructed anomaly propagation paths are structurally merged to form an anomaly propagation path set with fields as nodes and anomaly relationships as edges. A path scoring function is introduced to prioritize the paths by combining path length, hop count weight, and average popularity.

[0027] Preferably, the step of combining upstream and downstream data links and task flow diagrams to calculate the weights of potential influencing factors and construct a causal origination graph includes:

[0028] The field nodes involved in each path of the anomaly propagation path set are analyzed. Based on the predefined data flow mapping relationship table, the upstream and downstream source and target nodes of each field in the data link are retrieved, and the field-level upstream and downstream dependency links are constructed.

[0029] By combining the field call records of each processing task in the task flow diagram, the frequency of reference and scope of influence of abnormal fields in each task node are statistically analyzed, and the potential influence factor weight of each field in the global task flow is calculated using the TF-IDF weighting method.

[0030] Using field nodes as graph nodes and data flow and task dependencies as edges, a unified causal tracing graph is constructed by integrating anomaly propagation path graph, upstream and downstream field link graph, and task field call graph. The causal tracing graph is a directed weighted graph.

[0031] For key field nodes in the causal origination graph, based on their position in the abnormal path and the weight score of the influencing factor, the corresponding correction strategy label is matched to form a field-level correction label set and injected into the corresponding node.

[0032] Preferably, based on the causal origination map and correction strategy labels, the AI ​​Agent autonomously selects an appropriate combination of governance strategies to intervene online in the target data segment, including:

[0033] Based on the abnormal propagation weights, strategy label types, and historical correction effect scores of key field nodes in the causal tracing graph, a multi-objective decision model is invoked to construct a strategy adaptation priority matrix, which is used to indicate the matching degree and expected intervention benefits of each correction strategy with its applicable fields.

[0034] For each candidate correction strategy, a virtual intervention simulation is performed. The predicted output change of the target field after intervention is simulated based on the pre-trained model. The strategy value estimate is updated using the error convergence degree and stability as the reward function.

[0035] Based on the policy adaptation priority matrix and the policy value of the reinforcement learning output, the optimal combination of governance policies is selected.

[0036] The governance strategies are combined and applied to the original values ​​or semantic embedding vectors of the target data fragments to perform online intervention operations.

[0037] Preferably, trusted labels are generated for the processed data fragments and fed back to the AI ​​Agent model input interface, including:

[0038] Based on the results after the implementation of the governance strategy, the deviation convergence and stability change rate of key fields before and after governance are calculated for each data segment, and a credibility scoring index is defined. The credibility scoring index consists of the deviation reduction magnitude, the convergence degree of field value fluctuation range, and the persistence of intervention impact.

[0039] Based on the credibility scoring index, a multi-level credibility label level is set. The scoring results are divided into three levels: "high credibility", "medium credibility" and "low credibility" using a segmented scoring method. The corresponding data segment ID, intervention strategy type, field list and scoring details are recorded in the label.

[0040] The trusted tags are attached to the metadata structure of the corresponding data fragments, and a hierarchical caching mechanism is used to prioritize writing high-trust data into the AI ​​Agent model input queue, while low-trust data enters the asynchronous processing channel to wait for supplementary verification.

[0041] This invention also provides a data quality governance system based on AI Agent, comprising:

[0042] Data acquisition module: Acquires structured or semi-structured data streams, uses a preset data semantic encoder to embed semantic vectors into each data segment, and constructs an initial data semantic distribution map;

[0043] Evolutionary trajectory generation module: Based on the initial data semantic distribution map, extract cross-dimensional correlation feature factors, construct a multi-scale quality anomaly sensitivity factor matrix, and embed temporal evolution weights in the feature space to form a dynamic feature evolution trajectory;

[0044] Anomaly propagation path set construction module: The dynamic feature evolution trajectory is used to generate a multi-level quality risk heat map using an AI Agent, and an anomaly propagation path set is constructed based on the deviation-dense region;

[0045] Weight calculation module: Based on the set of abnormal propagation paths, combined with upstream and downstream data links and task flow graphs, calculate the weights of potential influencing factors, construct a causal tracing graph, and inject correction strategy labels into key nodes in the path.

[0046] Intervention module: Based on the causal origin map and correction strategy labels, the AI ​​Agent autonomously selects an appropriate combination of governance strategies to intervene online in the target data segment;

[0047] Update module: Generates trusted labels for the processed data fragments and feeds them back to the AI ​​Agent model input interface, enabling the AI ​​Agent model to prioritize and dynamically update high-quality data.

[0048] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0049] 1. This invention constructs an initial data semantic distribution map, extracts cross-dimensional dynamic evolution features, and combines it with AI agents for risk heat map perception, causal path construction, and multi-strategy intervention, achieving closed-loop governance of data quality issues throughout the entire process. Compared with traditional quality control methods based on rules or static indicators, this invention possesses strong dynamism, adaptability, and structural interpretability, significantly improving the accuracy and response speed of data anomaly identification and effectively reducing the risk of AI model misjudgment due to data anomalies.

[0050] 2. This invention, through a trusted labeling mechanism and an input priority feedback channel, enables the AI ​​Agent to prioritize high-quality data. Furthermore, by continuously optimizing the governance strategy selection logic through reinforcement learning, it achieves bidirectional collaborative optimization of model input quality and prediction accuracy. This method possesses good scalability and deployment adaptability, demonstrating significant engineering application value and industrial promotion potential. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0052] Figure 1 This is a mind map of the method of the present invention.

[0053] Figure 2 This is a mind map of the system modules of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0055] Example 1, please refer to Figure 1 As shown in this embodiment, a data quality governance method based on AI Agent includes:

[0056] Step S1: Obtain structured or semi-structured data streams, and use a preset data semantic encoder to embed semantic vectors into each data segment to construct an initial data semantic distribution map;

[0057] Step S2: Based on the initial data semantic distribution map, extract cross-dimensional correlation feature factors, construct a multi-scale quality anomaly sensitivity factor matrix, and embed temporal evolution weights in the feature space to form a dynamic feature evolution trajectory;

[0058] Step S3: Use an AI Agent to generate a multi-level quality risk heat map from the dynamic feature evolution trajectory, and construct an anomaly propagation path set based on the areas with dense deviations;

[0059] Step S4: Based on the set of abnormal propagation paths, combined with upstream and downstream data links and task flow graphs, calculate the weights of potential influencing factors, construct a causal tracing graph, and inject correction strategy labels into key nodes in the path.

[0060] Step S5: Based on the causal origination map and the correction strategy label, the AI ​​Agent autonomously selects an appropriate combination of governance strategies to intervene online in the target data segment;

[0061] Step S6: Generate trusted labels for the processed data fragments and feed them back to the AI ​​Agent model input interface to enable the AI ​​Agent model to prioritize and dynamically update high-quality data.

[0062] Data is collected from actual operating environments, and the sources include:

[0063] Configure and connect a temperature and humidity sensor and an image camera to collect readings and image description data every 10 seconds;

[0064] The daily transaction logs are retrieved from the business database, and table data with fields "User ID, Transaction Amount, Transaction Time, and Transaction Status" is obtained by retrieving the data in ascending order of time using SQL statements.

[0065] By calling the RESTful API interface of the city data open platform, weather information in JSON format is requested every 60 seconds, with fields including "temperature, humidity, air quality, and update time".

[0066] The data formats mentioned above are structured tables (CSV), semi-structured text (JSON), and mixed log streams (image data description + numerical values), which are all read and then uniformly entered into the data processing flow.

[0067] The data preprocessing process is as follows:

[0068] Field mapping: Create a standard field mapping table to uniformly map "userId" to "user ID" and "temp" to "temperature", and use a hash table to implement the mapping logic.

[0069] Null value marker: Adds a special placeholder "-9999" for missing fields;

[0070] Time alignment: All data is aligned to a granularity of "minutes". For sensor data collected more than once a minute, the last record within the last 60 seconds is used; for low-frequency API data, forward padding is applied to the timeline.

[0071] Format encapsulation: Organize each piece of processed data into key-value pairs, for example:

[0072] User ID: 23981

[0073] Transaction amount: 512.7

[0074] Trading Time: 2025-08-29 13:45:00

[0075] Temperature: 31.2

[0076] Air quality: Good.

[0077] The final result is a unified sequence of data slices, each slice representing a complete data fragment with standardized fields and timestamps.

[0078] Embedding data slices into semantic vectors aims to map different types of data into a unified vector space, facilitating semantic similarity calculation. This task is accomplished using a Transformer-based encoder. The specific implementation is as follows:

[0079] Field embedding matrix construction: Numerical fields (such as temperature, transaction amount) are normalized: the minimum and maximum values ​​of each field in the most recent 100 data points are calculated using a sliding window method, and linearly mapped to the interval [0,1]; text fields (such as air quality) are encoded using dictionary encoding, such as "excellent" as [1,0,0], "good" as [0,1,0], and "poor" as [0,0,1]; time fields are converted into three values: "hour, minute, and whether it is a holiday", and merged into one vector.

[0080] Input embedding generation: The above field values ​​are concatenated into a vector. For example, a data slice includes 6 fields, with a total vector length of 28 dimensions; a position vector is added, and the position code is calculated using sine and cosine functions, and then weighted and superimposed onto the original input item by item.

[0081] Transformer encoding: A four-layer encoder structure is constructed, with each layer containing a feedforward network, residual connections, and multi-head attention (a total of eight heads). Each head computes attention weights across different dimensions in parallel. The outputs are concatenated and linearly mapped to maintain a 256-dimensional dimension. All encoder weights are obtained through pre-training, with the training objective being to minimize the prediction error of data within the same time period (i.e., predicting the values ​​of fields such as "air quality" at the next moment). Each data slice, after being processed by the encoder, yields a 256-dimensional semantic vector, which is subsequently used for graph structure construction.

[0082] Calculate the similarity of all semantic vectors pairwise, using cosine similarity as the similarity metric. The specific calculation method is as follows:

[0083] Dot product calculation: Performing a dot product operation on two 256-dimensional semantic vectors;

[0084] Modulus normalization: Calculate the Euclidean norm of the two vectors respectively;

[0085] Division evaluation: Divide the dot product by the product of the moduli to obtain the similarity score, which is in the range of [0,1].

[0086] Using 0.85 as the threshold, if the similarity between two data slices is greater than or equal to 0.85, they are considered to have a semantic connection and recorded as an adjacency relationship. After construction, a semantic adjacency list is formed. If a node has no other nodes with a similarity greater than the threshold, the node is treated as an isolated point, and a semantic smoothing process will be added later.

[0087] Based on the semantic adjacency list, time hierarchy information is added, and all data slices are sorted by timestamp, with each level being 5 minutes long and increasing downwards.

[0088] The map construction process is as follows:

[0089] Horizontal dimension: Semantically similar data slices are clustered according to their adjacency relationship, and distributed horizontally;

[0090] Vertical axis dimension: Data slices within each time window are stacked in chronological order;

[0091] Depth dimension: The semantic vector value of each node serves as a node attribute, which is used for further processing by the graph neural network.

[0092] To enhance the expressive power of this map, an attention mechanism is introduced:

[0093] For each node, calculate the vector similarity between its neighboring nodes and normalize it using softmax;

[0094] Aggregate the semantic information of neighbors according to similarity weights and update the expression of the current node;

[0095] This operation is repeated three times to enhance the node's ability to recognize its semantic context.

[0096] The final output semantic distribution map of the data not only shows the semantic similarity structure between data, but also preserves the temporal evolution trend.

[0097] In the initial data semantic distribution map, each node contains a 256-dimensional semantic vector output by the semantic encoder. This vector expresses the contextual semantic information of the data fragment in the semantic space. In order to reduce the computational dimensionality, eliminate noise redundancy, and retain the core feature information, principal component analysis (PCA) needs to be performed on this vector set.

[0098] The specific process is as follows:

[0099] Let N data nodes correspond to vectors. to Each vector has a length of 256;

[0100] Construct a matrix M, where the rows are data nodes and the columns are the dimensions of the semantic vectors;

[0101] Center the matrix M, that is, subtract the mean of each column from each column;

[0102] Calculate the covariance matrix C;

[0103] Perform eigenvalue decomposition on C and select the top K principal components with eigenvalues ​​arranged from largest to smallest.

[0104] The selection process stops when the cumulative explained variance reaches 95%, which means that the sum of the eigenvalues ​​of the selected K features divided by the sum of all eigenvalues ​​is greater than or equal to 0.95.

[0105] A low-dimensional mapping basis is constructed using K principal components, which maps the original 256-dimensional vector to a K-dimensional feature space, where K is typically between 10 and 20.

[0106] Finally, a low-dimensional semantic representation vector for each node is obtained, which is used for subsequent cross-dimensional analysis and sensitivity calculation.

[0107] Under principal component mapping, the semantic information of each node is compressed into a K-dimensional vector, while each dimension still retains semantic relevance to the specific data field. To uncover hidden relationships between fields, this step calculates the mutual information values ​​between all fields and constructs a field correlation matrix.

[0108] The implementation details are as follows:

[0109] Suppose there are a total of m fields, including continuous fields (such as temperature, amount), category fields (such as weather level, transaction status), and time fields (such as hour, whether it is a working day);

[0110] For each pair of fields (x, y), the mutual information value MI(x, y) is estimated using the joint probability density function and the marginal probability density function, and an equal-binning and entropy gain-based distribution modeling strategy is adopted.

[0111] All mutual information values ​​are normalized and scaled to between 0 and 1;

[0112] Construct an m × m dimension association matrix A, where A[i][j] represents the semantic dependency strength between field i and field j.

[0113] The existence of pairs with high mutual information among the fields indicates that their changing trends may be highly consistent and susceptible to interference from the same quality problem. Therefore, this matrix provides the basic dimension for the subsequent construction of quality anomaly factors.

[0114] Since real-world data quality issues often manifest as temporal disturbances and fluctuations, such as sensor jitter and network latency, this step aims to characterize the abnormal sensitivity of fields to changes over time. The core task is to define and calculate the Variation Sensitivity Index (VSI), which is defined as follows:

[0115] VSI = Standard deviation of field perturbation within the moving window / Baseline mean deviation of the same field;

[0116] Implementation details include:

[0117] For each pair of high mutual information fields (x, y) in the correlation matrix, a sliding window of different scales is constructed with time windows of 5 minutes, 15 minutes and 60 minutes;

[0118] Within each window, calculate the standard deviation σ of the numerical perturbation of field x within that window. X At the same time, its mean was calculated compared with the historical benchmark. The difference ;

[0119] Sensitivity ,like If the value is close to 0, set a minimum threshold to avoid a denominator of 0.

[0120] Record the VSI value of each pair of fields at each time scale to form a set of field-time scale-sensitivity triplet;

[0121] Organize all triples into a tensor structure T[i][j][t], where i and j are field pair indices, t is the scale number, and the value is the corresponding VSI sensitivity.

[0122] This tensor set is the set of multi-scale quality anomaly sensitivity factors, which can reflect which field combinations are more susceptible to perturbation at different time granularities.

[0123] To express the anomalous evolution trends among fields, it is necessary not only to identify the sensitive factors themselves but also to track their changes over time. This step further embeds Temporal Evolution Weights (TEW) into the aforementioned sensitive factor tensors to construct dynamic evolution trajectories.

[0124] The specific method is as follows:

[0125] For each sensitivity factor (triplet T[i][j][t]), build a time series index. ;

[0126] Calculate the slope of its numerical change over n time slices, for example, by fitting it with simple linear regression, to obtain the evolution gradient g;

[0127] Introducing the definition of evolution weight: TEW = arctangent(g) × positive change coefficient;

[0128] Using TEW as the expression weight of the current trend of the sensitivity factor, multiply it with the original tensor elements to generate the evolution-weighted tensor T';

[0129] The evolution trajectory of each field pair on the time scale t can be represented by the corresponding vector in T', and connecting them in chronological order constitutes a dynamic feature evolution trajectory sequence.

[0130] This invention introduces an AI Agent to perform perceptual modeling, graphical visualization, deviation clustering, and path deduction on high-dimensional dynamic feature sequences, thereby expanding data quality governance from point anomaly detection to structured anomaly chain identification, and improving the integrity and foresight of the governance chain.

[0131] This invention takes the previously constructed dynamic feature evolution trajectory sequence as input, combines it with a graph neural network model to perform deviation modeling analysis, outputs a multi-level quality risk heat map, and identifies abnormal diffusion structures based on this to construct an abnormal propagation path set.

[0132] Based on graph neural network-based feature bias modeling, this step aims to input the dynamic feature evolution trajectory sequence into AIAgent for deep modeling, output the degree of feature bias, and spatially represent it in the form of a heatmap.

[0133] The input is a three-dimensional tensor structure, with the tensor dimension being: [number of field pairs] × [number of time windows] × [sensitivity factor strength].

[0134] Each tensor element represents the sensitivity factor strength value of a field pair within a certain time window (such as the ratio of the standard deviation of the disturbance to the mean deviation).

[0135] The time window is set to a sliding window with a step size of 1 minute and a window length of 5 minutes, 15 minutes, and 60 minutes.

[0136] A Graph Convolutional Network (GCN) is adopted, and the structure includes 3 convolutional layers and 2 fully connected layers;

[0137] Construct a feature map where nodes are combinations of fields and edges represent semantic similarity between fields (derived from the field association matrix).

[0138] The initial feature of each node is the evolution trajectory vector of this field combination under different time windows;

[0139] The goal of network training is to minimize the mean square error between the predicted bias intensity and the actual intensity of the sensitive factor.

[0140] Curve fitting is performed on the evolution trajectory of each field combination, and the first derivative over a continuous time period is calculated as the trend gradient.

[0141] If at a certain time t, the deviation value increases by more than twice its historical average compared to the previous time window, it is defined as a "mutation point" and the mutation intensity is recorded.

[0142] The initial values ​​of the heatmap are generated by combining the trend gradient and the intensity of abrupt changes as thermal weights.

[0143] Construct a two-dimensional heatmap matrix H, with time on the horizontal axis and field pair numbers on the vertical axis;

[0144] Each element H[i][t] represents the deviation strength of field i at time t;

[0145] Input the two-dimensional matrix into the AI ​​Agent's visualization module (such as Matplotlib or Plotly integrated rendering interface) to generate an interactive initial heatmap.

[0146] The initial heatmap contains numerous areas of deviation with wide spatial and temporal distribution but varying intensities. To accurately identify high-risk clusters, this step employs an adaptive threshold segmentation algorithm to extract hotspots from the heatmap.

[0147] Iterate through all elements in the heatmap H to obtain the set of heat values;

[0148] Calculate the 95th percentile value of the set, denoted as θ (theta);

[0149] θ is used as the global deviation intensity discrimination threshold.

[0150] Using all coordinates of H[i][t]≥θ in H as seed points, perform 4-neighborhood connectivity expansion;

[0151] Adjacent high-intensity points are grouped into a "deviation hotspot region", with the region number incrementing.

[0152] Each hotspot region record includes the combination of fields, the start and end time range, and the average popularity.

[0153] Based on the time span and peak popularity of hotspot areas, they are divided into three levels: low risk (popularity less than θ+10%), medium risk (θ+10% to θ+30%), and high risk (θ+30%).

[0154] The output structure is a set of hotspot regions. Each region contains a set of fields, a time window sequence, and a risk level label.

[0155] After marking hotspot areas, to track the potential sources and propagation logic of anomalies, it is necessary to construct path chains by combining the structural relationships between fields. Path modeling is based on a directed graph structure, where each path represents an anomaly propagation behavior.

[0156] In each high-risk area, the frequency of the starting and ending fields of the statistical field combination;

[0157] The starting field refers to the field that appears earliest in the path sequence and also has the fastest increase in popularity;

[0158] The endpoint field is the field that appears latest, has the highest popularity, and has no further continuation.

[0159] For each pair of field start and end points, find the sequence of field combinations that appear in between;

[0160] Directed edges are established based on the time sequence and the mutual information values ​​between fields, with the edge direction pointing from fields with earlier rising popularity to fields with later rising popularity;

[0161] Each path is denoted as Each hop in the path represents a potential abnormal interference or transmission.

[0162] Only paths with an average heat value higher than θ+20% and a path length of no less than 3 jumps are retained;

[0163] Delete all paths that are discontinuous in time or have too little fluctuation in popularity.

[0164] After obtaining several effective pathways, they need to be organized into a holistic communication structure, and their impact needs to be quantified.

[0165] All valid paths are merged into a propagation graph G, where fields are nodes and path jumps are edges;

[0166] Each side is assigned a weight, which is determined by the average heat value of the corresponding field combination.

[0167] Graph visualization is used for subsequent causal analysis and strategy backtracking.

[0168] Suppose a path p contains l fields for hop count, the average popularity is h, and the maximum hop count is k;

[0169] Path scoring α and β are the adjustment factors, with α=0.7 and β=0.3 by default; the higher the score, the more central and influential the path is in the abnormal propagation.

[0170] All paths are sorted by rating from highest to lowest;

[0171] The top 10% of high-scoring paths are marked as "core propagation paths" and used as input for subsequent AI Agent strategy decisions;

[0172] The final output structure of the path set includes path ID, field node sequence, rating, start and end time, and associated popularity value.

[0173] With the aforementioned set of abnormal propagation paths already constructed, in order to clarify the upstream and downstream sources and impact targets of each abnormal field, this step requires aligning the fields to the actual data link structure.

[0174] The system maintains a field mapping table to record the flow relationship of each field in the data pipeline;

[0175] Each record is formatted as follows: [Field Name] → [Source Node] → [Target Node]. For example, the "Transaction Status" field flows from the "Transaction Details Table" to the "Risk Assessment Model".

[0176] Analyze all field nodes in the anomaly propagation path set, call the field mapping table, and obtain the direct upstream source and downstream target of the field.

[0177] Recursively search for dependencies both upwards and downwards for all abnormal fields;

[0178] Create a set of directed edges at the field level, with the edge direction corresponding to the data flow direction;

[0179] The output structure is a link graph L = (F, E), where F is the set of fields and E is the set of directed dependency edges between fields.

[0180] This link diagram L reflects the propagation path of the abnormal field in the physical data flow layer, and will be integrated with the logical task flow diagram in the future.

[0181] The actual impact of anomaly propagation depends on the frequency and weight of the anomalous field in downstream tasks. This step defines the Potential Impact Factor Weight (PIFW) to quantify the interference of field anomalies on task execution.

[0182] Extract the field call records of each task node in the task flow graph;

[0183] Count the frequency of each field in the task nodes, that is, the number of times a field is referenced in a certain task;

[0184] Construct a task field call matrix M, where rows represent fields, columns represent task nodes, and values ​​represent the frequency of reference of that field in that task.

[0185] TF-IDF weighting method: Let the frequency of field f in task t be TF(f,t); the number of tasks in which field f appears in all tasks be DF(f); and the total number of tasks be N; then the weight of field f in task t is calculated as: weight(f,t) = TF(f,t)×log(N / (DF(f) + 1)); the maximum weight of each field in all tasks is taken as the potential influence factor weight PIFW(f) of that field; all weights are normalized to the range of 0 to 1.

[0186] Ultimately, PIFW(f) represents the importance of field f to the task execution result in the overall task flow. The higher the value, the greater the impact on system stability if an anomaly occurs in this field.

[0187] This step merges the three graph structures to generate a unified causal graph.

[0188] All nodes in the graph structure are uniformly called "field nodes", and each node carries the following attributes:

[0189] Field name; data source or task node; abnormal heat value (from heatmap); PIFW value.

[0190] Each hop edge in the original anomaly propagation path is preserved;

[0191] Add field dependency edges to the data link graph;

[0192] In the task flow graph, if two fields are referenced simultaneously in the same task, a "cooperative edge within the same task" is added.

[0193] All edges are directed and have weights set, with priority as follows: propagation path edges > data link edges > collaborative edges.

[0194] The nodes and edges mentioned above are organized using a directed graph structure to construct a causal origination graph G = (N, E);

[0195] The node set N represents all exception fields;

[0196] Edge set E represents the causal propagation path between fields;

[0197] The graph structure is saved as an adjacency list and visualized.

[0198] This diagram not only illustrates the transmission path of anomalies, but also reflects the functional hierarchy and centrality of each field in the system's propagation structure.

[0199] To achieve proactive governance of data quality, this step identifies key nodes in the causal origination graph and injects corresponding corrective strategies based on their attribute labels.

[0200] Node propagation role analysis: If a node is the starting point in multiple paths, it is marked as a "source node"; if a node is a path intersection point or a node with a high out-degree, it is marked as a "transfer critical node"; if a node is the end point of a path, it is marked as an "affecting end point".

[0201] Based on the overall PIFW value, nodes with a PIFW greater than 0.75 are selected as key candidate nodes;

[0202] The key node set K is selected from the high PIFW nodes that play the role of "source" or "transfer".

[0203] The system's preset correction strategies are as follows:

[0204] Data completion: Used to fill in missing field values, using context-based interpolation;

[0205] Value correction: used to correct numerical abrupt changes, using sliding window mean regression to replace outliers;

[0206] Confidence compression: Reduces the weight of outlier fields in the AI ​​model to avoid misleading results;

[0207] Time delay compensation: Used in scenarios where data arrives late, it synchronizes the data time window through a caching mechanism.

[0208] Strategy Tag Injection:

[0209] Select an appropriate correction strategy based on the anomaly type and path role of the key nodes;

[0210] Each key field node is appended with a "correction strategy label", the label structure of which is as follows:

[0211] Field name: "Transaction status"

[0212] Strategy type: "Value correction"

[0213] Application conditions: "Mutation intensity ≥ 2.0"

[0214] Application method: "Sliding window mean replacement".

[0215] The tags drive the AI ​​Agent to perform data intervention operations in subsequent system cycles.

[0216] In the causal origin graph, each key field node carries its anomalous attributes and the injected corrective strategy label. The AI ​​Agent first needs to identify the applicable strategy types and, combined with the current field state and the historical strategy execution effects, calculate the adaptation priority of each strategy.

[0217] Strategy label extraction: Read the correction strategy labels of the target field nodes from the graph. The labels include strategy type (such as value correction, completion, etc.), applicable conditions (such as mutation intensity ≥2), preset parameters, etc.

[0218] Number the strategies, and let the set of strategies be . Each of them It represents a specific executable strategy.

[0219] Fit factor evaluation: Calculate the fit score for each strategy on the current field. The fit score consists of three parts:

[0220] Match score: The degree of match between the applicable conditions in the strategy label and the abnormal features of the current field (such as popularity, PIFW value);

[0221] Historical performance score: Statistics on the success rate of governance using the same fields and strategies in the past;

[0222] Cost adjustment factor: A penalty factor for the computational resources or time overhead required to execute a strategy.

[0223] Strategy priority matrix construction: Construct a two-dimensional priority matrix M with fields as rows and strategies as columns;

[0224] Element M[i][j] represents the strategy. For fields The strategy adaptation score ranges from 0 to 1.

[0225] This matrix is ​​used for initial value guidance and policy selection inputs in subsequent reinforcement learning models.

[0226] While the policy priority matrix provides a static assessment of fit, it cannot dynamically reflect the actual effect of the policy in the current context. Therefore, a reinforcement learning-based virtual policy simulation evaluation mechanism is introduced to estimate governance benefits by simulating governance intervention results online, thereby improving the accuracy of policy selection.

[0227] Current status It consists of field anomaly feature vectors, including information such as the field's popularity value, PIFW value, historical anomaly rate, and number of strategy executions;

[0228] The state vector is standardized and then used as the policy input.

[0229] Action space construction: Action To select a combination of one or more correction strategies;

[0230] If the total strategy type is n, then the action space size is (Excluding empty actions), each action is a binary encoded vector of the strategy combination.

[0231] Reward function design: After executing a certain strategy combination in a virtual environment, simulate the change in the value of this field after intervention;

[0232] The reward value R consists of the following two parts:

[0233] Reduction in prediction error: The difference in prediction error between the model before and after the intervention;

[0234] Intervention stability: The degree of decrease in the popularity fluctuation of the field within the next 3 time windows after intervention;

[0235] If any strategy leads to a sharp increase in field bias or a larger prediction error, then let R be a negative value.

[0236] Reinforcement learning model training:

[0237] Use reinforcement learning frameworks based on Q-learning or DDPG;

[0238] Each training round updates the policy-state value table Q(S, A);

[0239] The learning rate was set to 0.01, and the discount factor γ was set to 0.95.

[0240] After completing several rounds of simulation, the expected return score for each strategy combination is output.

[0241] The model training process can be pre-deployed and run asynchronously, ensuring that the main process responds online without being blocked.

[0242] The AI ​​Agent selects the most suitable policy combination for the target field based on the policy priority matrix and the value score output by the reinforcement learning model.

[0243] The steps for choosing a strategy combination are as follows:

[0244] Construction of joint scoring function: Define strategy In the field The fitness of the model is A(i,j), and the corresponding score of the reinforcement learning model is V(i,j).

[0245] The combined score is: , where r is the balance coefficient, which is set to 0.6 by default;

[0246] For each field, select the strategy combination with the highest joint score, denoted as . .

[0247] Multi-strategy joint governance rules: If If it is a single strategy, such as "value correction", then only that strategy will be executed; if it is a combination of multiple strategies, such as "data completion + confidence compression", then they will be executed in order of priority.

[0248] If there are conflicts between strategies (such as "complete" and "compress" acting on the same field), the execution order can be coordinated through preset rules or a compatible subset can be selected.

[0249] Policy parameter refinement: Based on the parameter templates provided in the correction label (such as a correction window size of 3 and a confidence compression coefficient of 0.7), the policy is configured as a specific executable task;

[0250] All parameters and execution commands are encapsulated in JSON format and passed to the online intervention module.

[0251] After selecting the strategy combination, the AI ​​Agent performs online governance operations on the target data segment and records the feedback results for subsequent learning.

[0252] If the strategy applies to the original field value (such as value correction), the original value is directly replaced; if it applies to the semantic vector (such as confidence compression), the weights of the outlier dimension are adjusted in the vector space; all changes retain a snapshot of the original value for retrospective purposes.

[0253] Update the latest deviation value of the field node in the semantic distribution graph;

[0254] If the deviation value decreases by more than the set threshold (e.g., 30%), the governance is considered successful, and the strategy is recorded as a high-value strategy.

[0255] After each intervention, record the current strategy combination, the deviation index before and after the intervention, and the change in model prediction error; input the feedback information into the reinforcement learning model to update the strategy value function and improve the quality of subsequent strategy recommendations.

[0256] After completing the governance operations on the target data segment, to further improve the AI ​​Agent's sensitivity and adaptability to changes in data quality, this invention proposes a mechanism for generating data trust labels based on the governance effect. This mechanism supports the quantitative evaluation of the governance effect of each intervened data segment and assigns different levels of trust labels accordingly. These labels are then fed back to the input channel of the AI ​​Agent model, enabling the model to prioritize and dynamically update high-quality data.

[0257] After completing data governance (such as strategies for completion, correction, and confidence compression), it is necessary to evaluate the actual contribution of governance operations to improving data quality. To this end, a Trust Score (TS) metric is defined to reflect the effectiveness of data governance.

[0258] TS is composed of the following three weighted parts:

[0259] Deviation Convergence (CD): Represents the percentage reduction in outlier values ​​of a field after remediation compared to before remediation. Calculation method:

[0260] Deviation convergence = (deviation value before treatment - deviation value after treatment) / deviation value before treatment;

[0261] Value Fluctuation Convergence Ratio (SDR): The ratio of the standard deviation of the field after treatment over the next three time windows to the standard deviation before treatment is used as an indicator to show the degree of improvement in the stability of the field value.

[0262] Value fluctuation convergence rate = 1 - (subsequent standard deviation / standard deviation before treatment);

[0263] Persistence of Intervention Impact (EIS): If the probability of anomalies in a field after intervention remains below a system threshold (e.g., 0.2) for multiple consecutive time windows, the intervention effect is considered stable. EIS can be expressed as the ratio of the number of stable windows to the total number of observation windows.

[0264] Credible scoring function: After normalizing the above three terms to 0-1, a weighted combination is used: Recommended parameter settings: =0.4, γ = 0.3, γ3 = 0.3, supporting dynamic adjustment. The TS value is segmented into three levels: TS ≥ 0.8: marked as "High Trust"; 0.5 ≤ TS < 0.8: marked as "Medium Trust"; TS < 0.5: marked as "Low Trust". The scoring results provide the basis for trust label generation.

[0265] Based on the above calculations, the AI ​​Agent generates a structured, trusted label for each governed data fragment. The label is embedded in the data fragment's metadata in JSON or key-value format.

[0266] The tag field structure is as follows:

[0267] Data Fragment ID: Used to identify the original data fragment, usually consisting of a timestamp, source node, and unique sequence number;

[0268] Governance strategy list: Records the names and parameters of the applied corrective strategies, such as {"strategy": "value correction", "window": 3};

[0269] Trust level: High trust / Medium trust / Low trust;

[0270] Credibility Scoring Details: Displays the scores for each of the above CD, SDR, and EIS categories;

[0271] Generation time: Tag generation timestamp;

[0272] Whether retraining is allowed: Mark whether the data can be used in subsequent model updates.

[0273] The generated tags will be sent to the subsequent data scheduling module along with the data fragments.

[0274] In order to prioritize feeding high-reliability data fragments back to the model input interface of the AI ​​Agent, this implementation design a reliability-driven data feedback mechanism to achieve dynamic scheduling of label-based input data.

[0275] The system divides data fragments into three cache pools:

[0276] High-reliability pool: Only accepts data with TS ≥ 0.8 and has an input priority of 1;

[0277] Medium Trust Pool: Receives data with a frame rate of 0.5 ≤ TS < 0.8, with a priority of 2;

[0278] Low-trust pool: Used only for asynchronous storage, for later review or manual confirmation, with a priority of 3.

[0279] When the AI ​​Agent model updates a batch, it first samples from the high-confidence pool; if the data in the high-confidence pool is insufficient, it supplements the data from the medium-confidence pool proportionally; low-confidence data is only used in the "model stability test" stage.

[0280] All cache pools are emptied every 15 minutes; successfully processed data is removed from the feedback queue to avoid duplicate input.

[0281] This invention further proposes a model input weight adjustment mechanism based on trusted labels to improve the AI ​​Agent's ability to perceive and respond to high-quality data.

[0282] Input layer weight adjustment: In the input processing layer of the AI ​​Agent model, a "trust level weighting module" is introduced; the feature vector of each input data is multiplied by a trust weight coefficient before entering the model;

[0283] High credibility: Weighting coefficient = 1.2;

[0284] Zhongxin: Weighting coefficient = 1.0;

[0285] Low credibility: Weighting coefficient = 0.8;

[0286] Dynamic update strategy: After each model training or inference, the influence coefficient of different levels of input data on the output is calculated; if the error introduced by low confidence data exceeds the set threshold (such as the error increase exceeding 5%), the system will automatically reduce its weight to below 0.6, limiting its participation in the model update frequency.

[0287] All weight parameters support hot updates without interrupting the AI ​​Agent. It also supports fine-tuning of weight coefficients based on actual model accuracy feedback, constructing adaptive input adjustment curves.

[0288] Example 2, please refer to Figure 2As shown in this embodiment, a data quality governance system based on an AI Agent includes:

[0289] Data acquisition module: Acquires structured or semi-structured data streams, uses a preset data semantic encoder to embed semantic vectors into each data segment, and constructs an initial data semantic distribution map;

[0290] Evolutionary trajectory generation module: Based on the initial data semantic distribution map, extract cross-dimensional correlation feature factors, construct a multi-scale quality anomaly sensitivity factor matrix, and embed temporal evolution weights in the feature space to form a dynamic feature evolution trajectory;

[0291] Anomaly propagation path set construction module: The dynamic feature evolution trajectory is used to generate a multi-level quality risk heat map using an AI Agent, and an anomaly propagation path set is constructed based on the deviation-dense region;

[0292] Weight calculation module: Based on the set of abnormal propagation paths, combined with upstream and downstream data links and task flow graphs, calculate the weights of potential influencing factors, construct a causal tracing graph, and inject correction strategy labels into key nodes in the path.

[0293] Intervention module: Based on the causal origin map and correction strategy labels, the AI ​​Agent autonomously selects an appropriate combination of governance strategies to intervene online in the target data segment;

[0294] Update module: Generates trusted labels for the processed data fragments and feeds them back to the AI ​​Agent model input interface, enabling the AI ​​Agent model to prioritize and dynamically update high-quality data.

[0295] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A data quality governance method based on AI Agent, characterized in that: include: Step S1: Obtain structured or semi-structured data streams, and use a preset data semantic encoder to embed semantic vectors into each data segment to construct an initial data semantic distribution map; Step S2: Based on the initial data semantic distribution map, extract cross-dimensional correlation feature factors, construct a multi-scale quality anomaly sensitivity factor matrix, and embed temporal evolution weights in the feature space to form a dynamic feature evolution trajectory; Step S3: Use an AI Agent to generate a multi-level quality risk heat map from the dynamic feature evolution trajectory, and construct an anomaly propagation path set based on the areas with dense deviations; Step S4: Based on the set of abnormal propagation paths, combined with upstream and downstream data links and task flow graphs, calculate the weights of potential influencing factors, construct a causal tracing graph, and inject correction strategy labels into key nodes in the path. Step S5: Based on the causal origination map and the correction strategy label, the AI ​​Agent autonomously selects an appropriate combination of governance strategies to intervene online in the target data segment; Step S6: Generate trusted labels for the processed data fragments and feed them back to the AI ​​Agent model input interface to enable the AI ​​Agent model to prioritize and dynamically update high-quality data. Specifically, this includes: Based on the results after the implementation of the governance strategy, the deviation convergence and stability change rate of key fields before and after governance are calculated for each data segment, and a credibility scoring index is defined. The credibility scoring index consists of the deviation reduction magnitude, the convergence degree of field value fluctuation range, and the persistence of intervention impact. Based on the credibility scoring index, a multi-level credibility label level is set. The scoring results are divided into three levels: "high credibility", "medium credibility" and "low credibility" using a segmented scoring method. The corresponding data segment ID, intervention strategy type, field list and scoring details are recorded in the label. The trusted tags are attached to the metadata structure of the corresponding data fragments, and a hierarchical caching mechanism is used to prioritize writing high-trust data into the AI ​​Agent model input queue, while low-trust data enters the asynchronous processing channel to wait for supplementary verification.

2. The data quality governance method based on AI Agent according to claim 1, characterized in that: The construction of the initial data semantic distribution map includes: By connecting to multiple heterogeneous data source nodes, including but not limited to IoT sensors, API interfaces, databases or log files, structured and semi-structured data can be collected in real time. Data preprocessing is performed on the collected data fragments of various types to generate a data slice sequence with a unified input format; The data slice sequence is input into a preset data semantic encoder, which adopts a multi-layer bidirectional Transformer structure and integrates a context attention mechanism to model the implicit semantic relationships in the data. Using the semantic vector embedding results output by the data semantic encoder, a semantic adjacency graph is constructed in the vector space based on semantic similarity, and an initial data semantic distribution map is generated by combining it with time series labels.

3. The data quality governance method based on AI Agent according to claim 2, characterized in that: The extraction of cross-dimensional correlation feature factors and the construction of a multi-scale quality anomaly sensitivity factor matrix include: In the initial data semantic distribution map, principal component analysis is performed on the semantic vectors of all nodes. The top K principal component vectors with a cumulative explained variance of more than 95% are selected and used as low-dimensional feature basis. On the low-dimensional feature basis, the mutual information values ​​between different field types are calculated respectively, and a cross-dimensional correlation matrix between fields is constructed through standardization. For each pair of fields in the correlation matrix, the change sensitivity under different time windows is calculated. The change sensitivity is the ratio of the standard deviation of the feature perturbation within the moving average window to the baseline difference, thus obtaining a set of sensitive factors under different time scales. The sensitive factor set is organized into a three-dimensional tensor according to field dimension, time scale, and perturbation intensity to construct a multi-scale quality anomaly sensitive factor matrix. Then, an evolution weight is assigned to each sensitive factor using a sliding time window mechanism, and finally, the corresponding dynamic feature evolution trajectory sequence is generated.

4. The data quality governance method based on AI Agent according to claim 3, characterized in that: The step of generating a multi-level quality risk heatmap using an AI Agent based on the dynamic feature evolution trajectory includes: The dynamic feature evolution trajectory sequence is input into the feature deviation perception model built on graph neural network, and the evolution trend gradient and mutation intensity of each field pair in the continuous time window are calculated to generate an initial two-dimensional heat map with time dimension as the horizontal axis, field combination as the vertical axis, and deviation intensity as heat. An adaptive threshold segmentation algorithm is applied to the initial two-dimensional heat map to extract deviation hotspot areas. The threshold is dynamically calculated based on the deviation intensity of the 95th percentile according to the overall heat distribution, and multiple independent sets of high-risk areas are obtained. In each high-risk area, the historical change path of the corresponding field combination in the evolution trajectory is tracked. A directed path graph containing the starting field, the ending field, and the intermediate field chain is constructed by combining the field correlation matrix and the time sequence. Each hop in the path represents a potential anomalous transmission. All constructed anomaly propagation paths are structurally merged to form an anomaly propagation path set with fields as nodes and anomaly relationships as edges. A path scoring function is introduced to prioritize the paths by combining path length, hop count weight, and average popularity.

5. The data quality governance method based on AI Agent according to claim 4, characterized in that: The process of combining upstream and downstream data links and task flow diagrams to calculate the weights of potential influencing factors and construct a causal origination graph includes: The field nodes involved in each path of the anomaly propagation path set are analyzed. Based on the predefined data flow mapping relationship table, the upstream and downstream source and target nodes of each field in the data link are retrieved, and the field-level upstream and downstream dependency links are constructed. By combining the field call records of each processing task in the task flow diagram, the frequency of reference and scope of influence of abnormal fields in each task node are statistically analyzed, and the potential influence factor weight of each field in the global task flow is calculated using the TF-IDF weighting method. Using field nodes as graph nodes and data flow and task dependencies as edges, a unified causal tracing graph is constructed by integrating anomaly propagation path graph, upstream and downstream field link graph, and task field call graph. The causal tracing graph is a directed weighted graph. For key field nodes in the causal origination graph, based on their position in the abnormal path and the weight score of the influencing factor, the corresponding correction strategy label is matched to form a field-level correction label set and injected into the corresponding node.

6. The data quality governance method based on AI Agent according to claim 5, characterized in that: Based on the causal origination map and correction strategy labels, the AI ​​Agent autonomously selects an appropriate combination of governance strategies to intervene online in the target data segment, including: Based on the abnormal propagation weights, strategy label types, and historical correction effect scores of key field nodes in the causal tracing graph, a multi-objective decision model is invoked to construct a strategy adaptation priority matrix, which is used to indicate the matching degree and expected intervention benefits of each correction strategy with its applicable fields. For each candidate correction strategy, a virtual intervention simulation is performed. The predicted output change of the target field after intervention is simulated based on the pre-trained model. The strategy value estimate is updated using the error convergence degree and stability as the reward function. Based on the policy adaptation priority matrix and the policy value of the reinforcement learning output, the optimal combination of governance policies is selected. The governance strategies are combined and applied to the original values ​​or semantic embedding vectors of the target data fragments to perform online intervention operations.

7. A data quality governance system based on AI Agent, used to implement the data quality governance method based on AI Agent as described in any one of claims 1-6, characterized in that: include: Data acquisition module: Acquires structured or semi-structured data streams, uses a preset data semantic encoder to embed semantic vectors into each data segment, and constructs an initial data semantic distribution map; Evolutionary trajectory generation module: Based on the initial data semantic distribution map, extract cross-dimensional correlation feature factors, construct a multi-scale quality anomaly sensitivity factor matrix, and embed temporal evolution weights in the feature space to form a dynamic feature evolution trajectory; Anomaly propagation path set construction module: The dynamic feature evolution trajectory is used to generate a multi-level quality risk heat map using an AI Agent, and an anomaly propagation path set is constructed based on the deviation-dense region; Weight calculation module: Based on the set of abnormal propagation paths, combined with upstream and downstream data links and task flow graphs, calculate the weights of potential influencing factors, construct a causal tracing graph, and inject correction strategy labels into key nodes in the path. Intervention module: Based on the causal origin map and correction strategy labels, the AI ​​Agent autonomously selects an appropriate combination of governance strategies to intervene online in the target data segment; Update module: Generates trusted labels for the treated data fragments and feeds them back to the AI ​​Agent model input interface, enabling the AI ​​Agent model to prioritize and dynamically update high-quality data.

Images