Data maintenance management method and system based on artificial intelligence

Abstract

The invention relates to the technical field of electronic data digital processing, and discloses a data maintenance management method and system based on artificial intelligence. The method comprises the following steps: acquiring multi-dimensional features such as an access time sequence, metadata, an association relationship and a service context of data; constructing a hybrid model fusing a long-short-term memory network and a graph attention network, and predicting a future value score of the data; dynamically generating a hierarchical storage, archiving or deleting strategy based on the score and storage cost-performance model; and through the access feedback after the execution of the monitoring strategy, the prediction model is optimized in a closed-loop manner by using a reinforcement learning mechanism. The system comprises a data feature acquisition module, a value prediction module, a strategy generation module and an execution feedback module. According to the method, self-adaptive, refined and intelligent management of the full life cycle of the data is realized, the utilization efficiency of storage resources is remarkably improved, and the operation and maintenance cost is reduced.

Description

Technical Field

[0001] This invention belongs to the field of electronic data digital processing technology, specifically relating to a data maintenance and management method and system based on artificial intelligence. Background Technology

[0002] As digital transformation deepens, data has become a core asset for enterprises, and the importance of its full lifecycle management is increasingly prominent. Traditional data maintenance strategies generally rely on static rules preset by administrators, such as performing archiving, deletion, or degradation operations at fixed time intervals. Such methods assume that the value of data decays linearly or stepwise over time, ignoring the actual impact of data content semantics, dynamic usage behavior, and business context on value evolution.

[0003] In complex and ever-changing business environments, the rate of value decay varies significantly among different data items: some critical business logs may quickly become invalid after being accessed frequently in a short period, while certain historical contract texts, though inaccessible for a long time, still possess high value due to compliance or audit requirements. Fixed strategies cannot capture this non-uniform and non-linear pattern of value change, leading to the premature cleanup of high-value data or the prolonged occupation of storage resources by low-value data, resulting in both wasted storage costs and data availability risks.

[0004] AI-based data maintenance and management aims to achieve accurate prediction and adaptive control of data value through dynamic modeling. The core of this approach lies in integrating multi-dimensional information sources to construct a comprehensive evaluation mechanism that reflects the inherent semantics of data, its external usage patterns, and its business relevance. An ideal data value prediction model should understand the semantic importance of data content, identify the temporal evolution trends of behavioral characteristics such as access frequency, user roles, and query intent, and combine this with metadata such as the data's business domain, compliance level, and associated entities to form a quantitative prediction of the value and risk of a single piece of data or dataset at a specific future point in time.

[0005] While some existing technologies have attempted to incorporate machine learning for data classification or popularity prediction, they generally suffer from problems such as limited model dimensions, lack of semantic understanding, and insufficient integration with business metadata. Most solutions rely solely on access logs to build simple time-series models, neglecting the semantic value of the data content itself; others use keyword matching or rule engines for content analysis, which struggles to generalize to unfamiliar data types. More critically, existing methods fail to organically integrate content semantics, access patterns, and business context, leading to a disconnect between prediction results and actual business value. In scenarios with extremely high data sensitivity and compliance requirements, such as finance, healthcare, and government, this coarse-grained, static management approach is highly susceptible to serious consequences such as data breaches, compliance violations, or loss of critical information. Therefore, there is an urgent need for an intelligent data maintenance and management method that deeply integrates natural language processing, temporal behavior modeling, and business metadata reasoning to achieve dynamic perception, accurate prediction, and adaptive maintenance of data value. Summary of the Invention

[0006] To address the technical problem that existing data lifecycle management relies on preset, fixed time strategies, failing to accurately reflect the actual decay curve of data value, leading to unreasonable storage resource allocation and high management costs, this invention provides an artificial intelligence-based data maintenance and management method and system. The core technical solution of this invention lies in: constructing a multi-dimensional dynamic data feature acquisition system, comprehensively considering data access timing characteristics, content metadata, inter-entity relationships, and business context information; utilizing a deep learning hybrid model, specifically combining a long short-term memory network for capturing timing dependencies with a graph attention network for resolving complex inter-entity relationships, to accurately predict the intrinsic value of data in future time windows and generate a quantified data value score; based on this data value score, combined with a preset storage cost and access performance model, dynamically generating and executing optimal data tiered storage, archiving, compression, or deletion strategies; finally, by monitoring data access behavior after strategy execution, a closed-loop feedback mechanism is constructed to quantify the strategy execution effect into reward and penalty signals, continuously optimizing the data value prediction model, thereby achieving adaptive, intelligent, and refined management of the entire data lifecycle.

[0007] This invention provides a data maintenance and management method based on artificial intelligence, which includes the following steps: The characteristics of multi-source heterogeneous data in the data storage system are obtained, including data access log characteristics, data metadata characteristics, data association characteristics, and business context characteristics. Based on the characteristics of the multi-source heterogeneous data, a data value prediction model is constructed. The data value prediction model is used to predict the data value score of the target data object in a future preset time period. Based on the data value score, and combined with the preset storage resource cost model and data access performance model, a data maintenance strategy corresponding to the data value score is generated. The data maintenance strategy is executed, and the data access feedback events after the strategy is executed are monitored. The data access feedback events are then used to iteratively optimize the data value prediction model.

[0008] Furthermore, the acquisition of multi-source heterogeneous data features specifically includes: By deploying log interceptors at the underlying input / output interfaces of the data storage system, access events of data objects are captured in real time, forming a data access time sequence containing timestamps, data object identifiers, access types, access source identifiers, and executed operation identifiers, which serves as the characteristics of the data access logs. By parsing the internal structure of a data object or its accompanying description file, the creation time, modification time, data format, data size, and data owner information of the data object are extracted as data metadata features. By analyzing foreign key constraints in database management systems, object reference relationships in applications, or the association between fact tables and dimension tables in data warehouses, a weighted directed graph with data objects as nodes and references or dependencies as edges is constructed as the data association relationship feature. By calling external business systems through application programming interfaces (APIs), business entity information associated with data objects can be obtained. This business entity information includes customer level, order amount, or project priority, which serves as the business context feature.

[0009] Furthermore, the construction of the data value prediction model specifically includes: The data access time sequence of each data object is input into the Long Short-Term Memory network. Through the gated recurrent unit structure of the Long Short-Term Memory network, the time dependency of the access event is encoded, the dynamic access pattern of the data object is extracted, and the first time-series feature vector is generated. The weighted directed graph defined by the data association features is input into the graph attention network. Different attention weights are assigned to the neighboring nodes of each node through a multi-head attention mechanism. The neighboring node information is aggregated to generate a second association feature vector that represents the structural importance of the data object. The data metadata features and the business context features are one-hot encoded or embedded encoded to convert them into a numerical third static feature vector. The first temporal feature vector, the second associated feature vector, and the third static feature vector are concatenated and input into a fully connected neural network. The output layer of the fully connected neural network uses a linear activation function and finally outputs a continuous value between 0 and 1 as the data value score.

[0010] Furthermore, the generated data maintenance strategy specifically includes: Multiple storage tiers are preset, each corresponding to a storage medium, and unit data storage cost and standard access latency are defined for each storage tier; Define multiple data value scoring thresholds, divide the continuous value range of data value scores into multiple non-overlapping intervals, and uniquely map each interval to a storage level; For a target data object, compare its current storage level with the target storage level mapped according to its data value score; If the target storage level is different from the current storage level, a data migration instruction is generated. The data migration instruction includes a data object identifier, a source storage level identifier, and a target storage level identifier. If the value score of a data object is lower than the preset minimum archiving threshold, a data archiving instruction will be generated. If the value score of a data object is lower than the preset minimum survival threshold, a data deletion instruction is generated.

[0011] Furthermore, the step of executing the data maintenance strategy and monitoring the data access feedback events after the strategy execution, and using the data access feedback events to iteratively optimize the data value prediction model specifically includes: The generated data maintenance strategy instructions are submitted to the strategy execution engine, which calls the application programming interface of the underlying storage system to complete the data migration, archiving or deletion operations. Continuously monitor data access requests. When a data object that has been migrated to a low-performance storage tier is accessed frequently within a preset time window, or when its access latency exceeds the upper limit specified in the service level agreement, it is judged as a negative feedback event. When a data object that has not been accessed for a long time and resides in the high-performance storage tier has a data value score that is consistently higher than a preset threshold, it is judged as a potential misjudgment event and also as a negative feedback event. The data objects and their feature vectors associated with the negative feedback events are appended with negative reward values ​​to form reinforcement learning training samples. Conversely, when the overall storage cost of the system decreases and the average access latency does not increase significantly after the strategy is implemented, positive reward values ​​are added to all data objects that are correctly processed within this period to form positive reinforcement learning training samples. The network parameters of the data value prediction model are fine-tuned and updated periodically using the accumulated reinforcement learning training samples through a proximal policy optimization algorithm.

[0012] This invention provides a data maintenance and management system based on artificial intelligence, comprising: The data feature acquisition module is used to acquire multi-source heterogeneous data features in the data storage system. The multi-source heterogeneous data features include data access log features, data metadata features, data association relationship features, and business context features. The data value prediction module is used to construct a data value prediction model based on the characteristics of the multi-source heterogeneous data. The data value prediction model is used to predict the data value score of the target data object within a future preset time period. The maintenance strategy generation module is used to generate a data maintenance strategy corresponding to the data value score based on the data value score and in combination with a preset storage resource cost model and data access performance model. The strategy execution and feedback module is used to execute the data maintenance strategy, monitor the data access feedback events after the strategy execution, and use the data access feedback events to iteratively optimize the data value prediction model.

[0013] In one embodiment of the present invention, the data feature acquisition module integrates a log interceptor, a metadata parser, a graph construction engine, and a business interface adapter. The log interceptor captures all read and write requests to data blocks non-intrusively using kernel probes or file system hooks. The graph construction engine periodically scans the database schema definition language file and application configuration files, automatically discovering and updating the relationships between data objects.

[0014] In one embodiment of the present invention, the data value prediction module adopts a hybrid neural network architecture. The Long Short-Term Memory (LSTM) network comprises two stacked LSM units with a hidden layer dimension of 256. The graph attention network comprises three graph attention layers, each employing eight attention heads, with an output node embedding dimension of 128. The fully connected neural network comprises three hidden layers with 512, 256, and 64 neurons respectively, using a modified linear unit (MRU) as the hidden layer activation function.

[0015] In one embodiment of the present invention, the maintenance strategy generation module embeds a cost-benefit analysis unit. Before generating a data migration instruction, the cost-benefit analysis unit calculates the computing resources required to perform the migration operation, the network bandwidth consumption, and the instantaneous performance impact on the service. Only when the expected long-term storage cost savings exceed the immediate cost of performing the migration operation is the data migration instruction finally generated and issued, thereby avoiding system instability caused by frequent, small-scale data migrations.

[0016] In one embodiment of the present invention, the policy execution and feedback module includes a reinforcement learning agent. The agent treats the output of the data value prediction model as a policy network, quantifies negative feedback events as negative rewards with an absolute value of 1, and quantifies positive feedback events as positive rewards between 0 and 1. The agent employs an experience replay mechanism, storing training samples in a fixed-capacity buffer. During each model update, a small batch of data is randomly sampled from the buffer for training, thereby breaking the temporal correlation of data samples and improving the convergence stability of the model.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By integrating four types of features—access time sequence, content metadata, relationship, and business context—a comprehensive and three-dimensional understanding of data value is constructed. Compared with single-dimensional strategies that rely solely on access time or last modification time, the value assessment of this invention is more accurate and can accurately identify data that is accessed infrequently but has extremely high business value, or temporary data that is accessed frequently but has extremely low value.

[0018] 2. The deep learning model, which combines long short-term memory networks and graph attention networks, can simultaneously capture the dynamic evolution trend of data value and the static structural importance in complex data ecosystems. This solves the technical problem that traditional statistical models cannot handle high-dimensional, nonlinear, and strongly correlated data, and significantly improves the accuracy of value prediction.

[0019] 3. A strategy generation mechanism based on cost-benefit analysis was introduced. When making decisions on data migration, not only is the value of the data itself considered, but also the operational costs and performance impact are quantified. This ensures that every data maintenance operation is the globally optimal solution, effectively avoiding system performance fluctuations and resource waste caused by frequent migrations without regard to cost.

[0020] 4. A closed-loop feedback system from strategy execution to model optimization was constructed. Through a reinforcement learning mechanism, the system possesses the ability to self-evolve and continuously adapt. The system can learn from its own erroneous decisions and dynamically adjust the value assessment model to adapt to constantly changing business needs and data usage patterns, achieving full automation and intelligence in data maintenance and management. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of a data maintenance and management method and system based on artificial intelligence proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the data value prediction model that integrates long short-term memory network and graph attention network in this invention; Figure 3 This is a logical flowchart of the feature acquisition and fusion of multi-source heterogeneous data in this invention; Figure 4 This is a flowchart illustrating the logical process of generating dynamic maintenance strategies based on data value scoring and cost-benefit analysis in this invention. Figure 5 This is a schematic diagram of the reinforcement learning optimization framework for the policy execution and closed-loop feedback mechanism in this invention; Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the terminal storage system and the intelligent management module in this invention. Detailed Implementation

[0022] Please refer to Figures 1 to 6 This invention provides a data maintenance and management method and system based on artificial intelligence, aiming to solve the technical problems in existing technologies where data lifecycle management relies on preset, fixed time strategies, which cannot accurately reflect the actual decay curve of data value, leading to unreasonable allocation of storage resources and high management costs. The specific embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0023] The AI-based data maintenance and management method includes the following steps: acquiring multi-source heterogeneous data characteristics from the data storage system; constructing a data value prediction model based on the multi-source heterogeneous data characteristics; generating a corresponding data maintenance strategy based on the data value score output by the data value prediction model, combined with a preset storage resource cost model and a data access performance model; executing the data maintenance strategy, monitoring data access feedback events after strategy execution, and iteratively optimizing the data value prediction model using the data access feedback events.

[0024] In the step of acquiring the multi-source heterogeneous data characteristics in the data storage system, the multi-source heterogeneous data characteristics include data access log characteristics, data metadata characteristics, data association characteristics, and business context characteristics. Specifically, by deploying a log interceptor at the underlying input / output interface of the data storage system, access events of data objects are captured in real time, forming a data access time sequence containing timestamps, data object identifiers, access types, access source identifiers, and executed operation identifiers, which serves as the data access log characteristics. The log interceptor is implemented using kernel probes or file system hook technology, enabling it to non-intrusively intercept all read and write requests to data blocks without modifying application logic, and record the original access events in a unified format to a high-performance log buffer. This log buffer adopts a circular queue structure, possessing high throughput and low latency characteristics, ensuring that it will not become a system bottleneck under high-concurrency access scenarios.

[0025] Simultaneously, by parsing the internal structure of the data object or its accompanying description file, the creation time, modification time, data format, data size, and data owner information of the data object are extracted as the data metadata features. The metadata parsing process is executed by a metadata parser, which supports various mainstream file formats and database table structures, and can automatically identify and extract key metadata fields. For unstructured data, such as text, image, or audio / video files, the metadata parser calls the corresponding format parsing library to extract inherent attributes such as file header information, encoding parameters, and resolution. For structured data, such as relational database tables or key-value pair storage, the metadata parser directly queries the system directory table or schema definition language file to obtain field definitions, index information, and constraints.

[0026] Furthermore, by analyzing foreign key constraints in the database management system, object references in the application, or the association between fact tables and dimension tables in the data warehouse, a weighted directed graph is constructed, with data objects as nodes and references or dependencies as edges, serving as the data association features. This process is completed by a graph construction engine. The graph construction engine periodically scans the database schema definition language file, application configuration files, and dependency declarations in the service registry to automatically discover logical dependencies and physical references between data objects. The weight of each edge is dynamically calculated based on reference frequency, transaction coupling, or business importance. For example, if two data tables are jointly queried in multiple transactions, the weight of their associated edges is assigned a higher value. The constructed weighted directed graph is stored in an in-memory graph database in the form of an adjacency matrix or sparse tensor for subsequent graph neural network processing.

[0027] Furthermore, external business systems are invoked via application programming interfaces (APIs) to obtain business entity information associated with the data object. This business entity information includes customer level, order amount, or project priority, serving as the business context features. This process is implemented by a business interface adapter. The business interface adapter encapsulates standard communication protocols with enterprise resource planning (ERP) systems, customer relationship management (CRM) systems, and project management platforms. It can reverse-engineer the business entity to which the data object belongs based on its identifier and retrieve the corresponding context attributes. All obtained business context information undergoes anonymization and standardization to ensure compliance with data security requirements and is converted into a unified numerical or categorical encoding format for subsequent model processing.

[0028] In the step of constructing the data value prediction model, the data value prediction model is used to predict the data value score of the target data object within a preset future time period. This model adopts a hybrid neural network architecture, integrating Long Short-Term Memory (LSTM) networks with temporal modeling capabilities, graph attention networks with structure awareness capabilities, and fully connected networks with static feature integration capabilities. Specifically, the data access time sequence of each data object is input into the LTM network. Through the gated recurrent unit structure of the LTM network, the temporal dependence of access events is encoded, the dynamic access pattern of the data object is extracted, and a first temporal feature vector is generated. The LTM network contains two stacked LTM units with a hidden layer dimension of 256. The input sequence is first mapped to a dense vector through an embedding layer, and then fed into the first LTM unit step by step, with its output serving as the input to the second layer. Finally, the hidden state at the last time step is taken as the temporal representation of the data object.

[0029] Simultaneously, the weighted directed graph defined by the data association features is input into the graph attention network. A multi-head attention mechanism is used to assign different attention weights to the neighboring nodes of each node, aggregating neighboring node information to generate a second association feature vector representing the structural importance of the data object. The graph attention network consists of three graph attention layers, each using eight attention heads, with an output node embedding dimension of 128. In each layer, each node calculates its attention coefficient based on its own features and the features of its neighboring nodes, using the following formula: ; in, For nodes The current embedding, The weight matrix is ​​a learnable matrix. For the attention mechanism parameter vector, Represents a node The neighborhood group, This represents a vector concatenation operation. The aggregated node representation is then passed to the next layer after nonlinear activation, and the final output node embedding is the second associated feature vector.

[0030] Furthermore, the data metadata features and the business context features are converted into a numerical third static feature vector through one-hot encoding or embedding encoding. For discrete features such as data format and owner / department, one-hot encoding is used; for continuous features such as data size and order amount, normalization is performed to the 0-1 range; for high-cardinality categorical features such as customer identifiers, they are mapped to a low-dimensional dense vector through an embedding layer. All static features are concatenated to form a fixed-length third static feature vector.

[0031] Finally, the first temporal feature vector, the second associated feature vector, and the third static feature vector are concatenated and input into a fully connected neural network. The output layer of the fully connected neural network uses a linear activation function, ultimately outputting a continuous value between 0 and 1 as the data value score. The fully connected neural network contains three hidden layers with 512, 256, and 64 neurons respectively, using a modified linear unit as the hidden layer activation function. This network is trained end-to-end to learn how to fuse three types of heterogeneous features to accurately predict the comprehensive value of data within future time windows.

[0032] In the step of generating a data maintenance strategy, a data maintenance strategy corresponding to the data value score is generated based on the data value score and a preset storage resource cost model and data access performance model. Specifically, multiple storage tiers are preset, each corresponding to a storage medium, and unit data storage cost and standard access latency are defined for each storage tier. Typical storage tiers include: the first tier is a caching tier based on dynamic random access memory, with the highest unit cost but the lowest access latency; the second tier is a high-performance tier based on solid-state drives (SSDs); the third tier is a standard tier based on hard disk drives (HDDs); and the fourth tier is a cold storage tier based on tape or cloud archiving services. The cost and latency parameters for each tier are preset by the system administrator according to the actual hardware configuration and service level agreement.

[0033] Multiple data value scoring thresholds are defined, dividing the continuous range of data value scores into multiple non-overlapping intervals, each interval uniquely mapping to a storage level. For example, a score greater than or equal to 0.8 maps to the first level, 0.6 to 0.79 to the second level, 0.3 to 0.59 to the third level, and scores below 0.3 are archived or deleted. For a target data object, its current storage level is compared with the target storage level mapped according to its data value score. If the target storage level is different from the current storage level, a data migration instruction is generated, which includes a data object identifier, a source storage level identifier, and a target storage level identifier.

[0034] Before generating migration instructions, the cost-benefit analysis unit embedded in the maintenance strategy generation module calculates the computing resources, network bandwidth consumption, and instantaneous performance impact on the business required to perform the migration operation. Only when the expected long-term storage cost savings exceed the immediate cost of performing the migration operation will the data migration instruction be generated and issued. This cost-benefit analysis is based on the following formula: ; in, and These are the unit storage costs for the source and target levels, respectively. For the size of the data object, For the expected length of stay, Cost per unit bandwidth For transmission time, Cost per unit performance disturbance Duration of performance impact on business operations. Only when... When the value is greater than 0, the migration operation is considered economically feasible.

[0035] If the value score of a data object is lower than the preset minimum archiving threshold (e.g., 0.1), a data archiving instruction is generated to compress the data and move it to the cold storage layer; if the value score of a data object is lower than the preset minimum survival threshold (e.g., 0.05) and meets security conditions such as no related dependencies and no business binding, a data deletion instruction is generated.

[0036] In the steps of executing the data maintenance strategy and monitoring data access feedback events after strategy execution, and using the data access feedback events to iteratively optimize the data value prediction model, the generated data maintenance strategy instructions are first submitted to the strategy execution engine. The strategy execution engine then calls the application programming interface of the underlying storage system to complete data migration, archiving, or deletion operations. After the operation is completed, the system continuously monitors data access requests and identifies the strategy execution effect.

[0037] If a data object that has been migrated to a low-performance storage tier is accessed frequently (e.g., accesses exceeding a threshold of 10 times) within a preset time window (e.g., 24 hours), or if its access latency exceeds the upper limit specified in the service level agreement (e.g., 500 milliseconds), it is considered a negative feedback event. Similarly, if a data object that has not been accessed for a long time and resides in the high-performance storage tier has a data value score that is consistently higher than a preset threshold (e.g., 0.7) for multiple evaluation cycles, it is considered a potential misjudgment event and is also considered a negative feedback event.

[0038] The data objects associated with the negative feedback events and their feature vectors are appended with negative reward values ​​(absolute value of 1) to form reinforcement learning training samples; conversely, when the overall storage cost of the system decreases and the average access latency does not increase significantly (the change is less than 5%) after the policy is executed, positive reward values ​​(between 0 and 1, linearly scaled according to the cost saving ratio) are appended to all data objects that are correctly processed in this period to form positive reinforcement learning training samples.

[0039] The policy execution and feedback module includes a reinforcement learning agent. This agent treats the output of the data value prediction model as the policy network and uses a proximal policy optimization algorithm to fine-tune and update the network parameters. The agent employs an experience replay mechanism, storing training samples in a fixed-capacity buffer. During each model update, a small batch of data is randomly sampled from the buffer for training, thereby breaking the temporal correlation of data samples and improving the model's convergence stability. The model update cycle can be set to once daily or once weekly, dynamically adjusted based on system load and data change frequency.

[0040] The AI-based data maintenance and management system includes a data feature acquisition module, a data value prediction module, a maintenance strategy generation module, and a strategy execution and feedback module. The data feature acquisition module integrates a log interceptor, a metadata parser, a graph construction engine, and a business interface adapter, responsible for the collection and preprocessing of four types of features, respectively. The data value prediction module implements the aforementioned hybrid neural network architecture to generate data value scores. The maintenance strategy generation module embeds a cost-benefit analysis unit to ensure the economic rationality of the strategies. The strategy execution and feedback module includes a strategy execution engine and a reinforcement learning agent to achieve closed-loop optimization.

[0041] Through the above methods and systems, this invention achieves a comprehensive, dynamic, and accurate assessment of data value, and on this basis generates an economical, efficient, and adaptive data maintenance strategy, effectively solving the problems of resource waste and management rigidity caused by traditional fixed strategies, and significantly improving the overall performance and intelligence level of the storage system.

Claims

1. An artificial intelligence-based data maintenance management method, characterized by, The method comprises the following steps: acquiring multi-source heterogeneous data features in a data storage system, the multi-source heterogeneous data features comprising data access log features, data metadata features, data correlation relationship features, and business context features; constructing a data value prediction model based on the multi-source heterogeneous data features, the data value prediction model being used to predict a data value score of a target data object in a future preset time period; generating a data maintenance strategy corresponding to the data value score according to the data value score, in combination with a preset storage resource cost model and a data access performance model; executing the data maintenance strategy and monitoring data access feedback events after the execution of the strategy, and iteratively optimizing the data value prediction model using the data access feedback events. 2.The artificial intelligence-based data maintenance management method of claim 1, wherein The method of acquiring multi-source heterogeneous data features in a data storage system comprises the following steps: capturing access events of data objects in real time through a log interceptor deployed at a bottom input / output interface of the data storage system, to form a data access time sequence sequence comprising a timestamp, a data object identifier, an access type, an access source identifier, and an execution operation identifier, as the data access log features; extracting creation time, modification time, data format, data size, and data owner information of the data objects by analyzing internal structures of the data objects or accompanying description files thereof, as the data metadata features; constructing a weighted directed graph with data objects as nodes and reference or dependency relationships as edges by analyzing foreign key constraints in a database management system, object reference relationships in an application program, or fact table and dimension table correlations in a data warehouse, as the data correlation relationship features; obtaining business entity information associated with the data objects by calling external business systems through application program interfaces, the business entity information comprising customer levels, order amounts, or project priorities, as the business context features. 3.The AI-based data maintenance management method of claim 2, wherein, The method of constructing a data value prediction model comprises the following steps: inputting the data access time sequence sequence of each data object into a long short-term memory network, encoding the time dependence of the access events through a gated recurrent unit structure of the long short-term memory network, extracting dynamic access patterns of the data objects, and generating a first time sequence feature vector; inputting the weighted directed graph defined by the data correlation relationship features into a graph attention network, assigning different attention weights to neighbor nodes of each node through a multi-head attention mechanism, aggregating neighbor node information, and generating a second correlation feature vector representing the importance of the structure of the data objects; one-hot encoding or embedding encoding the data metadata features and the business context features, and converting them into a numerical third static feature vector; concatenating the first time sequence feature vector, the second correlation feature vector, and the third static feature vector, and inputting them into a fully connected neural network, the output layer of the fully connected neural network adopting a linear activation function, and finally outputting a continuous value between 0 and 1 as the data value score. 4.The AI-based data maintenance management method of claim 3, wherein, The method of inputting the data access time sequence sequence of each data object into a long short-term memory network to generate a first time sequence feature vector comprises the following steps: After the data access time sequence is mapped to a dense vector by an embedding layer, it is input into a network structure containing two layers of stacked long short-term memory units, with a hidden layer dimension of 256; The hidden state of the last time step is taken as the first time sequence feature vector. 5.The artificial intelligence-based data maintenance management method of claim 4, wherein, Inputting the weighted directed graph into the graph attention network to generate a second correlation feature vector, comprising: in each layer of the graph attention layer, calculating an attention coefficient according to the node itself feature and the neighbor node feature, and the formula is: ; wherein, is the current embedding of the node , is a learnable weight matrix, is an attention mechanism parameter vector, denotes the neighbor set of the node , denotes a vector concatenation operation, the aggregated node representation is passed to the next layer after a non-linear activation, and the final output node embedding is the second association feature vector. 6.The artificial intelligence-based data maintenance management method of claim 5, wherein, The data metadata features and the business context features are encoded to generate a third static feature vector, including: Discrete features are one-hot encoded, continuous features are normalized to the interval [0, 1], and high-base classification features are mapped to low-dimensional dense vectors by an embedding layer; All encoded features are concatenated into a fixed-length vector as the third static feature vector. 7.The AI-based data maintenance management method of claim 6, wherein, A data maintenance strategy is generated, including: Multiple storage tiers are preset, each corresponding to a storage medium, and the unit data storage cost and standard access delay of each storage tier are defined; Multiple data value score thresholds are defined to divide the continuous value domain of the data value score into multiple non-overlapping intervals, and each interval uniquely maps to a storage tier; For a target data object, compare its current storage tier with the target storage tier mapped according to its data value score; If the target storage tier is different from the current storage tier, generate a data migration instruction, which includes the data object identifier, the source storage tier identifier, and the target storage tier identifier; If the data object's value score is below the preset minimum archiving threshold, generate a data archiving instruction; If the data object's value score is below the preset minimum survival threshold, generate a data deletion instruction. 8.The AI-based data maintenance management method of claim 7, wherein, Before generating the data migration instruction, it also includes: Calculate the computing resources, network bandwidth consumption, and instantaneous performance impact on the business required to perform the migration operation; Only when the expected long-term storage cost savings exceed the immediate cost of performing the migration operation, generate and issue the data migration instruction. 9.The AI-based data maintenance management method of claim 8, wherein, Execute the data maintenance strategy and monitor the data access feedback events after strategy execution, and use the data access feedback events to iteratively optimize the data value prediction model, including: Submit the generated data maintenance strategy instructions to the strategy execution engine to call the application programming interface of the underlying storage system to complete the migration, archiving, or deletion of data; When a data object that has been migrated to a low-performance storage tier is accessed frequently within a preset time window, or its access delay exceeds the upper limit specified in the service level agreement, it is determined to be a negative feedback event; When a data object that has not been accessed for a long time and resides in a high-performance storage tier has a data value score that is consistently above a preset threshold for multiple evaluation periods, it is determined to be a negative feedback event; Attach a negative reward value to the data object and its feature vector associated with the negative feedback event to form a reinforcement learning training sample; When the overall storage cost of the system after strategy execution decreases and the average access delay does not significantly increase, attach a positive reward value to the data objects that are correctly disposed of to form a positive reinforcement learning training sample; Periodically use the accumulated reinforcement learning training samples to fine-tune the network parameters of the data value prediction model using the proximal policy optimization algorithm.

10. An artificial intelligence-based data maintenance management system, characterized by, ​ The data feature acquisition module is configured to acquire multi-source heterogeneous data features in the data storage system, wherein the multi-source heterogeneous data features include data access log features, data metadata features, data correlation relationship features, and business context features. The data value prediction module is configured to construct a data value prediction model based on the multi-source heterogeneous data features, wherein the data value prediction model is configured to predict a data value score of a target data object in a future preset time period. The maintenance strategy generation module is configured to generate a data maintenance strategy corresponding to the data value score according to the data value score, in combination with a preset storage resource cost model and a data access performance model. The strategy execution and feedback module is configured to execute the data maintenance strategy, monitor data access feedback events after the execution of the strategy, and iteratively optimize the data value prediction model based on the data access feedback events.

Images