An ai-driven edge node storage intelligent optimization method and system

Abstract

The application relates to the technical field of storage intelligent optimization, and provides an AI-driven edge node storage intelligent optimization method and system. The application is oriented to a distributed edge storage scene of a professional AI dialogue, based on historical access records and customer professional field labels of each node, time period access frequencies are counted and field time sequence feature sequences are constructed, so that the field access rules of different regional nodes are anchored. A time sequence convolution network carrying a field attention branch is used to train a hot data prediction model, high correlation field features are strengthened through attention weighting to adapt to node access emphasis, time sequence dependence is captured through dilated convolution and residual connection to improve prediction accuracy, and the hotness ranking of each field and the corresponding data block in the future is output. According to the hotness ranking, the node is matched with the exclusive field attribute, and cold and hot data are scheduled to the corresponding storage nodes in a hierarchical manner, so that the utilization rate of high-speed storage is improved, the cross-node access delay is reduced, and the storage load of each edge node is balanced.

Description

Technical Field

[0001] This invention relates to the field of intelligent storage optimization technology, and more specifically, to an AI-driven intelligent storage optimization method and system for edge nodes. Background Technology

[0002] The content in this section only provides background information related to this invention and may not constitute prior art.

[0003] With the deep integration of edge computing and vertical industry scenarios, collaborative storage technology for distributed edge nodes has become a core solution for reducing core network transmission pressure and improving content response efficiency. To balance storage scheduling efficiency with data transmission security between distributed nodes, edge storage architectures combining blockchain technology and content popularity prediction algorithms are gradually becoming a research hotspot. This architecture can optimize storage resource allocation by predicting trending content, while ensuring trusted collaboration between nodes in untrusted environments.

[0004] In existing technologies, Chinese patent CN116915781A discloses a blockchain-based edge collaborative caching system and method. This system generates content popularity distribution through a time-series prediction model, constructs a collaborative caching domain between adjacent nodes by combining the cache space status of each edge node, and configures multiple smart contracts to achieve full-process identity and content verification for data transmission, thus improving the system's cache hit rate and security performance to some extent. However, this solution has significant adaptability deficiencies in vertical application scenarios such as AI dialogue models, which target different professional fields and have high specialization. In such scenarios, different edge nodes are responsible for covering customer groups in different regions. Due to differences in regional industrial development directions, the access needs of customers in different regions are significantly differentiated in their professional fields, and the content popularity evolution patterns of different vertical fields are different. The popularity prediction in this solution is based only on general request content and access frequency characteristics, without incorporating the professional field attributes of customers for differentiated feature modeling, and therefore cannot accurately capture the popularity change trends specific to vertical fields. This results in high-volume, domain-specific content failing to be efficiently stored in the corresponding attribute nodes, which not only fails to meet the high-concurrency, dedicated access needs of professional scenarios but also wastes distributed storage resources. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-driven intelligent optimization method and system for edge node storage to improve the aforementioned problems. To achieve this objective, the technical solution adopted by this invention is as follows:

[0006] Firstly, this application provides an AI-driven intelligent optimization method for edge node storage, including:

[0007] S101, obtain the historical access record set of each distributed edge node, the professional field classification label of the corresponding customer, and the storage capacity threshold of the edge node; the historical access record set includes the access time of each data block and the initiating customer identifier, and the professional field classification label is used to mark the professional direction of each customer;

[0008] S102, based on the historical access record set and professional field classification labels, count the access frequency of the corresponding data blocks of each professional field in each time period, and construct the corresponding field time series feature sequence;

[0009] S103, input the domain temporal feature sequence into the temporal convolutional network with domain attention branch for iterative training to obtain the hot data prediction model;

[0010] The training set consists of data from the historical access record set divided into preceding continuous time periods, and the test set consists of data from subsequent continuous time periods. Each training sample uses the domain temporal feature sequence of the preceding continuous time period as the input feature and the actual access popularity of each domain in the corresponding subsequent continuous time period as the supervision label.

[0011] The domain attention branch receives the input domain temporal feature sequence, calculates the association weights of each domain feature, performs weighted processing on the input features, and then transmits them to the temporal convolutional network. The temporal convolutional network captures the access dependencies between time periods through dilated convolutional layers and residual connection layers, outputs the access popularity ranking of each domain in future time periods, and combines the historical access frequency of each data block in each domain to generate the access popularity ranking of each domain and its corresponding data block.

[0012] S104, extract the target domain time series features to be processed for each edge node, input the target domain time series features into the hot data prediction model, and output the current access popularity ranking results of each professional domain and its corresponding data block for each edge node in multiple future time periods.

[0013] S105: Based on the current access popularity ranking results over multiple future time periods, assign corresponding edge nodes with the highest access popularity professional fields as attributes; any edge node will store data blocks with the same attributes as itself and whose access popularity ranking is higher than or equal to a preset threshold in the local high-speed storage medium; any edge node will migrate data blocks with different attributes or whose access popularity ranking is lower than the preset threshold to other edge nodes with the same attributes and whose storage does not exceed the storage capacity threshold.

[0014] Furthermore, professional category tags are obtained through the following methods:

[0015] Semantic recognition is used to extract the semantic features of the data blocks accessed by each customer in the historical access record set;

[0016] The semantic features of the content are matched and compared with the pre-built feature libraries of various professional fields. The matching coverage of each customer in each professional field is calculated, and the professional field with the highest matching coverage is selected as the professional field classification label for the corresponding customer.

[0017] Furthermore, the access frequency of data blocks corresponding to each professional field within each time period is statistically analyzed to construct the corresponding field time series feature sequence, specifically including:

[0018] The historical access record set is divided into time periods according to a unified time granularity. Each access record is matched to the corresponding professional field based on the professional field classification label. The access count of all data blocks under each professional field is summarized in each time period. After removing abnormal access records that exceed the normal access fluctuation range, the access statistics of each professional field are sequentially spliced ​​together in chronological order to form a continuous domain time series feature sequence.

[0019] Furthermore, the domain-attention branch receives the input domain temporal feature sequence and calculates the association weights of features from each domain, specifically including:

[0020] Extract the temporal feature representations corresponding to each professional field from the input domain temporal feature sequence. The temporal feature representations are composed of multiple access frequencies of the professional field in multiple consecutive time periods.

[0021] For any two different professional fields, calculate the similarity of their respective time-series features in terms of access frequency change trends, and use the similarity as the access correlation between the two professional fields.

[0022] For each professional field, the access correlation between each professional field and the other professional fields is aggregated to obtain the comprehensive correlation strength between professional fields. Based on the comprehensive correlation strength between professional fields, the corresponding feature weights are generated. The comprehensive correlation between professional fields is directly proportional to the feature weights.

[0023] The generated feature weights are used to perform weighted adjustment on the temporal feature representation of the corresponding professional field to obtain weighted enhanced domain temporal features, which are then transmitted to a temporal convolutional network.

[0024] Furthermore, the similarity of their respective temporal features in terms of access frequency change trends is calculated, specifically including:

[0025] Smoothing is performed on the temporal feature representations corresponding to two different professional fields to obtain two sets of continuous access frequency change trends;

[0026] Extract the trend switching nodes in the two sets of access frequency changes. The trend switching nodes correspond to the time periods when the access frequency changes.

[0027] Compare the degree of temporal matching between the trend switching nodes of the two sets of access frequency changes, as well as the degree of consistency in the overall direction of rise and fall;

[0028] Based on the weighted summation of matching degree and consistency degree, the temporal feature representations of the two professional fields are obtained to show the similarity in the trend of access frequency changes.

[0029] Furthermore, iterative training of temporal convolutional networks also includes:

[0030] The target professional fields and target data blocks output by the hot data prediction model are sorted by access popularity, and this is used as the prediction ranking list.

[0031] Obtain the measured access popularity of the target professional field and the target data block, and sort them based on the measured access popularity to obtain the actual sorted list;

[0032] The ranking of each data block in the predicted ranking list is compared with that in the actual ranking list. For data blocks whose rankings are inconsistent, the difference between the ranking of the data block in the predicted ranking list and the ranking of the data block in the actual ranking list is used as the ranking deviation cost. The ratio of the actual access frequency of the data block to the total actual access frequency of all data blocks in the subsequent consecutive time period is used as the weighting coefficient to weight the ranking deviation cost, and the loss value is obtained by summing them up.

[0033] Using the loss value, backpropagation is used to update the network parameters of the domain attention branch and the temporal convolutional network.

[0034] Furthermore, the storage capacity threshold is obtained by subtracting the preset reserved storage space from the total capacity of the local high-speed storage media of the edge node.

[0035] Furthermore, it also includes:

[0036] If the growth of any edge node reaches a preset quantity threshold within a preset time period, it is determined that the access popularity ranking of the edge node has changed, and steps S101 to S105 are re-executed based on historical data within the preset time period.

[0037] Secondly, this application also provides an AI-driven intelligent optimization system for edge node storage, comprising:

[0038] The data acquisition module is used to acquire the historical access record set of each distributed edge node, the professional field classification label of the corresponding customer, and the storage capacity threshold of the edge node; the historical access record set includes the access time of each data block and the initiating customer identifier, and the professional field classification label is used to mark the professional direction of each customer;

[0039] The feature sequence construction module is used to count the access frequency of data blocks corresponding to each professional field within each time period based on the historical access record set and professional field classification labels, and construct the corresponding field time series feature sequence.

[0040] The prediction model building module is used to input the domain temporal feature sequence into a temporal convolutional network with a domain attention branch for iterative training to obtain a hot data prediction model;

[0041] The training set consists of data from the historical access record set divided into preceding continuous time periods, and the test set consists of data from subsequent continuous time periods. Each training sample uses the domain temporal feature sequence of the preceding continuous time period as the input feature and the actual access popularity of each domain in the corresponding subsequent continuous time period as the supervision label.

[0042] The domain attention branch receives the input domain temporal feature sequence, calculates the association weights of each domain feature, performs weighted processing on the input features, and then transmits them to the temporal convolutional network. The temporal convolutional network captures the access dependencies between time periods through dilated convolutional layers and residual connection layers, outputs the access popularity ranking of each domain in future time periods, and combines the historical access frequency of each data block in each domain to generate the access popularity ranking of the corresponding data blocks in each domain.

[0043] The prediction results module is used to extract the target domain time-series features to be processed at each edge node, input the target domain time-series features into the hot data prediction model, and output the data block access popularity ranking results of each edge node in each professional field in multiple future time periods.

[0044] The optimization module is used to sort data blocks by access popularity and store data blocks whose access popularity is higher than a preset threshold in the local high-speed storage medium; and to migrate data blocks whose access popularity is lower than the preset threshold to the adjacent edge nodes whose access popularity is higher than the preset threshold and whose remaining storage capacity meets the storage requirements. After migration, the storage occupancy of the edge nodes does not exceed their own storage capacity threshold.

[0045] Thirdly, this application also provides an electronic device, including:

[0046] Memory, used to store computer programs;

[0047] A processor is used to implement the method steps as described in the first aspect when executing a computer program.

[0048] The beneficial effects of this invention are as follows:

[0049] This invention targets distributed edge storage scenarios for AI dialogue with prominent professional domain attributes. By collecting historical access records of each edge node, customer professional domain classification tags, and storage capacity thresholds, it statistically analyzes the access frequency of corresponding data blocks in each domain by time period and constructs a domain temporal feature sequence. This accurately anchors the domain access patterns formed by differences in industrial development among nodes in different regions, laying a solid feature foundation for hot data prediction. Specifically, it uses a temporal convolutional network with a domain attention branch to train a hot data prediction model. The domain attention branch is used to calculate the correlation weights of features in each professional domain and complete feature weighting, strengthening the feature contribution of high-frequency access domains in the region and adapting to the domain access emphasis of different nodes. At the same time, dilated convolutional layers and residual connection layers capture long-term access dependencies, which can accurately output the access popularity ranking of each professional domain and corresponding data block in the future multi-time period, taking into account the prediction accuracy of both domain dimension and single data block dimension. Ultimately, based on the predicted popularity, each edge node is matched with a unique professional domain attribute. Data blocks with the same attribute and popularity reaching a preset threshold are retained in local high-speed storage, while the remaining data is migrated to nodes with matching attributes and sufficient storage capacity. This not only fully meets the professional access needs of regional customers and improves the effective utilization of storage, but also reduces the cross-node data retrieval latency for professional AI dialogues, while achieving balanced optimization of storage load across edge nodes. Attached Figure Description

[0050] Figure 1 A flowchart of an AI-driven intelligent optimization method for edge node storage provided by the present invention;

[0051] Figure 2 This is a schematic diagram of the temporal convolutional network architecture with a domain attention branch in this invention;

[0052] Figure 3 This is a schematic diagram illustrating the similarity of the respective temporal characteristics in the trend of access frequency changes in this invention.

[0053] Figure 4 This is a schematic diagram illustrating the iterative training of the temporal convolutional network using a loss function in this invention.

[0054] Figure 5 A schematic diagram of an AI-driven edge node storage intelligent optimization system provided by the present invention;

[0055] Figure 6 This is a schematic diagram of an electronic device provided by the present invention.

[0056] In the diagram: 201, Data acquisition module; 202, Feature sequence construction module; 203, Prediction model construction module; 204, Prediction result module; 205, Optimization module; 301, Processor; 302, Memory. Detailed Implementation

[0057] 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, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0058] like Figure 1 As shown in the figure, an AI-driven intelligent optimization method for edge node storage proposed in this embodiment of the invention includes:

[0059] S101, obtain the historical access record set of each distributed edge node, the professional field classification label of the corresponding customer, and the storage capacity threshold of the edge node; the historical access record set includes the access time of each data block and the initiating customer identifier, and the professional field classification label is used to mark the professional direction of each customer.

[0060] Specifically, the historical access record set contains the access time and initiating customer identifier for each data block. This set is extracted in batches from the gateway access logs of each edge node, pulled and deduplicated at fixed intervals through the node's standardized log interface, retaining access timestamps accurate to the second and the customer's unique identifier, while also associating them with the unique number of the accessed data block. This allows for precise backtracking of each customer's access behavior trajectory, providing a traceable original basis for subsequent domain-based statistics on access frequency.

[0061] Professional domain classification tags are used to mark the professional direction of each customer. These tags can classify the customer's business attributes and realize domain-based aggregation of access behavior. The principle is that customer access data in the same professional domain has a high degree of content similarity and temporal correlation. Domain-based aggregation can reduce the input dimension of subsequent prediction models and improve the accuracy of hot data identification.

[0062] The professional domain classification labels are obtained as follows: First, semantic recognition is used to extract the semantic features of the data blocks accessed by each customer from the historical access record set. Specifically, a pre-trained general semantic representation model is invoked to encode the titles, abstracts, and text fragments of the data blocks, generating standardized semantic representations as content semantic features. Then, the content semantic features are matched against pre-built feature libraries for each professional domain, and the matching coverage of each customer in each professional domain is calculated. The professional domain with the highest matching coverage is selected as the corresponding customer's professional domain classification label. The professional domain feature library can be constructed by training on labeled typical literature and standard datasets in each domain. During matching, the similarity between the semantic features and the central features of each domain is calculated, and the matching coverage is calculated as the percentage of matches for all data blocks accessed by each customer in the corresponding domain. This method automatically completes domain classification based on behavioral data without requiring customers to actively fill in business information, adapting to the characteristics of a large number of customers and difficulties in information collection in distributed edge scenarios.

[0063] The matching coverage is expressed as follows:

[0064] (1)

[0065] In the formula, For customers In the Matching coverage of each professional field; For customers A collection of all data blocks that have been accessed in the past; For customers Total number of data blocks accessed in history; For data blocks The semantic features and the first Cosine similarity of central features of each professional field; For data block indexing; For data blocks Standardized semantic representation vectors (i.e., content semantic features). For the first The central feature vector of a feature library for each professional field; The preset semantic similarity matching threshold; This is an indicator function. When the condition in the parentheses is true, the function takes the value 1; when the condition in the parentheses is false, the function takes the value 0.

[0066] The storage capacity threshold is obtained by subtracting a preset reserved storage space from the total capacity of the local high-speed storage media of the edge node. This threshold represents the maximum available capacity that the edge node can use to store hot data. In other words, the nominal total capacity of the high-speed media such as solid-state drives on each edge node is pre-calculated, and the reserved space for storing system files, metadata, and temporary cache is deducted. This ensures that necessary operational buffer space is reserved for the edge node, preventing hot data from filling the storage media and causing node malfunctions, thus guaranteeing the stability of storage scheduling.

[0067] S102, based on the historical access record set and professional field classification labels, count the access frequency of the corresponding data blocks of each professional field in each time period, and construct the corresponding field time series feature sequence.

[0068] Specifically, subsequent temporal convolutional networks can only recognize continuous sequence inputs with fixed time granularity. Scattered access records cannot be directly used for model training and prediction. By aggregating statistics and constructing sequences, individual access behaviors can be elevated to domain-level temporal dimension patterns, providing a standardized data foundation for subsequent hot data prediction.

[0069] First, the historical access record set is divided into time periods according to a unified time granularity. In specific implementation, the time granularity can be configured according to the type of business carried by the edge node. In normal scenarios, 1 hour can be selected as the basic granularity. In scenarios with drastic business fluctuations, it can be adjusted to 15 minutes, and in scenarios with stable business, it can be set to 24 hours. In the process of division, the whole hour is used as the time period boundary to ensure that the time period division standard of all distributed edge nodes is completely unified.

[0070] Secondly, each access record is matched to its corresponding professional field based on the professional field classification tags. Specifically, the initiating customer identifier in each access record is used as an index to retrieve the customer professional field classification tag generated in S101. The accessed data block corresponding to that record is then grouped into the statistical group of the corresponding professional field. For cases where the same data block is accessed by customers in different fields, it is counted separately within the statistical scope of the corresponding field. The principle behind this is that customers within the same professional field have highly similar access needs. Domain aggregation can gather scattered individual access behaviors into group behavioral characteristics, effectively filtering out random access fluctuations from individual customers. Taking a dialogue AI scenario as an example, customers can be divided into three professional fields: general question answering, code generation, and copywriting. A model weight block access record from a customer in the code generation field is then grouped into the statistical group of the code generation field. If the same weight data block is simultaneously accessed by a customer in the copywriting field, it is also counted in the access statistics of the copywriting field.

[0071] Then, within each time period, the access count of all data blocks under each professional field is summarized. In practice, a sliding window streaming statistical method can be used. When the window closes at the end of each time period, the access count of all data blocks under each professional field within the window is accumulated. This not only calculates the total access frequency for each professional field during that time period but also simultaneously retains the unique access count of each data block within its corresponding field. The expression for the unique access count is:

[0072] (2)

[0073] In the formula, For the first During the time period, data blocks In the Number of unique visits in each professional field; For the relevant professional field The customer base; Index for customers; Index for access records; For customers The set of all access records for data block b; For the first The access timestamp of each access record; For the first The time interval of each period.

[0074] Finally, after removing abnormal access records that exceed the normal access fluctuation range, the access statistics results of each professional field are sequentially concatenated according to the time period to form a continuous domain time-series feature sequence. Specifically, the three-standard-deviation method is used to identify abnormal data. This involves calculating the historical mean and standard deviation based on the access frequency of the preceding ten consecutive identical time periods. Access records exceeding the range of the mean plus or minus three standard deviations are identified as abnormal. Typical anomalies include short-term high-frequency crawler requests, batch stress test calls, and duplicate accesses caused by error retries. After identification, the corresponding records are removed from the time-series statistics. Then, using the time period start timestamp as the sorting basis, the access frequencies of each domain within each time period are arranged sequentially from morning to night. Each sequence element simultaneously contains the time period identifier and the access frequency values ​​of all professional fields within the corresponding time period.

[0075] The historical average is:

[0076] (3)

[0077] In the formula, This is the historical average. This refers to the number of preceding consecutive time periods used for statistical analysis. For the first The first field Total access frequency for the time period; It is an index variable for continuous time periods.

[0078] The standard deviation is:

[0079] (4)

[0080] In the formula, For the first The first field Standard deviation of access frequency corresponding to the time period.

[0081] S103: The domain temporal feature sequence is input into a temporal convolutional network with a domain attention branch for iterative training to obtain a hot data prediction model. The training set is divided into preceding continuous time periods from the historical access record set, and the test set is divided into subsequent continuous time periods. Each training sample uses the domain temporal feature sequence of the preceding continuous time period as input features and the actual access popularity of each domain in the corresponding subsequent continuous time period as supervision labels. The domain attention branch receives the input domain temporal feature sequence, calculates the association weights of each domain feature, performs weighted processing on the input features, and then transmits them to the temporal convolutional network. The temporal convolutional network captures the access dependencies between time periods through dilated convolutional layers and residual connection layers, outputs the access popularity ranking of each domain in future time periods, and combines this with the historical access frequency of each data block within each domain to generate the access popularity ranking of each domain and its corresponding data block.

[0082] Specifically, this step aims to learn generalizable heat prediction capabilities from historical domain access patterns, and its architecture is as follows: Figure 2 As shown, this step targets AI dialogue model scenarios with strong professional domains. Since different edge nodes are responsible for customers in different regions, and different regions have significant professional domain focuses due to industrial development layouts, for example, edge nodes clustered in science and technology innovation parks mainly receive access for code generation and technical Q&A, while areas concentrated in cultural and creative industries have a higher proportion of access for copywriting and content generation. Simple general temporal prediction models cannot capture the differences in correlation between domains. Therefore, a temporal convolutional network architecture with domain attention branches is adopted. The feature weights of highly correlated domains are strengthened through the attention mechanism. Combined with the long sequence dependency capture capability of temporal convolution, the model can adapt to the differences in domain access brought about by regional industrial distribution, improve the accuracy of hot data prediction, and provide a reliable decision basis for subsequent edge storage optimization.

[0083] The training set is composed of data from consecutive periods preceding each historical access record, while the test set consists of data from consecutive periods following each period. Each training sample uses the domain temporal feature sequence from the preceding consecutive period as input features, and the actual access popularity of each domain within the corresponding subsequent consecutive period as the supervision label. This division follows the causal principle of training with temporal data. In practice, 30 consecutive days of historical access data can be selected; the data from the first 25 days is used as the training set, and the data from the last 5 days is used as the test set. The time window for input features can be set to 12 consecutive periods, corresponding to 12 hours of historical access data. The supervision label is the ranking result of the actual access popularity of each domain in the subsequent three periods. This setup aligns with the business access patterns of AI dialogue models, allowing the model to learn periodic access patterns such as weekdays and commuting hours. Furthermore, due to regional industry focuses, access cycles in different domains may differ. The construction of samples from consecutive periods can fully preserve the temporal correlation features between domains, avoiding disruption of temporal dependencies and ensuring the predictive stability of the model after deployment at real edge nodes.

[0084] The architecture employs a domain attention branch to receive the input temporal feature sequence, calculate the correlation weights of features from each professional domain, perform weighted processing on the input features, and then transmit them to the temporal convolutional network. The temporal convolutional network captures the access dependencies between time periods through dilated convolutional layers and residual connection layers, outputting the access popularity ranking of each professional domain for each future time period. It also combines this with the historical access frequency of each data block within each professional domain to generate the access popularity ranking of each professional domain and its corresponding data block. The principle behind this architecture is that the access needs of AI dialogues in different professional domains are not entirely independent. For example, the access peak in the code generation domain often coincides with the synchronous growth in the technical question-and-answer domain, while the fluctuations in the copywriting domain are highly correlated with the marketing material generation domain. The domain attention branch is used to quantify the strength of this inter-domain correlation, allowing the model to focus on domain features highly correlated with the dominant industry of the target region, while weakening noise interference from irrelevant domains. In specific implementation, the temporal convolutional network can be configured with three dilated convolutional layers, with the dilation coefficient increasing sequentially to gradually expand the receptive field to capture long-term access dependencies. Residual connections are added after each convolutional layer to avoid gradient vanishing. The network ultimately outputs the overall popularity ranking of each professional domain for the next three time periods. By combining the historical access counts of individual data blocks within each domain retained in S102, and multiplying the overall domain popularity coefficient by the historical frequency of the individual data block, the comprehensive access popularity value of each data block is calculated and sorted. This design, tailored to the characteristics of regional industry focus, enables edge nodes deployed in different regions to automatically enhance the prediction accuracy of their local dominant domain, without the need to train models separately for each node, thus reducing model operation and maintenance costs in distributed scenarios.

[0085] The domain-attention branch receives the input domain-specific temporal feature sequence and calculates the association weights of features from each domain, specifically including:

[0086] S3011: Extract the temporal feature representations corresponding to each professional field from the input domain temporal feature sequence. Each temporal feature representation consists of multiple access frequencies for that professional field over several consecutive time periods. This step forms the data foundation for domain association calculation. In practice, according to the preset professional field classification dimensions, independent one-dimensional frequency sequences are extracted from the input temporal feature sequence for each field. Each sequence contains the access frequency values ​​for the corresponding field across all time periods within the input window. For example, for the three fields of general question answering, code generation, and copywriting, access frequency sequences for twelve consecutive time periods are extracted as their respective temporal feature representations. The split single-field temporal features can eliminate numerical interference from other fields and accurately reflect the access change trajectory of a single field.

[0087] S3012: For any two different professional fields, calculate the similarity of their respective temporal feature representations in terms of access frequency change trends, and use this similarity as the access correlation between the two professional fields. Unlike traditional correlation calculations based on numerical similarity, this step focuses on the similarity of access trends. The principle is that the domain access demand of AI dialogue models is often driven by the rhythm of industry activities. Related fields driven by the same industry cluster will show synchronous rise and fall trends, rather than completely identical access values. Trend similarity can better reflect the inherent business connections between fields. For example, although the number of accesses in a single time period differs between the fields of code generation and technical Q&A, their trends of synchronous increase during weekday working hours and synchronous decrease at night are highly consistent, thus exhibiting a high access correlation. This calculation method can effectively capture the domain linkage effect driven by regional industries, avoid misjudgment of correlation due to differences in the access base of a single field, and is suitable for distributed edge scenarios where different regions have different industry focuses.

[0088] In step S3012, the similarity of their respective temporal features in terms of access frequency change trends is calculated, such as... Figure 3 As shown, it specifically includes:

[0089] The first step involves smoothing the temporal feature representations corresponding to two different professional fields, resulting in two sets of continuous access frequency trends. The smoothing process filters out random access fluctuations within a single time period, highlighting the true trend changes. Specifically, a moving average method is used for smoothing, with a window size set to three time periods. The access frequency for each time period is averaged between the preceding and following time periods as the smoothed value. Because access to edge nodes involves occasional single-batch calls, test requests, and other noise, directly calculating trends based on the original frequencies would result in numerous false trend shifts. Smoothing restores the true industry access rhythm, improving the reliability of subsequent trend matching.

[0090] The smoothed representation using the moving average method is as follows:

[0091] (5)

[0092] In the formula, For the first The first field Access frequency after time-smoothing; , , The first , , Within the time period Total access frequency for each professional field; The number of time periods set for the window size.

[0093] The second step involves extracting trend switching nodes from the changes in access frequency in both sets of data. These nodes correspond to the time periods when the access frequency changes. Trend switching nodes are key feature points reflecting changes in access rhythm. Specifically, the access frequency values ​​of adjacent time periods in the smoothed sequence are compared sequentially. If the frequency of the current time period is higher than the previous time period, it is marked as an upward trend; otherwise, it is marked as a downward trend. The time period corresponding to the change in state is the trend switching node. By extracting these switching nodes, a continuous numerical sequence can be transformed into a discrete sequence of key time points, simplifying subsequent trend matching calculations and accurately capturing the inflection point characteristics of domain access, aligning with the phased changes in industry demand.

[0094] The third step compares the temporal matching degree of the trend switching nodes in the two sets of access frequency changes, as well as the consistency of the overall upward and downward directions. This step quantifies trend similarity from two dimensions: temporal matching degree reflects whether the access inflection points of the two domains appear synchronously, and overall direction consistency degree reflects whether the rise and fall rhythms are unified over a long period. In specific implementation, temporal matching degree refers to the proportion of the number of nodes in the first professional domain whose time difference with the trend switching node of the second professional domain does not exceed a preset threshold within a preset statistical time window, out of the total number of trend switching nodes in the first professional domain. Consistency degree is calculated by the proportion of time periods in which the two domains have the same rise and fall status within the entire input window. This dual-dimensional evaluation method can comprehensively measure the trend correlation between domains, capturing both the synchronicity of short-term inflection points and the consistency of long-term trends, ensuring that the calculated access correlation degree can truly reflect the business linkage relationship between domains.

[0095] Among them, the degree of time matching The expression is as follows:

[0096] (6)

[0097] In the formula, Within the entire input time window, the first The number of trend switching nodes in each field; The number of nodes in the first professional field whose time difference with the trend switching node in the second professional field does not exceed a preset threshold.

[0098] The expression for the degree of consistency is as follows:

[0099] (7)

[0100] In the formula, For the first The and the first The overall consistency of access trends across different professional fields; Enter the starting time period number of the time window; The length of the time window for the input feature; For the first The first field Access trend status markers for different time periods; For the first The first field Access trend status markers for different time periods.

[0101] The expression for accessing the trend status flag is:

[0102] (8)

[0103] In the formula, A value of 1 indicates an upward trend; A value of -1 indicates a decreasing state; A value of 0 indicates a balanced state.

[0104] The fourth step involves a weighted summation of the matching and consistency scores to determine the similarity in the access frequency trends of the two professional fields. In practice, the weight of the matching score is set higher, followed by the overall consistency score. The weighted sum represents the access correlation between the two professional fields; a higher value indicates a stronger correlation in the access trends of the two fields. The reason for emphasizing inflection point synchronization in the weighting is that the core of hot data prediction is to anticipate the inflection point of rising access popularity. Fields with synchronized inflection points have higher reference value for popularity prediction. This setting allows attention weights to be more biased towards related fields with synchronous growth potential, improving the model's accuracy in predicting periods of rising popularity.

[0105] The formula for calculating the degree of similarity is as follows:

[0106] (9)

[0107] In the formula, For the first The and the first The degree of similarity in the changing trends of access frequency across different professional fields; The weighting coefficients represent the degree of temporal matching. Weighting coefficients for overall consistency.

[0108] S3013: For each professional domain, the access correlation between each professional domain and all other professional domains is aggregated to obtain the comprehensive correlation strength between domains. Based on this comprehensive correlation strength, corresponding feature weights are generated for each professional domain. The comprehensive correlation is directly proportional to the feature weights. The aggregation operation is implemented by summing and averaging: the access correlation of a particular domain with all other domains is added together, and the average value is taken as the comprehensive correlation strength of that domain. Then, normalization is applied to map the comprehensive correlation strengths of all domains to the range of zero to one, generating corresponding feature weights. The principle is that domains with higher correlation to other domains reflect the overall business access trend more accurately and contain richer effective predictive information, thus receiving higher feature weights. For scenarios with a focus on regional industries, the core domains corresponding to the local leading industries often have strong correlations with multiple surrounding domains, automatically receiving higher weights and thus dominating the prediction results in subsequent convolution calculations, achieving adaptive adaptation of the model to local industry characteristics.

[0109] The expression for the overall correlation strength is as follows:

[0110] (10)

[0111] In the formula, For the first The overall correlation strength between professional fields; This represents the total number of professional fields.

[0112] The expression for the feature weights is:

[0113] (11)

[0114] In the formula, For the first The feature weights corresponding to each professional field have a value range of [0,1]. This represents the minimum overall correlation strength across all domains. This represents the maximum value of the overall correlation strength across all domains.

[0115] S3014 utilizes the generated feature weights to perform weighted adjustment on the temporal feature representations of the corresponding professional domains, obtaining weighted enhanced domain temporal features, which are then transmitted to the temporal convolutional network. Weighted adjustment involves multiplying the full-time temporal feature value of each domain by the corresponding feature weight; the feature amplitude of highly correlated domains is amplified, while the feature amplitude of low-correlation domains is correspondingly reduced. Its expression is as follows:

[0116] (12)

[0117] In the formula, For the first Time-series characteristics of each professional field after weighted enhancement; For the first The original temporal characteristics of each professional field.

[0118] After weighted processing, the feature sequences input to the temporal convolutional network will highlight the changing patterns of the core associated domains and suppress noise interference from weakly associated domains. This allows the convolutional layers to learn effective temporal dependencies more accurately and avoids the impact of access fluctuations in irrelevant domains on the prediction results.

[0119] Among them, the temporal convolutional network is trained iteratively, such as Figure 4 As shown, it includes:

[0120] First, the target professional fields and target data blocks output by the hot data prediction model are sorted by access popularity, and this sorted list is used as the prediction ranking list.

[0121] Secondly, the measured access popularity of the target professional field and the target data block is obtained, and the data is sorted based on the measured access popularity to obtain the actual sorted list.

[0122] Then, the ranking of each data block in the predicted ranking list is compared with that in the actual ranking list. For data blocks with inconsistent rankings, the difference between their rankings in the predicted and actual ranking lists is used as the ranking deviation cost. The proportion of the actual access frequency of this data block to the total actual access frequency of all data blocks in subsequent consecutive time periods is used as a weighting coefficient to weight the ranking deviation cost, and the sum is the loss value. The core of this loss calculation method is the ranking optimization idea, while introducing access volume weighting. The principle is that the ranking error of popular data blocks will lead to a greater loss of storage performance. For example, if the first-ranked hot data is predicted to be the tenth, it will cause the data to not be stored in high-speed storage, resulting in a large number of access requests hitting low-speed storage, which will seriously affect the response latency of AI dialogue. The ranking deviation of unpopular data has little impact on overall performance. In specific implementation, the ranking difference is directly taken as the absolute difference between the predicted and actual rankings, and the weighting coefficient is the proportion of the actual access frequency of the data block to the total access frequency. The weighted deviation cost of all data blocks (the product of the ranking difference and the weighting coefficient) is the total loss value of this iteration. This loss function design allows the model to prioritize the sorting accuracy of high-access data, which aligns with the core objective of edge storage optimization: to maximize cache hits and improve the overall response speed of AI dialogue models within limited high-speed storage capacity.

[0123] The expression for the loss weighting coefficient is as follows:

[0124] (13)

[0125] In the formula, For data blocks The corresponding loss weighting coefficient; For subsequent consecutive time periods, data blocks The actual number of visits; This represents the total number of actual accesses to all data blocks within subsequent consecutive time periods. A collection of data blocks.

[0126] The expression for the loss value is:

[0127] (14)

[0128] In the formula, This is the loss value; For data blocks The cost of sorting bias.

[0129] The loss value is used to backpropagate and update the network parameters of the domain attention branch and the temporal convolutional network. Specifically, gradient descent is used for backpropagation, calculating the parameter gradient layer by layer from the network output. Simultaneously, the similarity calculation parameters in the domain attention branch and the convolutional kernel parameters in the temporal convolutional network are updated. Parameters are updated after each batch of training samples, until the ranking accuracy on the test set reaches a preset threshold, at which point training stops. Simultaneously updating the parameters of both modules allows the attention branch and convolutional layers to adapt to each other. Adjusting the attention weights, in conjunction with feature extraction from the convolutional layers, optimizes the ranking prediction performance, ultimately resulting in a hot data prediction model that can be directly deployed on various edge nodes.

[0130] Taking an edge node deployed in a science and technology industrial park as an example, the customers in this area are mainly software R&D companies, focusing on code generation and technical Q&A, with a small amount of access in the copywriting field. During training, after inputting a continuous twelve-hour sequence of domain time-series features, the domain attention branch first extracts the time-series feature representations of the three domains. After smoothing by moving average, it extracts the trend switching nodes for each time period. The calculation shows that code generation and technical Q&A have a high degree of trend similarity, while code generation and copywriting have a low degree of similarity, and the similarity between technical Q&A and copywriting is also low. After aggregation, the code generation domain has the highest comprehensive correlation strength and obtains the highest feature weight, while the copywriting domain has the lowest weight. The weighted features are input into a temporal convolutional network, which captures the periodic access patterns within 24 hours through three layers of dilated convolution, and outputs the popularity ranking of each domain and data block for the next three hours. During training iterations, if the predicted position of the top-ranked code generation model core weight block matches the actual position, the loss is minimal. If the predicted position of this data block deviates, it will generate a higher weighted loss due to its high access volume, driving the model to prioritize correcting the ranking deviation of this type of high-value data. The final trained model can accurately predict peak access times in the code generation field of the park and store the corresponding hot data in high-speed storage in advance, effectively improving the response speed of AI dialogue.

[0131] S104: Extract the target domain time-series features to be processed from each edge node, input the target domain time-series features into the hot data prediction model, and output the current access popularity ranking results of each professional domain and its corresponding data block for each edge node in multiple future time periods.

[0132] S105: Based on the current access popularity ranking results over multiple future time periods, assign corresponding edge nodes with the highest access popularity professional fields as attributes; any edge node will store data blocks with the same attributes as itself and whose access popularity ranking is higher than or equal to a preset threshold in the local high-speed storage medium; any edge node will migrate data blocks with different attributes or whose access popularity ranking is lower than the preset threshold to other edge nodes with the same attributes and whose storage does not exceed the storage capacity threshold.

[0133] Specifically, for AI dialogue model scenarios with strong professional focus, different edge nodes are responsible for customers in different regions, and different regions have significant professional focus due to industrial development layouts. Therefore, based on the current access popularity ranking results over multiple future time periods, the edge nodes are assigned attributes based on the top professional fields with the highest access popularity. For example, edge nodes clustered in science and technology innovation parks mainly receive access for code generation and technical Q&A, while areas concentrated in cultural and creative industries have a higher proportion of access for copywriting and content generation. Therefore, assigning attribute tags corresponding to high-popularity fields to edge nodes allows the node's storage scheduling strategy to align with the local business structure, fundamentally avoiding the misallocation and waste of high-speed storage resources caused by general caching strategies. In specific implementation, the top two professional fields with the highest cumulative access popularity over the next three consecutive time periods are selected as the attributes of the corresponding edge nodes. The local storage scheduling module of the node writes the attribute information into the node metadata table. The attribute tags are updated synchronously with each round of popularity prediction results, rather than remaining fixed. This ensures that the node attributes always match the recent access patterns, allowing the storage resources of the edge node to be tilted towards the local core business domain, accurately adapting to the domain-specific access characteristics of AI dialogue models, and improving the targeting of caching strategies from the node positioning level.

[0134] Any edge node will store data blocks with the same attributes as itself and whose access popularity ranking is higher than or equal to a preset threshold on local high-speed storage media. The domain matching the node attributes is the core access direction of local customers. Data blocks that meet the popularity ranking criteria are high-frequency access hot data in the future. Storing them on local high-speed storage media such as solid-state drives can reduce the data reading latency of AI dialogue requests and improve the interaction response speed. In specific implementation, the preset threshold can be set to the top 30% of popularity in the corresponding professional domain. During the scheduling process, all data blocks to be cached are first traversed, and the domain and node attributes of the data blocks are matched first. If the attributes are consistent, the popularity ranking in the domain is compared. Data blocks that meet the threshold requirements are written to the high-speed storage partition in descending order of popularity until the used capacity of the high-speed storage reaches the storage capacity threshold set in S101, at which point the writing automatically stops.

[0135] Any edge node will migrate data blocks with attributes different from its own or whose access popularity ranking is below a preset threshold to other edge nodes with the same attributes and whose storage does not exceed the storage capacity threshold. Storing domain data that is not related to the node's attributes or data blocks whose access popularity does not meet the standard locally will consume valuable high-speed storage resources, while other edge nodes with the corresponding domain attributes have a higher demand for accessing this type of data. Cross-node migration can achieve global overall optimization of distributed storage resources and avoid the problem of both idle and insufficient storage resources on a single node. In practice, before migration, the remaining high-speed storage capacity of all edge nodes with corresponding domain attributes across the entire network is queried through the metadata broadcast channel of the distributed edge network. The target node is selected according to the priority of remaining capacity from high to low. The migration process adopts an incremental transmission method, transmitting only the data blocks not stored on the target node to avoid repeated transmission and wasting bandwidth resources. After the migration is completed, the global data index table is updated synchronously, and the index pointers of the corresponding data blocks are retained locally. If there is a temporary cross-domain access requirement in the future, the data location can be quickly located through the index and retrieved across nodes. This design can fully utilize the distributed storage capacity of edge nodes across the entire network, allowing hot data from various domains to be centrally stored on nodes with corresponding business attributes. This not only alleviates the storage pressure on a single node, but also improves the data sharing efficiency between nodes with the same industry attributes, adapting to the multi-domain and widely distributed deployment characteristics of AI dialogue models.

[0136] Furthermore, if the growth of any edge node reaches a preset threshold within a preset time period, it is determined that the access popularity ranking of that edge node has changed, and steps S101 to S105 are re-executed based on historical data within the preset time period. The regional industrial structure, customer scale, and business needs will dynamically change over time, and fixed node attributes and caching strategies will gradually deviate from actual access patterns. Setting a dynamically triggered full-process update mechanism ensures that storage optimization strategies always match real business access characteristics, avoiding a continuous decline in cache hit rate due to model aging and strategy solidification. In specific implementation, the preset time period can be set to 60 consecutive calendar days, and the growth threshold can be set to a cumulative increase of 40% in the access frequency percentage. When the access percentage is continuously increasing for three consecutive days and cumulatively reaches the threshold, a full-process update is automatically triggered, re-fetching historical access records within the corresponding preset time period, updating customer professional field classification tags, fine-tuning the hot data prediction model, and re-dividing node attributes and data cache distribution. This mechanism achieves closed-loop adaptive edge storage optimization, adapting to dynamic changes in regional business without manual operation and maintenance intervention, ensuring long-term stable response performance of the AI ​​dialogue model under business fluctuations and industrial adjustment scenarios.

[0137] like Figure 5 As shown, based on the same inventive concept, this embodiment provides an AI-driven edge node storage intelligent optimization system, including:

[0138] The data acquisition module 201 is used to acquire the historical access record set of each distributed edge node, the professional field classification label of the corresponding customer, and the storage capacity threshold of the edge node; the historical access record set includes the access time of each data block and the initiating customer identifier, and the professional field classification label is used to mark the professional direction of each customer;

[0139] The feature sequence construction module 202 is used to count the access frequency of data blocks corresponding to each professional field in each time period based on the historical access record set and professional field classification labels, and construct the corresponding field time series feature sequence.

[0140] The prediction model building module 203 is used to input the domain temporal feature sequence into a temporal convolutional network with a domain attention branch for iterative training to obtain a hot data prediction model.

[0141] The training set consists of data from the historical access record set divided into preceding continuous time periods, and the test set consists of data from subsequent continuous time periods. Each training sample uses the domain temporal feature sequence of the preceding continuous time period as the input feature and the actual access popularity of each domain in the corresponding subsequent continuous time period as the supervision label.

[0142] The domain attention branch receives the input domain temporal feature sequence, calculates the association weights of each domain feature, performs weighted processing on the input features, and then transmits them to the temporal convolutional network. The temporal convolutional network captures the access dependencies between time periods through dilated convolutional layers and residual connection layers, outputs the access popularity ranking of each domain in future time periods, and combines the historical access frequency of each data block in each domain to generate the access popularity ranking of the corresponding data blocks in each domain.

[0143] Prediction Result Module 204 is used to extract the target domain time series features to be processed for each edge node, input the target domain time series features into the hot data prediction model, and output the data block access popularity ranking results for each professional field in multiple future time periods for each edge node.

[0144] The optimization module 205 is used to store data blocks whose access popularity is higher than a preset threshold in the local high-speed storage medium according to the data block access popularity sorting results; and to migrate data blocks whose access popularity is lower than the preset threshold to the adjacent edge nodes whose access popularity is higher than the preset threshold and whose remaining storage capacity meets the storage requirements. After migration, the storage occupation of the edge nodes does not exceed their own storage capacity threshold.

[0145] like Figure 6 As shown, based on the same inventive concept, this embodiment provides an electronic device, including:

[0146] Memory 302 is used to store computer programs;

[0147] Processor 301 is used to implement the method steps as described in the first aspect when executing a computer program.

[0148] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0149] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An AI-driven intelligent optimization method for edge node storage, characterized in that, include: S101, obtain the historical access record set of each distributed edge node, the professional field classification label of the corresponding customer, and the storage capacity threshold of the edge node; the historical access record set includes the access time of each data block and the initiating customer identifier, and the professional field classification label is used to mark the professional direction of each customer; S102, based on the historical access record set and the professional field classification label, count the access frequency of the corresponding data block of each professional field in each time period, and construct the corresponding field time series feature sequence; S103, The domain temporal feature sequence is input into a temporal convolutional network with a domain attention branch for iterative training to obtain a hot data prediction model; The data from the historical access record set, divided into preceding continuous time periods, is the training set, and the data from subsequent continuous time periods is the test set. Each training sample uses the domain temporal feature sequence of the preceding continuous time period as the input feature, and the actual access popularity of each domain in the corresponding subsequent continuous time period is used as the supervision label. The domain attention branch receives the input domain temporal feature sequence, calculates the association weights of each professional domain feature, performs weighted processing on the input features, and then transmits them to the temporal convolutional network. The temporal convolutional network captures the access dependencies between time periods through dilated convolutional layers and residual connection layers, outputs the access popularity ranking of each professional domain in future time periods, and combines the historical access frequency of each data block in each professional domain to generate the access popularity ranking of each professional domain and its corresponding data block. S104, extract the target domain time series features to be processed for each edge node, input the target domain time series features into the hot data prediction model, and output the current access popularity ranking results of each professional domain and its corresponding data block for each edge node in multiple future time periods. S105, according to the current access popularity ranking results in the future multiple time periods, assign corresponding edge nodes based on the multiple professional fields with the highest access popularity as attributes; any edge node stores data blocks with the same attributes as itself and whose access popularity ranking is higher than or equal to a preset threshold in the local high-speed storage medium; any edge node migrates data blocks with different attributes or whose access popularity ranking is lower than the preset threshold to other edge nodes with the same attributes and whose storage does not exceed the storage capacity threshold.

2. The AI-driven intelligent optimization method for edge node storage according to claim 1, characterized in that, The professional field classification tags were obtained through the following methods: Semantic recognition is used to extract the semantic features of the data blocks accessed by each customer in the historical access record set; The semantic features of the content are matched and compared with the pre-set feature libraries of each professional field. The matching coverage of each customer in each professional field is calculated, and the professional field with the highest matching coverage is selected as the professional field classification label of the corresponding customer.

3. The AI-driven intelligent optimization method for edge node storage according to claim 1, characterized in that, The method involves statistically analyzing the access frequency of data blocks corresponding to each professional field within each time period and constructing a corresponding field-specific time-series feature sequence, specifically including: The historical access record set is divided into time periods according to a unified time granularity. Each access record is matched to the corresponding professional field according to the professional field classification label. The access count of all data blocks under each professional field is summarized in each time period. After removing abnormal access records that exceed the normal access fluctuation range, the access statistics of each professional field are sequentially spliced ​​together in chronological order to form a continuous domain time series feature sequence.

4. The AI-driven intelligent optimization method for edge node storage according to claim 1, characterized in that, The domain-attention branch receives the input domain temporal feature sequence and calculates the association weights of features from each professional domain, specifically including: Extract the temporal feature representation corresponding to each professional field from the input temporal feature sequence of the field. The temporal feature representation is composed of multiple access frequencies of the professional field in multiple consecutive time periods. For any two different professional fields, calculate the degree of similarity in the access frequency change trend of their respective time-series features, and use the degree of similarity as the access correlation between the two professional fields; For each professional field, the access correlation between each professional field and the other professional fields is aggregated to obtain the comprehensive inter-field correlation strength of that professional field, and the feature weight corresponding to the professional field is generated according to the comprehensive inter-field correlation strength. The comprehensive inter-field correlation is proportional to the feature weight. The generated feature weights are used to perform weighted adjustment on the temporal feature representation of the corresponding professional field to obtain weighted enhanced domain temporal features, and the weighted enhanced domain temporal features are transmitted to the temporal convolutional network.

5. The AI-driven intelligent optimization method for edge node storage according to claim 4, characterized in that, The calculation of the respective temporal features represents the degree of similarity in the trend of access frequency changes, specifically including: Smoothing is performed on the temporal feature representations corresponding to the two different professional fields to obtain two sets of continuous access frequency change trends; Extract the trend switching nodes in the two sets of access frequency change trends, where the trend switching nodes correspond to the time periods when the access frequency rises or falls. Compare the temporal matching degree of the trend switching nodes of the two sets of access frequency changes, as well as the consistency of the overall rise and fall directions; Based on the weighted sum of the matching degree and the consistency degree, the similarity of the temporal feature representations of the two professional fields in terms of access frequency change trends is obtained.

6. The AI-driven intelligent optimization method for edge node storage according to claim 1, characterized in that, The iterative training of the temporal convolutional network also includes: The access popularity of the target professional fields and target data blocks output by the hot data prediction model is sorted to form a prediction ranking list; Obtain the measured access popularity of the target professional field and the target data block, and sort them based on the measured access popularity to obtain the actual sorted list; The ranking of each data block in the predicted ranking list is compared with that in the actual ranking list. For data blocks whose rankings are inconsistent, the difference between the ranking of the data block in the predicted ranking list and the ranking of the data block in the actual ranking list is used as the ranking deviation cost. The ratio of the actual access count of the data block to the total actual access count of all data blocks in the subsequent continuous time period is used as the weighting coefficient. The ranking deviation cost is weighted and accumulated to obtain the loss value. Using the loss value, the network parameters of the neighborhood attention branch and the temporal convolutional network are updated via backpropagation.

7. The AI-driven intelligent optimization method for edge node storage according to claim 1, characterized in that, The storage capacity threshold is obtained by subtracting the preset reserved storage space from the total capacity of the local high-speed storage medium of the edge node.

8. The AI-driven intelligent optimization method for edge node storage according to claim 1, characterized in that, Also includes: If the growth of any edge node within a preset time period reaches a preset quantity threshold, it is determined that the access popularity ranking of the edge node has changed, and steps S101 to S105 are re-executed based on historical data within the preset time period.

9. An AI-driven intelligent optimization system for edge node storage, using the AI-driven intelligent optimization method for edge node storage as described in claim 1, characterized in that, include: The data acquisition module is used to acquire the historical access record set of each distributed edge node, the professional field classification label of the corresponding customer, and the storage capacity threshold of the edge node; the historical access record set includes the access time of each data block and the initiating customer identifier, and the professional field classification label is used to mark the professional direction of each customer; The feature sequence construction module is used to calculate the access frequency of data blocks corresponding to each professional field in each time period based on the historical access record set and the professional field classification label, and construct the corresponding field time series feature sequence. The prediction model building module is used to input the domain temporal feature sequence into a temporal convolutional network with a domain attention branch for iterative training to obtain a hot data prediction model. The data from the historical access record set, divided into preceding continuous time periods, is the training set, and the data from subsequent continuous time periods is the test set. Each training sample uses the domain temporal feature sequence of the preceding continuous time period as the input feature, and the actual access popularity of each domain in the corresponding subsequent continuous time period is used as the supervision label. The domain attention branch receives the input domain temporal feature sequence, calculates the association weight of each professional domain feature, performs weighted processing on the input features, and then transmits them to the temporal convolutional network. The temporal convolutional network captures the access dependency relationship between time periods through dilated convolutional layers and residual connection layers, outputs the access popularity ranking of each professional domain in future time periods, and generates the access popularity ranking of the corresponding data blocks in each professional domain by combining the historical access frequency of each data block in each professional domain. The prediction result module is used to extract the target domain time-series features to be processed for each edge node, input the target domain time-series features into the hot data prediction model, and output the data block access popularity ranking results for each edge node in each professional field in multiple future time periods. The optimization module is used to store data blocks whose access popularity is higher than a preset threshold in the local high-speed storage medium according to the data block access popularity sorting result; and to migrate data blocks whose access popularity is lower than the preset threshold to the adjacent edge nodes whose access popularity is higher than the preset threshold and whose remaining storage capacity meets the storage requirements. After migration, the storage occupation of the edge nodes does not exceed their own storage capacity threshold.

10. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement, when executing the computer program, an AI-driven edge node storage intelligent optimization method as described in any one of claims 1 to 8.

Images