A data cache optimization method in an edge computing environment

Abstract

The application discloses a data cache optimization method in an edge computing environment, and relates to the technical field of edge computing data cache optimization, which comprises collecting original access logs from a target edge node and performing structured processing. Semantic distillation, periodic pattern mining and event intensity analysis are respectively performed on data topic labels, access heat time series and update event time series to obtain data semantic fingerprints, access heat evolution trajectories and update event intensity spectra. These features are input into a feature coupling device for collaborative evolution to generate a comprehensive cache feature map. A cache value evaluation network is constructed based on the map, data value is dynamically calibrated, and a hierarchical value list is formed, and then an adaptive strategy containing a cache order and a space pre-allocation plan is derived, which is finally converted into a configuration instruction set and issued to the target node. The application realizes accurate matching of cache strategies and multi-dimensional dynamic features of data, and improves the overall efficiency of the cache system.

Description

Technical Field

[0001] This invention belongs to the field of edge computing data caching optimization technology, specifically a data caching optimization method in an edge computing environment. Background Technology

[0002] In edge computing environments, data caching is a crucial element in optimizing service performance. Current mainstream methods are mainly based on principles such as least recently used or least frequently used, relying on simple statistics of historical access records, such as the access time or frequency of data items. These techniques abstract caching decisions into a ranking problem based on a single or limited static metric, failing to deeply analyze and utilize the multi-dimensional information contained in access behavior.

[0003] Existing solutions have shortcomings. Traditional methods fail to identify the semantic attributes and business context differences of the data itself, and apply a uniform value measurement standard to all data. They lack the ability to analyze complex patterns of access popularity evolving over time, and struggle to capture dynamic features such as periodicity and trends. Furthermore, these methods often ignore the temporal regularity of data update events and their impact on cache effectiveness, easily retaining data nearing expiration. In resource-constrained and highly variable edge environments, these limitations directly lead to low caching efficiency and increased redundant data transmission.

[0004] A technique is needed to systematically extract multi-dimensional features such as data semantics, fine-grained access timing patterns, and update event intensity from raw access logs. A method is also needed to perform nonlinear fusion and collaborative analysis of these heterogeneous features, and to achieve dynamic and accurate cache value assessment through an adaptive model, thereby generating an optimization strategy that matches the real-time state of edge nodes. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art;

[0006] To this end, the present invention proposes a data caching optimization method in an edge computing environment, comprising:

[0007] Raw cached data access logs are collected from the target edge computing node, restructured into a structured form, and a cached data access record set is generated. Semantic distillation, periodic pattern mining, and event intensity analysis are performed on the data topic tags, multi-level access heat time series, and adjacent update event time series in the cached data access record set to obtain data semantic fingerprints, access heat evolution trajectories, and update event intensity spectra.

[0008] The data semantic fingerprint, the access popularity evolution trajectory, and the update event intensity spectrum are input into the feature coupling device for co-evolution to generate a comprehensive cache feature map;

[0009] Based on the comprehensive cache feature map, a cache value evaluation network is constructed to dynamically calibrate the value, forming a hierarchical value list that includes data identification and quantified cache value.

[0010] Based on the hierarchical value list, an adaptive caching strategy for the target edge computing node is generated through a caching strategy derivation engine. The adaptive caching strategy includes the cache object retention order and space pre-allocation plan.

[0011] The adaptive caching strategy is translated into a set of device-executable cache configuration update instructions, and the set of cache configuration update instructions is pushed to the cache controller of the target edge node.

[0012] Furthermore, obtaining the data semantic fingerprint, the access popularity evolution trajectory, and the updated event intensity spectrum includes:

[0013] Raw cached data access logs containing data identifiers, content topic descriptions, and access timestamps are collected from the target edge computing node to form a raw log sequence;

[0014] The original log sequence is restructured in a structured manner to generate a set of cached data access records in a unified format;

[0015] Perform multiple rounds of semantic distillation on the data topic tags to obtain data semantic fingerprints, including:

[0016] The data topic tags are input into a preset domain semantic network to perform topic concept mapping and output a preliminary concept set;

[0017] The initial concept set is then grouped into synonymous concepts and elevated to hierarchical concepts to extract the core concept cluster.

[0018] The core concept cluster is embedded into a context-aware encoder to generate context-enhanced concept vectors.

[0019] The context-enhanced concept vector is reduced and encoded using sparse coding techniques to generate the data semantic fingerprint, which has uniqueness and representational power.

[0020] Furthermore, periodic pattern mining is performed on the multi-level access popularity time series to obtain the access popularity evolution trajectory, including:

[0021] Multiple observation granularities are defined on the multi-level access popularity time series. The original access popularity data is windowed at different observation granularities to form multi-scale time slices.

[0022] Trend fitting and fluctuation decomposition are performed on the access popularity data within each multi-scale time slice to separate the time series of trend components, periodic components and residual components.

[0023] By aligning and fusing the trend components, periodic components, and residual components corresponding to different observation granularities, the evolution trajectory of the access popularity is constructed, which reflects the coordinated changes of access popularity in the three dimensions of trend, periodicity, and randomness.

[0024] Furthermore, the step of co-evolving the data semantic fingerprint, the access popularity evolution trajectory, and the update event intensity spectrum into the feature coupling device to generate a comprehensive cache feature map includes:

[0025] An interactive evolution layer is established in the feature coupling device. The interactive evolution layer receives the data semantic fingerprint, access popularity evolution trajectory and update event intensity spectrum as the initial node state.

[0026] In the interactive evolution layer, dynamic connection relationships are established between feature nodes. These dynamic connection relationships simulate the information flow and state influence between the data semantic fingerprint, access popularity evolution trajectory, and update event intensity spectrum based on preset interaction rules.

[0027] After a preset evolution cycle, the state of each feature node in the interactive evolution layer tends to stabilize. At this time, the stable node states and their connection relationships are output in a graph form, forming the comprehensive cached feature graph.

[0028] Furthermore, based on the comprehensive cache feature map, a cache value evaluation network is constructed to dynamically label the value, forming a hierarchical value list that includes data identification and quantified cache value, including:

[0029] The feature nodes in the comprehensive cache feature map are used as the input layer nodes of the cache value evaluation network;

[0030] A multi-layer value-perceiving neuron is constructed, and the features of the input layer nodes are nonlinearly transformed and aggregated through the multi-layer value-perceiving neuron, and then passed to the output layer layer by layer.

[0031] In the output layer, each raw data identifier corresponds to a value labeling neuron, which converts the received aggregated information into a continuous quantized cached value.

[0032] All data identifiers and their corresponding quantized cache value values ​​are automatically grouped and sorted according to the range of the quantized cache value values ​​to generate the hierarchical value list.

[0033] Further, the step of generating an adaptive caching strategy for the target edge computing node based on the hierarchical value list through a caching strategy derivation engine includes:

[0034] The hierarchical value list, along with a real-time snapshot of the resource capacity obtained from the target edge computing node, is input into the caching strategy derivation engine.

[0035] In the cache strategy derivation engine, a balance model between resource constraints and value demands is established. The balance model uses the quantified cache value in the hierarchical value list as the quantified expression of value demands, and the available cache space in the resource capacity snapshot as the boundary condition of resource constraints.

[0036] Solving the balance model yields the set of cache objects that maximize the overall cache value and the corresponding ideal allocation space under the resource constraint boundary conditions.

[0037] The set of cached objects and their corresponding ideal allocation spaces are adapted to the physical characteristics of the edge node storage medium to generate the cached object retention order and space pre-allocation plan, which together constitute the adaptive caching strategy.

[0038] Furthermore, the adaptation adjustment of the cache object set and its corresponding ideal allocation space according to the physical characteristics of the edge node storage medium includes:

[0039] Identify the type of storage medium used by the target edge computing node and its read / write performance profile;

[0040] Analyze the expected access pattern of each object in the cached object set, and extract the expected access pattern from the access popularity evolution trajectory;

[0041] The matching degree between the expected access pattern and the read / write performance feature profile is calculated to recommend the most suitable storage medium region for each cache object.

[0042] Within the recommended optimal storage medium area, the final allocated physical space size is fine-tuned based on the ideal allocated space size and the physical block size characteristics of the storage medium area, thus completing the adaptation adjustment.

[0043] Further, translating the adaptive caching strategy into a device-executable cache configuration update instruction set includes:

[0044] The cache object retention order and space pre-allocation plan in the adaptive caching strategy are analyzed to identify the cache entries that need to be added, modified, retained or removed and their corresponding specific operations.

[0045] Consult the cache control instruction manual that matches the target edge computing node model, and map the identified specific operations to standard underlying cache control commands;

[0046] Based on the execution sequence implied by the retention order of the cached objects, the underlying cache control commands of the standard are arranged into a sequence of instructions with dependencies and execution order, forming the cache configuration update instruction set.

[0047] Furthermore, the step of arranging the standard underlying cache control commands into a sequence of instructions with dependencies and execution order according to the execution timing implied by the retention order of the cached objects includes:

[0048] The retention order of the cached objects is analyzed to determine the dependency chain between the cached objects. The dependency chain indicates that the removal or migration of some cached objects is a prerequisite for the pre-allocation of space for other cached objects.

[0049] Based on the dependency chain, label the set of preceding commands for each standard underlying cache control command;

[0050] Based on the aforementioned set of preceding commands, a directed acyclic graph model is used to perform topological sorting on the underlying cache control commands of the standard, generating a linear instruction execution queue without circular dependencies, which serves as the instruction sequence.

[0051] Furthermore, the construction of multi-layer value-perceiving neurons, which involves non-linearly transforming and aggregating the features of the input layer nodes through the multi-layer value-perceiving neurons, and then transmitting them layer by layer to the output layer, includes:

[0052] Construct a multi-layer value-perceiving neural network with gated recurrent units, wherein each layer of neurons contains a gated recurrent unit module;

[0053] The features of the input layer nodes are used as the initial temporal state and input into the gated recurrent unit module of the first layer of value-perceiving neurons.

[0054] Within the gated recurrent unit module of each neuron layer, the following operations are performed: the input features from the previous layer and the hidden state of the previous time step are selectively updated using the update gating mechanism; the hidden state of the previous time step is controlled by the reset gating mechanism to control the degree of influence of the hidden state of the previous time step on the current candidate state; and the new hidden state of the current time step is generated by combining the results.

[0055] The new hidden state output by each layer of gated loop unit module is transformed by a nonlinear activation function, and the transformed result is used as the information aggregation output of each layer.

[0056] The information from the previous layer is aggregated and output as the input feature sequence for the value-perceiving neurons in the next layer. For the neurons in the next layer, the gated loop operation and nonlinear transformation process are repeatedly executed, iterating layer by layer until the information is transmitted to the last layer of neurons, thereby realizing the layer-by-layer transmission and refinement of features between layers.

[0057] Compared with the prior art, the beneficial effects of the present invention are:

[0058] By performing semantic distillation, periodic pattern mining, and event intensity analysis on data topic tags, multi-level access popularity time series, and adjacent update event time series, semantic fingerprints, access popularity evolution trajectories, and update event intensity spectra are extracted from the original access logs. This process achieves a deep characterization of the multi-dimensional dynamic features of cached data. Semantic fingerprints encode the business attributes of the data, enabling caching decisions to differentiate based on semantic categories. The popularity evolution trajectory analyzes the structural patterns of access requests over time, including periodic and trend changes. The update event intensity spectrum quantifies the frequency and intensity of modifications at the data source, used to assess data stability. The extraction of these features allows caching strategies to comprehensively consider the business value, time sensitivity, and state validity of the data.

[0059] By inputting the aforementioned features into a feature coupling device for co-evolution, the nonlinear correlation between semantics, temporal popularity, and update intensity can be learned, generating a comprehensive cache feature map. A cache value evaluation network built upon this map enables dynamic labeling of data cache value. This network maps complex feature combinations into precise quantitative values ​​through a parameterized model, replacing static evaluation methods based on fixed formulas. This mechanism allows the value labeling process to adaptively respond to dynamic changes in access patterns and data states, thereby achieving optimal allocation of cache space based on the real-time value spectrum of the data. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating the steps of the data caching optimization method in the edge computing environment described in this invention.

[0061] Figure 2 To obtain a flowchart of the evolution trajectory of visit popularity;

[0062] Figure 3 A flowchart for creating a tiered value list;

[0063] Figure 4 Analysis diagram of cache value and space allocation in edge computing cache optimization scenarios;

[0064] Figure 5 A radar chart for multi-dimensional evaluation of data semantic fingerprints. Detailed Implementation

[0065] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0066] See Figure 1Raw cached data access logs are collected from the storage system of the target edge computing node. These logs record access events for each cached data item. The collected raw cached data access logs are cleaned and restructured, transforming them into a set of cached data access records containing standard fields. Each record in this set contains at least a data topic tag, a multi-level access heat time series reflecting different time granularities, and a neighboring update event time series recording adjacent update operations. Semantic distillation is performed on the data topic tags to extract their core semantic information and generate a data semantic fingerprint. Periodic pattern mining is performed on the multi-level access heat time series to analyze its access patterns and form an access heat evolution trajectory. Event intensity analysis is performed on the neighboring update event time series to evaluate update frequency and intensity and obtain an update event intensity spectrum. The obtained data semantic fingerprint, access heat evolution trajectory, and update event intensity spectrum are used as inputs and imported into a pre-set feature coupling device. In this device, features interact and co-evolve, ultimately generating a comprehensive cache feature map that can comprehensively reflect multiple characteristics of the cached data. The aforementioned feature coupling device is a dedicated data processing device adapted for edge computing cache feature analysis, consisting of a hardware layer and a software layer. The hardware layer includes an embedded processor, a data interaction module, and a timing computation unit, used to receive feature data and perform high-speed computation. The software layer includes an interaction evolution layer, a feature node management module, and a graph generation module, forming the core functional layer of the device. The interaction evolution layer is the core execution unit for feature coupling, used to realize the interaction and state evolution of feature nodes in data semantic fingerprinting, access popularity evolution trajectory, and update event intensity spectrum. The feature node management module maintains the initial state and dynamic connection relationships of feature nodes. The graph generation module transforms the stabilized feature nodes and their connection relationships into a visualized comprehensive cache feature graph. This device can be integrated into the cache controller of an edge computing node or deployed in an edge-side gateway to achieve feature coupling analysis of multiple edge nodes. Its core function is to receive heterogeneous cache feature data, achieve nonlinear correlation learning between features through preset interaction rules, and finally output a comprehensive cache feature graph. The preset state vector change rate threshold in the interactive evolution layer is a dynamically adaptable threshold that can be fine-tuned according to the application scenario of the target edge computing node: for scenarios with high requirements for cache feature analysis accuracy, such as industrial IoT and real-time monitoring, the threshold can be lowered to 0.5%, i.e., the average change rate over 10 consecutive time steps is <0.5%; for ordinary civilian edge computing scenarios, the threshold can be kept at 1%, achieving a balance between analysis accuracy and computational efficiency. Based on this comprehensive cache feature map, a cache value evaluation network is constructed. This network learns and calculates the features in the map, dynamically assigning a quantified cache value value to each cached data item, and then organizing all data identifiers and their quantified cache values ​​into a hierarchical value list according to their numerical ranges.Based on this tiered value list and combined with the real-time resource status obtained from the target edge computing node, a caching strategy inference engine is used to calculate and infer a customized caching strategy specifically for that node. This strategy explicitly defines the retention order of cached objects and the pre-allocation plan in storage space. This customized caching strategy is translated into a set of cache configuration update instructions that the target edge computing node's cache controller can recognize and execute. This instruction set is then pushed to the target edge node, where its cache controller loads and executes it, thereby completing the optimized update of the cache configuration.

[0067] See Figure 2In one embodiment of the present invention, raw cached data access logs containing data identifiers, content topic descriptions, and access timestamps are collected from target edge computing nodes to form a raw log sequence. The raw log sequence is then subjected to structured reorganization. Log entries are parsed according to a preset data pattern, and uniform fields are filled to generate a cached data access record set with a standardized format. Each record in this cached data access record set contains a data topic tag, a multi-level access popularity time series, and a nearby update event time series. Multiple rounds of semantic distillation are performed on the data topic tags to obtain data semantic fingerprints. Specifically, the data topic tags are input into a preset domain semantic network for topic concept mapping. The domain semantic network maps text tags to predefined concept nodes, outputting a preliminary concept set. This preliminary concept set is then subjected to synonym merging and hierarchical concept climbing, merging concepts with different expressions but the same meaning, and summarizing specific concepts to their superior general concepts, thereby extracting a core concept cluster. A context-aware encoder is used to embed this core concept cluster into a context, considering the co-occurrence relationships between elements in the concept cluster to generate context-enhanced concept vectors. Sparse coding techniques are used to reduce and encode the features of the context-enhanced concept vector, compressing the vector dimension and highlighting key features to generate a unique and representative semantic fingerprint. Periodic pattern mining is performed on multi-level access popularity time series to obtain the evolution trajectory of access popularity. Specifically, multiple observation granularities are defined on the multi-level access popularity time series, such as minute, hour, and day levels. Sliding windows are used to extract the original access popularity data at different observation granularities, forming multi-scale time slices. Trend fitting and fluctuation decomposition are performed on the access popularity data within each multi-scale time slice. Time series decomposition methods are applied to separate the time series of the trend component representing the long-term direction of change, the periodic component reflecting regular fluctuations, and the residual component representing random noise. The trend component, periodic component, and residual component corresponding to different observation granularities are aligned on the time axis and numerically fused to construct an evolution trajectory of access popularity that reflects the coordinated changes in access popularity across the three dimensions of trend, periodicity, and randomness. Perform event intensity analysis on the time series of neighboring update events, calculate the frequency and density of update events per unit time, assess the degree of event clustering, and obtain the update event intensity spectrum.

[0068] In practice, raw cached data access logs within a continuous time window are collected from the caching system of the target edge computing node deployed at the IoT gateway in an industrial park. Each record in the raw cached data access log contains data identifiers such as "DataID-001", "sensor temperature data aggregation result", and "2023-10-26 08:30:15", along with a content topic description and access timestamp information. All these records are arranged chronologically to form a raw log sequence. The raw log sequence is then structurally reorganized, extracting key fields from each record according to preset parsing rules and filling them into a uniformly formatted template to generate a cached data access record set. Each record in the cached data access record set contains a data topic tag of "temperature sensor-aggregation-alarm", a multi-level access heat time series consisting of access counts in minutes, hours, and days, and a time series of nearby update events recording the time points when the data was updated by neighboring operations.

[0069] Multi-round semantic distillation is performed on data topic tags to obtain data semantic fingerprints. In specific implementation, the data topic tags "temperature sensor-aggregation-alarm" are input into a pre-defined industrial IoT domain semantic network for topic concept mapping. The domain semantic network maps text words to predefined concept nodes, outputting a preliminary concept set containing "temperature," "sensor," "data aggregation," and "alarm." The preliminary concept set is then merged into synonymous concepts and elevated to a higher level. For example, "alarm" and "alarm" are merged into the same concept, and specific concepts such as "sensor" and "temperature" are elevated to their parent concept "monitoring parameters," thereby extracting the core concept cluster of "monitoring parameters," "data aggregation," and "alarm." A context-aware encoder embeds the core concept cluster into the context, considering the high correlation between "monitoring parameters" and "alarm" in industrial monitoring scenarios, generating context-enhanced concept vectors. Sparse coding techniques are used to reduce and encode the context-enhanced concept vectors, compressing the vector dimension and highlighting discriminative features to generate a unique and representative data semantic fingerprint. This process can be achieved by solving the following optimization problem:

[0070]

[0071] Where: vector Concept vectors and matrices representing context-enhanced concepts For a complete dictionary, vectors The sparse coding coefficients to be determined are the data semantic fingerprints, and the scalars are... The regularization parameter is used to control sparsity.

[0072] Periodic pattern mining is performed on multi-level visit popularity time series to obtain the evolution trajectory of visit popularity. In specific implementation, three observation granularities—minute, hour, and day—are defined on the multi-level visit popularity time series. At different observation granularities, the original visit popularity data, such as the minute-level visit frequency sequence, is truncated into fixed-length windows to form multi-scale time slices. Trend fitting and fluctuation decomposition are performed on the visit popularity data within each multi-scale time slice. Seasonal decomposition methods are applied to separate the time series of the trend component representing the long-term direction of change, the periodic component reflecting regular fluctuations, and the residual component representing random noise. The trend component, periodic component, and residual component corresponding to the minute, hour, and day observation granularities are time-axis aligned and numerically fused to construct the evolution trajectory of visit popularity that can reflect the coordinated changes in visit popularity across the three dimensions of trend, periodicity, and randomness.

[0073] Event intensity analysis is performed on the time series of adjacent update events. In specific implementations, the frequency of update events per unit time is calculated, the distribution density of update events on the time axis is analyzed, and the degree of clustering of update events is evaluated, ultimately obtaining an update event intensity spectrum characterizing the intensity of the update pattern. In some embodiments, it can be understood that the structured reconstructing of the original log sequence can be customized with a parser based on different log source formats. Optionally, the domain semantic network can be constructed based on an industry knowledge graph. It can be understood that the window length of the multi-scale time slice can be adjusted according to the actual data characteristics.

[0074] In one embodiment of the present invention, an interactive evolution layer is established in the feature coupling device. This interactive evolution layer receives data semantic fingerprints, access popularity evolution trajectories, and update event intensity spectra as initial node states. Each feature corresponds to a feature node and is assigned an initial state vector. Dynamic connections between feature nodes are established in the interactive evolution layer. These dynamic connections are generated according to preset interaction rules, which define the weights and conditions for information transmission between different types of feature nodes, simulating the information flow and state interactions between data semantic fingerprints, access popularity evolution trajectories, and update event intensity spectra. During the interactive evolution process, each feature node iteratively updates its own state based on its current state, information from other connected nodes, and the interaction rules. After a preset evolution period, the state changes of each feature node in the interactive evolution layer tend to level off and reach stability. At this point, the stable node state vectors and the node connections formed during the evolution process are output as a graph. Nodes are the vertices of the graph, and connections are the edges. Vertex attributes and edge weights are determined by the final state, thereby forming a comprehensive cached feature graph.

[0075] In practical implementation, an interactive evolution layer is established in the feature coupling device. The interactive evolution layer contains three feature nodes, corresponding to the data semantic fingerprint feature node, the access popularity evolution trajectory feature node, and the update event intensity spectrum feature node, respectively. The initial state vector of the data semantic fingerprint feature node is a 128-dimensional sparse vector, the initial state vector of the access popularity evolution trajectory feature node is a 75-dimensional vector composed of trend, period, and residual components, and the initial state vector of the update event intensity spectrum feature node is a 20-dimensional vector representing the intensity distribution. These initial state vectors together constitute the initial node states of the interactive evolution layer.

[0076] In the interactive evolution layer, dynamic connections are established between feature nodes. These connections are generated based on preset interaction rules, defined as follows: Bidirectional connections exist between data semantic fingerprint feature nodes and access popularity evolution trajectory feature nodes, with connection weights calculated based on the correlation between "alarm" type data semantics and "sudden periodic" access patterns; bidirectional connections exist between access popularity evolution trajectory feature nodes and update event intensity spectrum feature nodes, with connection weights calculated based on the correlation between "high-frequency periodic access" patterns and "low-intensity sparse update" patterns; unidirectional connections exist between data semantic fingerprint feature nodes and update event intensity spectrum feature nodes, with the connection direction pointing from the update event intensity spectrum feature node to the data semantic fingerprint feature node, and the connection weights calculated based on the negative correlation between "intensity spectrum peak value" and "data topic stability". The information flow and state influence among data semantic fingerprints, access popularity evolution trajectories, and update event intensity spectra are simulated. At each evolution time step, each feature node aggregates state information transmitted from all other feature nodes with which it has connections. The amount of information transmitted is determined by the connection weights and the state of the source nodes. After a preset evolution cycle, such as 50 time steps, the rate of change of the state vectors of the data semantic fingerprint feature nodes, access popularity evolution trajectory feature nodes, and update event intensity spectrum feature nodes in the interaction evolution layer is lower than a threshold. This threshold is set according to the accuracy requirements of the cached data feature analysis of the edge computing nodes. The selection rule is as follows: the average rate of change of each dimension element in the feature node state vector is used as the criterion. When the average rate of change of each dimension element in the state vector within 10 consecutive evolution time steps is less than 1%, it is determined that the preset threshold has been reached and the state tends to be stable. At this time, the stable node state vector and the node connection weights formed during the evolution process are output as a graph to form a comprehensive cache feature graph. The comprehensive cache feature graph uses three feature nodes as vertices and dynamic connection relationships as edges. The vertex attributes are described by the final state vector, and the edge weights are described by the final connection weights. In some embodiments, the node state update process can be described by the following formula:

[0077]

[0078] Where: vector Represents feature nodes In evolutionary time step The state, matrix It is a feature node Self-updating weight matrix, set Representation and feature nodes The set of all connected neighbor feature nodes, scalar Indicates at time step From feature nodes To feature nodes Dynamic connection weights, matrix It is the information transfer weight matrix, a function Represents a non-linear activation function. Representing neighbor feature nodes In the evolutionary stage The state vector at each time step. This can be understood as the dynamic connection weights. It can be dynamically calculated during the evolution process based on preset interaction rules and node states.

[0079] See Figure 3In one embodiment of the present invention, feature nodes in the comprehensive cache feature map are used as input layer nodes of the cache value evaluation network, and the feature vector of each node is used as the input value of the input node. A multi-layer value-aware neural network is constructed, wherein each layer of neurons contains a gated recurrent unit module. The feature vector sequence of the input layer nodes is used as the initial temporal state and input to the gated recurrent unit module of the first layer of value-aware neurons. Within the gated recurrent unit module of each layer of neurons, gated recurrent operations are performed: an update gate mechanism is used to selectively update the input features from the previous layer and the hidden state of the previous time step, determining how much information from the previous time step is retained; a reset gate mechanism is used to control the degree of influence of the hidden state of the previous time step on the current candidate state, determining how much past information is used to calculate the new candidate state; the results of the update gate and the reset gate are combined with the current input to generate the new hidden state of the current time step. The new hidden state output by each layer of gated recurrent unit module is transformed through a non-linear activation function, and the transformed result is used as the information aggregation output of that layer. The information from the previous layer is aggregated and output as the input feature sequence for the next layer of value-perceiving neurons. For each neuron in the next layer, gated loop operations and nonlinear transformations are repeatedly executed, iterating layer by layer. Through the progressive transmission and refinement of multiple neurons, feature information is aggregated and abstracted until it reaches the output layer. In the output layer, each original data identifier corresponds to a value-calibrating neuron. This neuron converts the received final aggregated information into a continuous quantized cached value value using a linear transformation or regression function. All data identifiers and their corresponding quantized cached value values ​​are automatically grouped and sorted according to their numerical ranges, generating a hierarchical value list.

[0080] In the specific implementation, feature nodes in the comprehensive cache feature map are used as input layer nodes of the cache value evaluation network. The comprehensive cache feature map contains three feature nodes, and the state vector of each feature node is used as the input value. The data semantic fingerprint feature node corresponds to a 128-dimensional vector, the access popularity evolution trajectory feature node corresponds to a 75-dimensional vector, and the update event intensity spectrum feature node corresponds to a 20-dimensional vector. These three vectors are concatenated to form a 223-dimensional input feature vector, which serves as the input to the input layer nodes and constitutes the initial temporal state. A multi-layer value-aware neural network with a three-layer structure is constructed, where each layer of neurons contains a gated recurrent unit module. The first, second, and third layer value-aware neurons all adopt the same gated recurrent unit structure but have independent network parameters. The feature vector sequence of the input layer nodes is used as the initial temporal state and input into the gated recurrent unit module of the first layer value-aware neuron.

[0081] Within the gated recurrent unit module of each layer of neurons, a gated recurrent operation is performed. An update gating mechanism is used to selectively update the input features from the previous layer and the hidden state from the previous time step. The update gating determines how much information from the previous time step is retained in the current state. A reset gating mechanism is used to control the influence of the hidden state from the previous time step on the current candidate state. The reset gating determines how much past information is used to calculate the new candidate state. The results of the update and reset gating are combined with the current input to generate the new hidden state for the current time step. The new hidden state output from each layer of the gated recurrent unit module is transformed through a nonlinear activation function, using a hyperbolic tangent function. The transformed result is used as the information aggregation output of that layer. The information aggregation output of the previous layer is used as the input feature sequence for the next layer of value-aware neurons. For the next layer of neurons, the gated recurrent operation and the nonlinear transformation process are repeated, iterating layer by layer. The calculation of update gating, reset gating, and candidate states in the gated recurrent operation can be described as follows:

[0082] ;

[0083] ;

[0084] ;

[0085] Where: vector Indicates the first Update the gated output of the layer-gated loop unit, vector Indicates the first Reset gated output of layer-gated loop unit, vector Indicates the first Candidate states of layer-gated recurrent units, vector This represents the input feature sequence from the previous layer, a vector. Indicates the first The hidden state of the previous time step of the layer-gated loop unit, matrix ,matrix ,matrix These are the corresponding trained weight matrices, with symbols... Represents matrix multiplication, symbol This represents the vector concatenation operation, with the symbol... Represents element-wise multiplication, function sum function This is the activation function.

[0086] Through the progressive transmission and refinement of multiple layers of value-perceiving neurons, feature information is gradually aggregated and abstracted. The first layer of value-perceiving neurons has an output dimension of 150, the second layer has an output dimension of 80, the third layer has an output dimension of 40, and so on, until the information is transmitted to the output layer. In the output layer, each original data identifier corresponds to a value-calibrating neuron. The value-calibrating neuron transforms the received final aggregated information—the output vector of the last layer of value-perceiving neurons—into a continuous, quantized, cached value value through a linear transformation function. The form of the linear transformation function is as follows: scalar The output is a vector representing the quantized cache value. The final aggregated information vector, vector For the weight vector, scalar For bias terms, Represents the vector transpose symbol. This represents performing a transpose operation on the weight vector w. All data identifiers and their corresponding quantized cache value values ​​are automatically grouped and sorted according to their numerical range. In one grouping example, data with values ​​greater than 0.8 are assigned to the high-value group, data with values ​​between 0.5 and 0.8 are assigned to the medium-value group, and data with values ​​less than 0.5 are assigned to the low-value group. Within each group, data is arranged in descending order of value, generating a hierarchical value list.

[0087] In one embodiment of the present invention, a hierarchical value list and a resource capacity snapshot obtained in real time from the target edge computing node are input into the cache strategy derivation engine. The resource capacity snapshot includes information such as the current available cache space size and memory utilization. In the cache strategy derivation engine, a balance model between resource constraints and value demands is established. This balance model uses the quantified cache value in the hierarchical value list as the quantified expression of value demands, and the available cache space in the resource capacity snapshot as the boundary condition of resource constraints. The balance model is solved, and an optimization algorithm is used to find the set of cache objects that maximizes the overall cache value demand and the ideal allocation space corresponding to each object under the condition of satisfying the resource constraint boundary. The set of cache objects and the corresponding ideal allocation space are adapted according to the physical characteristics of the edge node's storage medium. Specifically, this includes: identifying the type of storage medium used by the target edge computing node and its read / write performance profile, such as the high random read / write performance of SSDs and the good sequential read / write performance of HDDs. The expected access pattern of each object in the cache object set is analyzed. This expected access pattern is extracted from the access heat evolution trajectory, such as whether random access or sequential access is dominant, and whether it is read-intensive or write-intensive. The expected access patterns are matched with read / write performance profiles to calculate the degree of matching, recommending the most suitable storage medium region for each cached object. For example, objects with frequent random reads are preferentially recommended to the SSD region. Within the recommended most suitable storage medium region, the allocated space size is fine-tuned to align with the physical block size characteristics of the storage medium region, based on the ideal allocation space size and the physical block size characteristics of the storage medium region, to obtain the final allocated physical space size, thus completing the adaptation adjustment. After the above adjustments, the cached object retention order and space pre-allocation plan are generated, which together constitute the adaptive caching strategy.

[0088] In implementation, a hierarchical value list and a resource capacity snapshot obtained in real time from the target edge computing node are input into the cache strategy derivation engine. The hierarchical value list contains data identifiers and corresponding quantified cache value values, while the resource capacity snapshot contains information on the current available cache space size, total memory, storage media type, and remaining capacity. Within the cache strategy derivation engine, a balance model between resource constraints and value demands is established. This model uses the quantified cache value in the hierarchical value list as the quantified expression of value demands and the available cache space in the resource capacity snapshot as the boundary condition for resource constraints. The balance model is solved using a 0-1 knapsack problem optimization algorithm, ensuring that the total allocation space does not exceed the boundary of the available cache space. This calculates the set of cache objects that maximizes the overall cache value demand and the ideal allocation space for each object in the cache object set. The ideal allocation space is a preliminary spatial partition based on the object size and its value ratio.

[0089] The cached object set and its corresponding ideal allocation space are adapted to the physical characteristics of the edge node's storage medium. In specific implementation, the storage medium type and read / write performance profile used by the target edge computing node are identified. The target edge computing node is equipped with both SSD solid-state drives and HDD hard disk drives. The read / write performance profile of the SSD solid-state drives shows low random read latency and high random write speed; the read / write performance profile of the HDD hard disk drives shows high sequential read throughput and high sequential write throughput but poor random access performance. The expected access pattern of each object in the cached object set is analyzed. The expected access pattern is extracted from the access popularity evolution trajectory. The characteristics of the expected access pattern include the degree of randomness of access, read / write ratio, and request size distribution. The matching degree between the expected access pattern and the read / write performance profile is calculated. The matching degree calculation is implemented through a scoring function. The scoring function comprehensively considers the degree of fit between performance indicators such as random read latency and sequential throughput and the object access pattern, and recommends the most suitable storage medium region for each cached object. Refer to Table 1, which shows a list of cached objects and some of their attributes and recommendation results:

[0090] Table 1: List of cached objects, their partial attributes, and recommendation results.

[0091] ;

[0092] Within the recommended optimal storage medium region, the final allocated physical space size is fine-tuned based on the ideal allocation space size and the physical block size characteristics of the storage medium region. For example, for DataID-001 recommended to the SSD high-speed zone, its ideal allocation space is 150MB, while the physical erase block size of the SSD high-speed zone is 256KB. The fine-tuning process aligns 150MB upwards to integer multiples of 256KB, resulting in a final allocated physical space size of 150MB + 106KB (padded to integer multiples of 256KB). After completing the adaptation adjustment, a cache object retention order and space pre-allocation plan are generated. The cache object retention order is arranged in descending order based on the quantified cache value, and locally adjusted based on the performance priority of the storage medium region, such as prioritizing objects on the SSD when values ​​are similar. The space pre-allocation plan explicitly specifies the final allocated physical space size for each cache object and its specific storage medium region, together forming the adaptive caching strategy.

[0093] See Figure 4This is a diagram analyzing cache value and space allocation in an edge computing cache optimization scenario. It clearly presents the quantified value of four cache objects and their corresponding ideal allocation space. DataID-001, with the highest value, is allocated only 150MB of space, while DataID-004, with the lowest value, occupies the largest amount of storage resources. This is precisely the key scenario where caching strategies need optimization. Algorithms such as the 0-1 knapsack problem can be used to reallocate space, maximizing overall cache value under resource constraints. Quantifying cache value directly provides a basis for the retention order of cache objects. High-value objects should be retained first, providing data support for the subsequent cache strategy inference engine to generate adaptive strategies. By comparing the matching degree between value and space, objects with wasted resources can be quickly identified. For example, DataID-004 has the lowest value density and is the primary target for resource reclamation.

[0094] In one embodiment of the present invention, the cache object retention order and space pre-allocation plan in the adaptive caching strategy are analyzed to identify cache entries that need to be added, modified, retained, or removed, and their corresponding specific operations, such as allocating address space for data A, migrating data B to an SSD, and evicting data C. The cache control instruction manual matching the target edge computing node model is consulted to map the identified specific operations to standard underlying cache control commands, such as the ALLOCATE command for allocating space and the MIGRATE command for data migration. According to the execution sequence implied by the cache object retention order, the standard underlying cache control commands are arranged into an instruction sequence with dependencies and execution order. Specifically, this includes: analyzing the cache object retention order to determine the dependency chain between cache objects, where the dependency chain indicates that the removal or migration of certain cache objects is a prerequisite for the space pre-allocation of other cache objects; for example, old data D must be moved out before space can be made for new data E. Based on this dependency chain, the set of prerequisite commands necessary for the execution of each standard underlying cache control command is marked. Based on the set of preceding commands, a directed acyclic graph (DAG) model is used to perform topological sorting of standard underlying cache control commands. Commands are the vertices of the graph, and dependencies are the directed edges, generating a linear instruction execution queue without circular dependencies, which serves as the final cache configuration update instruction set. This instruction set ensures that commands are executed in the correct logical order.

[0095] In a specific implementation, Example 5 of a data caching optimization method in an edge computing environment is implemented as follows: The cache object retention order and space pre-allocation plan in the adaptive caching strategy are analyzed. The cache object retention order indicates the value priority sequence of data objects, and the space pre-allocation plan specifies the target storage area and space size for each data object. The cache entries that need to be added, modified, retained, or removed, and their corresponding specific operations, are identified. For example, the specific operation for data object DataID-005 is "allocate 128MB of space in the SSD high-speed zone," the specific operation for data object DataID-010 is "remove and release 200MB of space from the HDD sequential storage area," the specific operation for data object DataID-002 is "migrate from the current HDD area to the SSD high-speed zone and adjust the space to 256MB," and the specific operation for data object DataID-001 is "retain in the current location and space."

[0096] Consult the cache control instruction manual that matches the target edge computing node model. The cache control instruction manual defines all the underlying commands and their parameter formats that the cache controller of this edge node model can recognize. Map the identified specific operations to standard underlying cache control commands. The mapping process is based on a predefined rule table. For example, the "add space allocation" operation is mapped to the CMD_ALLOC command with the target device identifier, logical address, and size parameters; the "remove and release space" operation is mapped to the CMD_FREE command with the logical address parameter; the "data migration" operation is mapped to the CMD_MOVE command with the source address, destination address, and size parameters; and the "reserve" operation corresponds to no operation command or a status confirmation command CMD_NOP.

[0097] Based on the execution sequence implied by the cache object retention order, standard low-level cache control commands are arranged into a sequence of instructions with dependencies and execution order. In specific implementation, the cache object retention order is analyzed to determine the dependency chains between cache objects. These dependency chains indicate that the removal or migration of certain cache objects is a prerequisite for the pre-allocation of space for other cache objects. For example, the cache object retention order indicates that DataID-010 needs to be removed first to free up space for DataID-002, which needs to be migrated in later. The completion of the migration of DataID-002 is also a prerequisite for the successful allocation of space for DataID-005, thus forming a dependency chain.

[0098] Based on the set of preceding commands, a directed acyclic graph (DAG) model is used to perform topological sorting on the standard underlying cache control commands. Each command is treated as a vertex in the DAG. If command B's set of preceding commands includes command A, a directed edge is constructed from vertex A to vertex B. The constructed DAG is then topologically sorted to generate a linear sequence where all preceding commands of any command in the sequence precede that command. This topological sorting process ensures that the command execution sequence is free of cyclic dependencies, and the sorting result can be formally represented as finding a vertex sequence. This makes it possible for every directed edge in a directed acyclic graph... ,vertex They all appear at the vertices in the sequence. Previously. Among them, variables and variables Representing vertices (i.e., cache control commands) in a directed acyclic graph, variables Represents the number of vertices. A linear instruction execution queue without circular dependencies is generated as the final cache configuration update instruction set. The cache configuration update instruction set is an ordered list of commands. In some embodiments, dependency chain analysis can be extended to consider additional constraints introduced by the wear leveling strategy of the storage medium.

[0099] See Figure 5 This is a multi-dimensional radar chart for evaluating data semantic fingerprints. It quantifies the performance of five cache objects across four dimensions: context relevance, temporal consistency, spatial relevance, and semantic score. It holds significant value in edge computing cache optimization scenarios. As a key output of the "data semantic fingerprint generation" stage, it provides fine-grained semantic feature basis for the subsequent construction of a "comprehensive cache feature map" through multi-dimensional quantitative evaluation. This allows cache value evaluation to move beyond solely relying on access frequency, combining semantic and spatiotemporal features to enhance the comprehensiveness of the evaluation. The performance differences of different objects across these dimensions provide direct evidence for "cache object retention order" and "storage media adaptation." For example, objects with high temporal consistency can be preferentially allocated to high-speed SSD zones, while objects with high semantic scores can be retained as core cache for long-term use.

[0100] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A data caching optimization method in an edge computing environment, characterized in that, include: Raw cached data access logs are collected from the target edge computing node, restructured into a structured form, and a cached data access record set is generated. Semantic distillation, periodic pattern mining, and event intensity analysis are performed on the data topic tags, multi-level access heat time series, and adjacent update event time series in the cached data access record set to obtain data semantic fingerprints, access heat evolution trajectories, and update event intensity spectra. The data semantic fingerprint, the access popularity evolution trajectory, and the update event intensity spectrum are input into the feature coupling device for co-evolution to generate a comprehensive cache feature map; Based on the comprehensive cache feature map, a cache value evaluation network is constructed to dynamically calibrate the value, forming a hierarchical value list that includes data identification and quantified cache value. Based on the hierarchical value list, an adaptive caching strategy for the target edge computing node is generated through a caching strategy derivation engine. The adaptive caching strategy includes the cache object retention order and space pre-allocation plan. The adaptive caching strategy is translated into a device-executable cache configuration update instruction set, and the cache configuration update instruction set is pushed to the cache controller of the target edge node; The process of obtaining the data semantic fingerprint, access popularity evolution trajectory, and updated event intensity spectrum includes: Raw cached data access logs containing data identifiers, content topic descriptions, and access timestamps are collected from the target edge computing node to form a raw log sequence; The original log sequence is restructured in a structured manner to generate a set of cached data access records in a unified format; Perform multiple rounds of semantic distillation on the data topic tags to obtain data semantic fingerprints, including: The data topic tags are input into a preset domain semantic network to perform topic concept mapping and output a preliminary concept set; The initial concept set is then grouped into synonymous concepts and elevated to hierarchical concepts to extract the core concept cluster. The core concept cluster is embedded into a context-aware encoder to generate context-enhanced concept vectors. Sparse coding techniques are used to reduce and encode the context-enhanced concept vectors to generate the data semantic fingerprints that are unique and representative. Perform periodic pattern mining on the multi-level access popularity time series to obtain the access popularity evolution trajectory, including: Multiple observation granularities are defined on the multi-level access popularity time series. The original access popularity data is windowed at different observation granularities to form multi-scale time slices. Trend fitting and fluctuation decomposition are performed on the access popularity data within each multi-scale time slice to separate the time series of trend components, periodic components and residual components. The trend component, periodic component, and residual component corresponding to different observation granularities are aligned and fused respectively to construct the evolution trajectory of the access popularity that reflects the coordinated changes of access popularity in the three dimensions of trend, periodicity, and randomness. The step of co-evolving the data semantic fingerprint, the access popularity evolution trajectory, and the update event intensity spectrum into the feature coupling device to generate a comprehensive cache feature map includes: An interactive evolution layer is established in the feature coupling device. The interactive evolution layer receives the data semantic fingerprint, access popularity evolution trajectory and update event intensity spectrum as the initial node state. In the interactive evolution layer, dynamic connection relationships are established between feature nodes. These dynamic connection relationships simulate the information flow and state influence between the data semantic fingerprint, access popularity evolution trajectory, and update event intensity spectrum based on preset interaction rules. After a preset evolution cycle, the state of each feature node in the interactive evolution layer tends to stabilize. At this time, the stable node states and their connection relationships are output in a graph form, forming the comprehensive cached feature graph.

2. The data caching optimization method in an edge computing environment according to claim 1, characterized in that, The process involves constructing a cache value assessment network based on the comprehensive cache feature map to dynamically calibrate the value, forming a hierarchical value list that includes data identification and quantified cache value, including: The feature nodes in the comprehensive cache feature map are used as the input layer nodes of the cache value evaluation network; A multi-layer value-perceiving neuron is constructed, and the features of the input layer nodes are nonlinearly transformed and aggregated through the multi-layer value-perceiving neuron, and then passed to the output layer layer by layer. In the output layer, each raw data identifier corresponds to a value labeling neuron, which converts the received aggregated information into a continuous quantized cached value. All data identifiers and their corresponding quantized cache value values ​​are automatically grouped and sorted according to the range of the quantized cache value values ​​to generate the hierarchical value list.

3. The data caching optimization method in an edge computing environment according to claim 1, characterized in that, The step of generating an adaptive caching strategy for the target edge computing node based on the hierarchical value list through a caching strategy derivation engine includes: The hierarchical value list, along with a real-time snapshot of the resource capacity obtained from the target edge computing node, is input into the caching strategy derivation engine. In the cache strategy derivation engine, a balance model between resource constraints and value demands is established. The balance model uses the quantified cache value in the hierarchical value list as the quantified expression of value demands, and the available cache space in the resource capacity snapshot as the boundary condition of resource constraints. Solving the balance model yields the set of cache objects that maximize the overall cache value and the corresponding ideal allocation space under the resource constraint boundary conditions. The set of cached objects and their corresponding ideal allocation spaces are adapted to the physical characteristics of the edge node storage medium to generate the cached object retention order and space pre-allocation plan, which together constitute the adaptive caching strategy.

4. The data caching optimization method in an edge computing environment according to claim 3, characterized in that, The step of adapting the cache object set and its corresponding ideal allocation space according to the physical characteristics of the edge node storage medium includes: Identify the type of storage medium used by the target edge computing node and its read / write performance profile; Analyze the expected access pattern of each object in the cached object set, and extract the expected access pattern from the access popularity evolution trajectory; The matching degree between the expected access pattern and the read / write performance feature profile is calculated to recommend the most suitable storage medium region for each cache object. Within the recommended optimal storage medium area, the final allocated physical space size is fine-tuned based on the ideal allocated space size and the physical block size characteristics of the storage medium area, thus completing the adaptation adjustment.

5. The data caching optimization method in an edge computing environment according to claim 1, characterized in that, The step of translating the adaptive caching policy into a device-executable cache configuration update instruction set includes: The cache object retention order and space pre-allocation plan in the adaptive caching strategy are analyzed to identify the cache entries that need to be added, modified, retained or removed and their corresponding specific operations. Consult the cache control instruction manual that matches the target edge computing node model, and map the identified specific operations to standard underlying cache control commands; Based on the execution sequence implied by the retention order of the cached objects, the underlying cache control commands of the standard are arranged into a sequence of instructions with dependencies and execution order, forming the cache configuration update instruction set.

6. The data caching optimization method in an edge computing environment according to claim 5, characterized in that, The step of arranging the standard underlying cache control commands into a sequence of instructions with dependencies and execution order according to the execution timing implied by the retention order of the cached objects includes: The retention order of the cached objects is analyzed to determine the dependency chain between the cached objects. The dependency chain indicates that the removal or migration of some cached objects is a prerequisite for the pre-allocation of space for other cached objects. Based on the dependency chain, label the set of preceding commands for each standard underlying cache control command; Based on the aforementioned set of preceding commands, a directed acyclic graph model is used to perform topological sorting on the underlying cache control commands of the standard, generating a linear instruction execution queue without circular dependencies, which serves as the instruction sequence.

7. The data caching optimization method in an edge computing environment according to claim 2, characterized in that, The construction of multi-layer value-perceiving neurons involves non-linearly transforming and aggregating the features of the input layer nodes through these neurons, and then transmitting them layer by layer to the output layer. Construct a multi-layer value-perceiving neural network with gated recurrent units, wherein each layer of neurons contains a gated recurrent unit module; The features of the input layer nodes are used as the initial temporal state and input into the gated recurrent unit module of the first layer of value-perceiving neurons. Within the gated recurrent unit module of each neuron layer, the following operations are performed: the input features from the previous layer and the hidden state of the previous time step are selectively updated using the update gating mechanism; the hidden state of the previous time step is controlled by the reset gating mechanism to control the degree of influence of the hidden state of the previous time step on the current candidate state; and the new hidden state of the current time step is generated by combining the results. The new hidden state output by each layer of gated loop unit module is transformed by a nonlinear activation function, and the transformed result is used as the information aggregation output of each layer. The information from the previous layer is aggregated and output as the input feature sequence for the value-perceiving neurons in the next layer. For the neurons in the next layer, the gated loop operation and nonlinear transformation process are repeatedly executed, iterating layer by layer until the information is transmitted to the last layer of neurons, thereby realizing the layer-by-layer transmission and refinement of features between layers.

Images