A dynamic cooperative perception method for content deployment optimization

Abstract

The application claims a dynamic cooperative perception content deployment optimization method, belonging to the field of mobile edge network storage and content distribution, specifically comprising: an edge server collects historical traffic data of short video content, uses a GARCH model to extract volatility rate characteristics of the sequence, and inputs the same into a dynamic LSTM model with a fusion attention mechanism together with historical data to accurately predict content popularity at the next moment. On this basis, an edge controller monitors network link status in real time, dynamically calculates a cooperative radius based on network load to determine a cooperation range; then, combined with predicted popularity, neighbor node cache status and access time delay, the cooperative gain of content is calculated, and a hierarchical cooperative deployment strategy is used to store content decisions in local nodes or overflow to neighbor nodes within the cooperative radius. The application reduces service time delay in network congestion and traffic mutation scenarios, and realizes virtual expansion of edge storage capacity and improvement of cache hit rate through a neighbor cooperation mechanism.

Description

Technical Field

[0001] This invention relates to the field of content storage and distribution technology in mobile edge networks, and particularly to a content popularity prediction method that combines volatility characteristics and a dynamic collaborative caching deployment method based on network load awareness. Background Technology

[0002] Mobile edge computing (MEC) significantly alleviates the transmission pressure on the core network by offloading computing and storage resources to the network edge, providing users with a low-latency, high-bandwidth service experience. With the explosive growth of mobile internet content services (such as high-definition video, VR / AR data, and live streaming), edge servers can directly respond to user requests by caching popular content, significantly reducing content retrieval latency. However, edge server storage resources are typically limited, and content request traffic patterns in the network are highly dynamic and bursty. Accurately predicting content popularity trends and developing efficient content deployment strategies are crucial for improving the quality of service (QoS) and cache hit rate of edge networks under limited resources.

[0003] To improve the utilization of edge storage resources, existing research mainly focuses on two questions: "what to cache" and "where to store it." In terms of popularity prediction, current methods primarily rely on statistical models (such as the Zipf distribution) or traditional machine learning models (such as linear regression and ordinary LSTM) to fit historical access volumes. However, edge-side content request traffic data exhibits significant non-stationarity and drastic fluctuations (i.e., volatility clustering). Existing deep learning models often struggle to capture the risk of random disturbances within the data, leading to severe lag or decreased accuracy in predictions during traffic surges. Regarding content deployment, existing collaborative caching strategies typically share resources based on fixed edge server clusters or fixed collaborative hop counts (Radius). This static collaborative mechanism ignores real-time changes in network load: when the network is congested, an excessively large collaborative range can lead to signaling storms and latency degradation; when the network is idle, an excessively small collaborative range limits resource discovery capabilities. Furthermore, traditional deployment strategies often lack awareness of the caching status of neighboring nodes, resulting in adjacent nodes storing a large amount of duplicate popular content, wasting valuable storage resources within the region. In conclusion, how to achieve accurate popularity prediction and flexible collaborative deployment under the dual constraints of drastic traffic fluctuations and dynamic changes in network load has become an urgent problem to be solved.

[0004] Therefore, this invention proposes a dynamic collaborative sensing content deployment optimization method.

[0005] A search revealed application CN110493489A, belonging to the field of edge caching collaboration technology. The method described provides a cluster-based edge collaborative caching approach. Its main steps include: dividing nodes into fixed collaborative clusters based on geographical location or network topology; performing collaborative storage and retrieval within each cluster based on content request characteristics; and searching other nodes within the cluster when a local cache miss occurs. This method achieves a certain degree of resource sharing through cluster management, but its cluster structure is relatively fixed and lacks adaptability to dynamic changes in network load.

[0006] Unlike the methods described above, this invention fully considers the volatility characteristics of traffic data and the dynamic load of edge networks. For predicting content request traffic, existing technologies typically ignore the volatility characteristics of time series data. This invention, however, introduces a generalized autoregressive conditional heteroscedasticity model to extract volatility and integrates it into the attention mechanism LSTM, significantly improving the ability to detect sudden traffic spikes. For cache deployment, existing technologies use a fixed cooperation range, lacking adaptability to network congestion and easily leading to redundant storage. This invention proposes a dynamic cooperation radius mechanism based on network load awareness, which can expand and contract the cooperation range according to real-time link status and introduces a neighbor state factor to calculate the cooperation gain, achieving a hierarchical complementary deployment of "localizing extremely popular content and neighboring less popular content." This invention effectively expands the virtual storage capacity on the edge side and reduces network communication overhead while ensuring a good user experience. Summary of the Invention

[0007] This invention aims to address the problems existing in the prior art and proposes a dynamic collaborative awareness content deployment optimization strategy. The technical solution of this invention is as follows:

[0008] A dynamic collaborative awareness-based content deployment optimization strategy includes the following steps:

[0009] S1. The edge server collects historical traffic data of the content (i.e., the content request volume, which is reflected by the content likes), and uses a generalized autoregressive conditional heteroscedasticity model to extract the volatility characteristics of the sequence.

[0010] S2. The edge controller constructs a dynamic LSTM prediction model with an integrated attention mechanism, which integrates volatility features with historical data as input to predict the popularity of content in the next moment.

[0011] S3. The edge controller monitors the network link status in real time, calculates the network load index, and dynamically calculates the current collaborative radius accordingly;

[0012] S4. The edge controller combines the predicted popularity, the cache status of neighbors within the collaboration radius, and the access latency to calculate the collaboration gain score for each candidate content.

[0013] S5. The edge controller sorts the content based on the collaborative gain score and uses a hierarchical collaborative deployment strategy to store the content on local nodes or spill it to neighboring nodes.

[0014] Furthermore, in step S1, the edge server collects historical traffic data of the content and extracts the volatility characteristics of the sequence using a generalized autoregressive conditional heteroscedasticity model, specifically including:

[0015] S11. The edge server retrieves the historical like count sequence of the content within the past time window and performs normalization processing;

[0016] S12. Perform an autoregressive conditional heteroscedasticity model effect test on the historical series to verify the volatility and clustering characteristics of the data;

[0017] S13. The residual sequence is modeled using a generalized autoregressive conditional heteroscedasticity model, the conditional heteroscedasticity at each moment in the historical sequence is calculated, and a volatility characteristic sequence is generated.

[0018] Furthermore, in step S2, the edge controller constructs a dynamic LSTM prediction model with an attention mechanism to predict the popularity of content at the next time step, specifically including:

[0019] S21. Using the validation set of historical data, a grid search strategy is employed to iteratively optimize the time window size, number of network layers, number of fully connected layers, and number of neurons of the LSTM model to determine the optimal network structure parameters at the current moment.

[0020] S22. Construct a multidimensional feature vector and concatenate and fuse the volatility feature sequence obtained in step S1 with the original historical likes sequence;

[0021] S23. Input the fused feature vector into the LSTM network layer constructed in step S21 to extract long and short-term temporal dependency features;

[0022] S24. Introduce a time series attention mechanism to assign dynamic weights to the hidden layer states of the LSTM output, focus on key fluctuation moments, and output the popularity prediction value for the next moment through a fully connected layer.

[0023] Furthermore, in step S3, the edge controller monitors the network link status in real time and dynamically calculates the current cooperative radius accordingly, specifically including:

[0024] S31. The edge controller collects average round-trip latency data between edge nodes within the region;

[0025] S32. Calculate the network load index at the current moment based on the real-time latency data and the preset upper and lower latency bounds;

[0026] S33. Using a preset dynamic adjustment function or threshold rule, calculate the current dynamic coordination radius R based on the network load index. The higher the network load, the smaller the coordination radius R.

[0027] Furthermore, in step S4, the edge controller calculates the collaborative gain score for each candidate content, specifically including:

[0028] S41. Traverse the candidate content and detect whether there are neighbor nodes that have cached the content within the dynamic collaboration radius R, and obtain the neighbor state factor;

[0029] S42. Calculate the differential latency gain: If the content does not exist in the neighboring node, calculate the latency difference between the cloud and the local node as the high gain; if the content already exists in the neighboring node, calculate the latency difference between the neighboring node and the local node as the low gain.

[0030] S43. Introduce a weighting factor to sum the predicted popularity and differential latency gain to obtain a comprehensive synergistic gain score for the content.

[0031] Furthermore, in step S5, the edge controller adopts a hierarchical collaborative deployment strategy based on the collaborative gain score, specifically including:

[0032] S51. Sort all candidate items from high to low according to their synergistic gain scores and generate a priority queue;

[0033] S52. Perform the first level of local filling: Store high-scoring content to the local edge server in sequence until the local storage space is exhausted;

[0034] S53. Perform the second level of overflow coordination: For content that is not stored locally, if there is no neighbor cache copy within the coordination radius R, push it to the free neighbor node with the largest remaining space within the radius R for storage.

[0035] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the dynamic collaborative sensing content deployment optimization method as described in any one of the claims.

[0036] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a dynamic collaborative awareness content deployment optimization strategy as described in any one of the claims.

[0037] A computer program product includes a computer program that, when executed by a processor, implements the dynamic collaborative awareness content deployment optimization method as described in any one of the claims.

[0038] The advantages and beneficial effects of this invention are as follows:

[0039] 1. In step S2 of this invention, a dynamic LSTM model that integrates volatility features and attention mechanism is constructed. Compared with traditional prediction methods that rely solely on historical values, this invention can effectively quantify the random fluctuation risk of content traffic data and use the attention mechanism to focus on key fluctuation moments, significantly improving the model's prediction robustness and accuracy in non-stationary traffic scenarios.

[0040] 2. Step S3 of this invention proposes a dynamic coordination radius mechanism based on network load awareness, which solves the problem of poor adaptability of traditional fixed coordination ranges. By dynamically expanding and contracting the search range through real-time monitoring of link status, the radius is reduced to decrease signaling overhead when the network is congested, and expanded to improve resource discovery rate when the network is idle, thus achieving a dynamic balance between communication cost and service quality.

[0041] 3. Steps S4 and S5 of this invention propose a differentiated hierarchical deployment strategy based on collaborative awareness. By introducing a neighbor state factor to correct content value, the local storage priority of existing content within the region is proactively reduced, avoiding excessive redundant storage of popular content. This mechanism of "localizing extremely popular content and neighboring less popular content" effectively expands the virtual storage capacity of edge nodes and maximizes the cache hit rate within the region. Attached Figure Description

[0042] Figure 1 This is a flowchart of the preferred embodiment provided by the present invention;

[0043] Figure 2 System architecture diagram;

[0044] Figure 3 The structure diagram of the LSTM prediction model that incorporates volatility features;

[0045] Figure 4 This is a schematic diagram of the edge network topology under dynamic collaborative radius. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0047] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0048] 1. To address the issue of low prediction accuracy caused by the non-stationarity and drastic fluctuations of content traffic data, this paper proposes a volatility-aware deep learning prediction scheme. This scheme utilizes a generalized autoregressive conditional heteroscedasticity model to extract volatility features from historical sequences, quantifies the random risk of traffic data, and integrates it with historical data into a dynamic LSTM model incorporating an attention mechanism. By focusing on key fluctuation moments, it accurately predicts the content popularity of the next moment.

[0049] 2. To address the low collaboration efficiency caused by dynamic changes in edge network load and limited storage resources, this paper proposes a hierarchical collaborative deployment scheme based on dynamic collaboration radius. First, the collaboration radius is dynamically adjusted according to the real-time network load to determine the effective collaboration range. Second, a neighbor state factor is introduced to calculate differentiated collaboration gains, achieving hierarchical deployment of "localized extremely popular content and neighbored secondary popular content," maximizing the cache hit rate within the region.

[0050] like Figure 1 As shown, the present invention provides a dynamic collaborative sensing content deployment optimization method, characterized by comprising the following steps:

[0051] S1. Historical traffic data of the content collected by the edge server is used to extract the volatility characteristics of the sequence using a generalized autoregressive conditional heteroscedasticity model;

[0052] S2. The edge controller constructs a dynamic LSTM prediction model with an integrated attention mechanism, which integrates volatility features with historical data as input to predict the popularity of content in the next moment.

[0053] S3. The edge controller monitors the network link status in real time, calculates the network load index, and dynamically calculates the current collaborative radius accordingly;

[0054] S4. The edge controller combines the predicted popularity, the cache status of neighbors within the collaboration radius, and the access latency to calculate the collaboration gain score for each candidate content.

[0055] S5. The edge controller sorts the content based on the collaborative gain score and uses a hierarchical collaborative deployment strategy to store the content on local nodes or spill it to neighboring nodes.

[0056] To achieve the above process, this invention relies on a specific system physical architecture. For example... Figure 2As shown, the system architecture of this invention is divided into three layers from top to bottom: the cloud data center layer is responsible for the long-term storage and back-to-origin service of massive content; the edge collaboration layer is the core execution layer of this invention, which includes distributed edge servers and logically centralized edge controllers. The edge servers are interconnected through collaborative links, and the edge controller is responsible for executing prediction and deployment decisions; the mobile user layer accesses the nearest edge server through a wireless network to obtain content.

[0057] In this embodiment, step S1 involves the edge server collecting historical traffic data of the content and extracting the volatility characteristics of the sequence using a generalized autoregressive conditional heteroscedasticity model. Specifically, this includes the following steps:

[0058] S11. The edge server retrieves the historical like count sequence of content within a past time window and performs normalization processing. The edge server periodically compiles content statistics. in the past The number of likes within a time step (e.g., the past 24 hours, with each hour being a time step) is denoted as the sequence. To eliminate the difference in data magnitude, the max-min normalization method is used to map the data to... Interval:

[0059]

[0060] S12. Test the effects of the autoregressive conditional heteroscedasticity (ARH) model on the historical data to verify the volatility clustering characteristics of the data. To verify whether the data is suitable for a volatility model, first, a stationarity test (ADF test) is performed on the normalized series. Subsequently, the Lagrange multiplier method (LM test) is used to test the effects of the ASH model on the residual series. If the test results are significant (i.e., volatility clustering exists), it indicates that the series is suitable for modeling using the ASH model.

[0061] S13. A generalized autoregressive conditional heteroscedasticity (GARCH) model is used to model the residual sequence, calculating the conditional heteroscedasticity at each time step in the historical sequence to generate a volatility feature sequence. This paper uses the GARCH(1,1) model to extract the volatility features of the content. Assuming the residuals of the sequence... If it follows a normal distribution, then at time... The conditional variance (i.e., the square of volatility). It can be represented as:

[0062]

[0063] in, These are the model parameters. After solving for these parameters using the maximum likelihood estimation method, the conditional volatility at each time step can be calculated. The final generation is a volatility characteristic sequence. This sequence reflects the uncertainty risk of content popularity at different times.

[0064] In this embodiment, in step S2, the edge controller constructs a dynamic LSTM prediction model with an attention fusion mechanism to predict the popularity of content at the next time step, specifically including the following steps:

[0065] S21. Using a validation set of historical data, a grid search strategy is employed to iteratively optimize the time window size, number of network layers, number of fully connected layers, and number of neurons in the LSTM model to determine the optimal network structure parameters at the current moment. To enable the prediction model to adapt to the fluctuating characteristics of traffic data over different time periods, this invention does not use a fixed network structure. In specific implementation:

[0066] 1. Set the parameter space: Set the time window size of the LSTM model. ( Network layers ( ), number of fully connected layers ( and the number of neurons in the hidden layer ( These are the four key hyperparameters to be optimized.

[0067] 2. Iterative Training: Utilizing the validation set portion of the historical dataset, a grid search strategy is employed to traverse all combinations of the above parameters. For each set of parameters... A temporary LSTM model is constructed for rapid training.

[0068] 3. Determine the optimal structure: Calculate the root mean square error (RMSE) of the model on the validation set for each set of parameters. Select the set of parameters with the smallest RMSE as the optimal network structure parameters for the current time step, and construct the final dynamic LSTM model for prediction based on this.

[0069] S22. Construct a multi-dimensional feature vector by concatenating and fusing the volatility feature sequence obtained in step S1 with the original historical likes sequence. To enable the prediction model to simultaneously perceive both "historical trends" and "volatility risk," the normalized historical data obtained in step S1 is used... With volatility characteristics Perform feature concatenation to construct the timeline. Multidimensional input vector :

[0070]

[0071] The total input sequence of the model is .

[0072] S23. The fused feature vector is input into the LSTM network layer to extract long-short-term temporal dependencies. The LSTM unit is used to process long-sequence data, overcoming the gradient vanishing problem of traditional RNNs. The LSTM unit uses a forgetting gate. Input gate and output gate Controlling the flow of information. (Time) Hidden layer state From the state of the previous moment and current input Joint decision:

[0073]

[0074] After processing by the LSTM layer, a hidden state sequence containing temporal features is obtained. .

[0075] S24. A time-series attention mechanism is introduced, dynamically assigning weights to the hidden layer states of the LSTM output to focus on key fluctuation moments. The predicted popularity value for the next time step is then output through a fully connected layer. To enhance the model's focus on highly volatile moments, an attention mechanism is introduced. First, the popularity prediction value for each time step is calculated. Attention score This indicates the importance of that moment to the future prediction outcome:

[0076]

[0077] The scores are normalized using the Softmax function to obtain the weight coefficients. :

[0078]

[0079] The hidden states are weighted and summed based on their weight coefficients to obtain the final context vector. :

[0080]

[0081] Finally, the context vector Input a fully connected (Dense Layer) and output the predicted content popularity value for the next time step. By introducing volatility characteristics and an attention mechanism, this model can handle situations with drastic flow fluctuations ( (A larger value) will automatically be assigned a higher weight at that moment, thereby solving the problem of prediction lag.

[0082] To facilitate understanding of the specific calculation process for popularity prediction in steps S1 to S2 of this invention, a practical data example will be used for detailed explanation below. Assume the edge server is processing data numbered... To predict the popularity of content, a time window is set. (i.e., data from the past 5 hours), and the original likes sequence for this content over the past 5 time points is as follows: It can be seen that the sequence at time 4 ( It experienced explosive growth (300), exhibiting drastic fluctuations.

[0083] 1. Example of data normalization processing (corresponding to step S11): First, obtain the minimum value of the sequence. and maximum value Normalization calculation is performed according to formula (1):

[0084]

[0085]

[0086]

[0087]

[0088]

[0089] The final normalized history sequence is .

[0090] 2. Example of GARCH volatility feature extraction (corresponding to step S13) Assuming that the parameters obtained by fitting the GARCH(1,1) model through long-term historical training are: constant term ARCH term coefficient GARCH term coefficient Assuming Squared residuals at time step Conditional variance Now calculate Volatility at a given moment (i.e., the moment of a traffic surge): According to formula (2): .but Volatility eigenvalue at time Similarly, assuming the calculated complete volatility characteristic sequence is shown in the table below:

[0091]

[0092] 3. Example of multidimensional feature fusion and attention weight allocation (corresponding to steps S21-S23): Constructing a multidimensional input vector .For example The input vector at time t is Inputting this sequence into an LSTM network yields the hidden layer state sequence. At this point, an attention mechanism is introduced. Because Not only are the values ​​high at certain times ( ), and high volatility ( The score calculated by the attention mechanism The attention weight distribution will be significantly higher than at other times. Assume the attention weight distribution calculated using formulas (5) and (6) is... ) as follows:

[0093]

[0094] Analysis: As shown in the table above, because this invention introduces volatility characteristics, the model successfully identified... It is a key turning point that determines future trends, and therefore it is given a weight as high as 0.60 (Attention Weight).

[0095] Final context vector will be mainly by leading:

[0096] ;

[0097] 4. Final prediction result: After processing by the fully connected layer, the model outputs the normalized prediction value for the next time step ($t=6$). Inverse normalization reduction: This indicates that, despite The instantaneous popularity dropped to 120, but the model took into account various factors. The high burst and high volatility inertia predict that the heat will rebound in the next moment (179), rather than simply declining linearly. This result is more in line with the actual law of "burst-fall-rebound (long tail)" of the flow than the traditional model.

[0098] In this embodiment, in step S3, the edge controller monitors the network link status in real time and dynamically calculates the current cooperation radius accordingly, specifically including the following steps:

[0099] S31. The edge controller collects average round-trip latency data between edge nodes within the region. The edge controller periodically (e.g., every 5 minutes) sends lightweight probe packets to edge servers within the region, or collects transmission logs of service messages to calculate the current time. Average round-trip latency between edge nodes To eliminate transient jitter, an exponentially weighted moving average (EWMA) is used to smooth the latency data.

[0100] S32. Based on real-time latency data and preset latency upper and lower bounds, calculate the network load index at the current moment and set the minimum transmission latency under ideal network conditions as follows. (Depending on physical distance and link bandwidth), and the maximum tolerable latency allowed by the service is (Based on QoS standards, such as 100ms). Current network load index. Defined as the normalized position of the delay within the tolerance interval:

[0101]

[0102] in, The range of values ​​is truncated to between. The closer the value is to 1, the more congested the network is and the higher the cost of coordinated transmission.

[0103] S33. Using a preset dynamic adjustment function or threshold rule, calculate the current dynamic coordination radius R based on the network load index. The higher the network load, the smaller the coordination radius R. Considering the small-world characteristics of edge networks and multi-hop latency constraints, set a maximum coordination radius. (Skip). A smooth decay model based on a variant of the Sigmoid function is used to calculate the dynamic radius. :

[0104]

[0105] In the formula, Indicates rounding down; The sensitivity coefficient (e.g., 10) controls the response speed of the radius as the load changes. Set the load threshold center (e.g., 0.5).

[0106] This formula achieves adaptive control: when the network is idle ( At a lower level, Maintain at To maximize the scope of resource search; when the network is congested ( When it is relatively high, It rapidly decays to 1 or 0 to limit the range of collaborative communication and prevent network crashes.

[0107] In this embodiment, step S4, in which the edge controller calculates the collaborative gain score for each candidate content, specifically includes the following steps:

[0108] S41. Traverse the candidate content, detect whether there are neighbor nodes that have cached the content within the dynamic collaboration radius R, and obtain the neighbor state factor for each content to be deployed. The edge controller queries the global metadata table to check the area centered on the current node with a radius of [missing information]. Within the logical topology range, are there any other edge nodes that have already stored content? Define the neighbor state factor. :

[0109]

[0110] S42. Calculate the differential latency gain: If the content does not exist on neighboring nodes, calculate the latency difference between the cloud and the local machine as the high gain; if the content already exists on neighboring nodes, calculate the latency difference between the neighboring nodes and the local machine as the low gain. These represent the average latency for retrieving content from local storage, neighboring storage, and the cloud, respectively. Content Latency gain from local storage The calculation is as follows:

[0111]

[0112] This formula introduces a "cooperative penalty" mechanism: when At that time, the gain was only The smaller value reduces the priority of this content in local storage, thus avoiding duplicate storage of popular content with neighboring stores.

[0113] S43. Introduce a weighting factor to sum the predicted popularity and differential latency gain to obtain a comprehensive collaborative gain score for the content. This score is then used to calculate the popularity predicted in step S2. and the time delay gain obtained in step S42 Normalization was performed separately for each part. A comprehensive score was then calculated. :

[0114]

[0115] in, and As a weighting factor, and In this embodiment, a sensitivity analysis experiment was conducted to set... To achieve the best balance between popularity and latency gain.

[0116] In this embodiment, step S5 involves the edge controller employing a hierarchical collaborative deployment strategy based on the collaborative gain score, specifically including the following steps:

[0117] S51. Sort all candidate content according to their synergistic gain scores from highest to lowest, and generate a priority queue for the set of all calculated content scores. Sort the data in descending order to generate a deployment priority queue. The higher the score, the higher the "popularity" and the stronger the "scarcity" within the current collaboration domain.

[0118] S52. Perform the first level of local filling: Store high-rated content to the local edge server in sequence until the local storage space is exhausted. Traverse the priority queue and store the top-ranked content to the local edge server in sequence. Since the rating of content already existing in the neighboring area is lowered, this step prioritizes storing "extremely hot" content that the neighbor does not have, ensuring the fastest response for core content.

[0119] S53. Perform second-level overflow coordination: For content not stored locally, if there is no neighbor cache copy within the coordination radius R, push it to the free neighbor node with the largest remaining space within radius R for storage. For "second-hot content" remaining after the local storage space is full, perform overflow judgment:

[0120] 1. Check the status of its neighbors again. .

[0121] 2. If If a neighbor already has a copy, then skip the storage (use the existing neighbor copy to provide services and achieve deduplication).

[0122] 3. If (Neither neighbors nor other neighbors are present), and currently... Then in the radius The system finds the neighbor node with the largest remaining storage space within the specified range and pushes the content to that neighbor node for storage. This step virtually expands edge storage resources, ensuring that even frequently accessed content can be found at the edge, reducing the number of times the data needs to be retrieved from the origin cloud.

[0123] The following example illustrates the calculation process and effect of this strategy in steps S3 to S5: Assume there are 3 items to be deployed in the current edge network (A, B, C), and their normalized popularity predicted in step S2 is as follows: (Extremely hot) (Secondary heat) (Normal). Set delay parameters: , , Set weighting factors: , The local edge server has only enough remaining capacity to store one piece of content.

[0124] Scenario 1: Network Idle State (Low Load)

[0125] 1. S3 calculation radius: detected Calculate According to formula (9), (Large scope of collaboration).

[0126] 2. S4 Probe Status: Probe within a 3-hop radius and find that neighboring nodes have cached content A, but not B and C. .

[0127] 3. S4 calculates the gain and score:

[0128] (1) Content A: Score after normalization .

[0129] (2) Content B: Score after normalization .

[0130] (3) Content C: Score after normalization .

[0131] 4. S5 Tiered Deployment:

[0132] (1) Sort: .

[0133] (2) Local filling: Store content B (highest score and scarcest). Local space is full.

[0134] (3) Overflow coordination:

[0135] ①Inspection Item C: (No neighbors), and Execute the push to neighboring nodes.

[0136] ②Inspection Item A: (Neighbor has it). Skip and don't save.

[0137] (4) Results: B is stored locally, C is stored in the neighboring area, and A directly uses the neighbor's data. All requests are hit at the edge and there is no redundancy.

[0138] Scenario 2: Network congestion (High Load)

[0139] 1. S3 calculation radius: detected (Severe queuing), calculated According to formula (9), (Disable collaboration).

[0140] 2. S4 detection status: due to This is considered as having no neighbors available. .

[0141] 3. S4 calculates the gain and score:

[0142] (1) All content All (High gain).

[0143] (2) At this point, the score depends entirely on popularity: .

[0144] 4. S5 Tiered Deployment:

[0145] (1) Local filling: Store content A (most popular). Local space is full.

[0146] (2) Overflow collaboration: because No push notifications will be sent.

[0147] (3) Result: Only A is stored locally. When the network is congested, the local experience of the most popular content is prioritized, avoiding the burden on the network caused by collaborative communication.

[0148] This invention addresses the problem of low popularity prediction accuracy caused by the non-stationarity and drastic fluctuations in content access traffic in mobile edge networks by proposing a volatility-aware deep learning prediction scheme. This scheme utilizes a generalized autoregressive conditional heteroscedasticity model to extract volatility features from historical sequences, quantifying the random risk of traffic data. This data is then fused with historical data and input into a dynamic LSTM model incorporating an attention mechanism to accurately predict content popularity at the next moment. Compared to traditional prediction methods that rely solely on historical values, this invention, by introducing volatility features and an attention mechanism, effectively captures key turning points in sudden traffic surges, solving the prediction lag problem. Simultaneously, a grid search strategy is used to dynamically determine the network structure parameters of the LSTM, enabling it to adapt to traffic fluctuations across different time periods compared to fixed-structure models, significantly improving prediction robustness and accuracy. Furthermore, to address the low collaborative efficiency caused by dynamic changes in edge network load and limited storage resources, this invention proposes a hierarchical collaborative deployment strategy based on a dynamic collaborative radius. In optimizing the collaborative range, the collaborative radius is dynamically expanded and contracted based on real-time network link load awareness, solving the problems of high signaling overhead during network congestion and low resource discovery rate during network idle periods associated with fixed collaborative ranges. In terms of content deployment, a neighbor state factor is introduced to calculate differentiated collaborative gains, and a comprehensive value scoring index is constructed to determine the optimal storage location for content. This strategy achieves complementary storage of "localized extremely popular content and neighbored content," effectively avoiding excessive redundancy of popular content within the region. While ensuring a low-latency service experience for users, it expands the virtual storage capacity of edge nodes and maximizes the overall cache hit rate.

[0149] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0150] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0151] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0152] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A method for optimizing content deployment based on dynamic collaborative sensing, characterized in that, Includes the following steps: S1. The feature data of the content collected by the edge server is subjected to stationarity test and preprocessing, and the volatility features of the sequence are extracted using a generalized autoregressive conditional heteroscedasticity model. S2. The edge controller constructs a dynamic LSTM prediction model with an integrated attention mechanism, which integrates volatility features with historical data as input to predict the popularity of content in the next moment. S3. The edge controller monitors the network link status in real time, calculates the network load index, and dynamically calculates the current collaborative radius accordingly; S4. The edge controller combines the predicted popularity, the cache status of neighbors within the collaboration radius, and the access latency to calculate the collaboration gain score for each candidate content. S5. The edge controller sorts the content based on the collaborative gain score and uses a hierarchical collaborative deployment strategy to store the content on local nodes or spill it to neighboring nodes.

2. The method for optimizing dynamic collaborative sensing content deployment according to claim 1, characterized in that, In step S1, the edge server collects historical sequences of content features and extracts the volatility features of the sequences using a generalized autoregressive conditional heteroscedasticity model, specifically including: S11. The edge server obtains the historical feature sequence of the content within the past time window and performs normalization processing; S12. Perform an autoregressive conditional heteroscedasticity model effect test on the historical series to verify the volatility and clustering characteristics of the data; S13. The residual sequence is modeled using a generalized autoregressive conditional heteroscedasticity model, and the conditional heteroscedasticity at each moment in the historical sequence is calculated. This means that the volatility intensity in the historical sequence is not constant, but evolves dynamically over time and exhibits the characteristic of "volatility clustering". It accurately quantifies the dynamic uncertainty and sudden risks of flow at different moments and generates a volatility characteristic sequence.

3. The method for optimizing dynamic collaborative sensing content deployment according to claim 1, characterized in that, In step S2, the edge controller constructs a dynamic LSTM prediction model that incorporates an attention mechanism to predict the popularity of content at the next time step. The specific model structure includes an input layer, an LSTM layer, an attention layer, and an output layer, specifically including: S21. Using the validation set of historical data, a grid search strategy is used to iteratively optimize the time window size, number of network layers, number of fully connected layers, and number of neurons of the LSTM model to determine the optimal network structure parameters at the current moment. S22. Construct a multidimensional feature vector and concatenate and fuse the volatility feature sequence obtained in step S1 with the original historical likes sequence; S23. Input the fused feature vector into the LSTM network layer constructed in step S21 to extract long and short-term temporal dependency features; S24. Introduce a time series attention mechanism to assign dynamic weights to the hidden layer states of the LSTM output, focus on key fluctuation moments, and output the popularity prediction value for the next moment through a fully connected layer.

4. The method for optimizing dynamic collaborative sensing content deployment according to claim 1, characterized in that, In step S3, the edge controller monitors the network link status in real time and dynamically calculates the current cooperative radius accordingly, specifically including: S31. The edge controller collects average round-trip latency data between edge nodes within the region; S32. Calculate the network load index at the current moment based on the real-time latency data and the preset upper and lower latency bounds; S33. Using a preset dynamic adjustment function or threshold rule, calculate the current dynamic coordination radius R based on the network load index. The higher the network load, the smaller the coordination radius R.

5. The method for optimizing dynamic collaborative sensing content deployment according to claim 1, characterized in that, In step S4, the edge controller calculates the collaborative gain score for each candidate content, specifically including: S41. Traverse the candidate content and detect whether there are neighbor nodes that have cached the content within the dynamic collaboration radius R, and obtain the neighbor state factor; S42. Calculate the differential latency gain: If the content does not exist in the neighboring node, calculate the latency difference between the cloud and the local node as the high gain; if the content already exists in the neighboring node, calculate the latency difference between the neighboring node and the local node as the low gain. S43. Introduce a weighting factor to sum the predicted popularity and differential latency gain to obtain a comprehensive synergistic gain score for the content.

6. The content deployment optimization strategy for dynamic collaborative perception according to claim 1, characterized in that, In step S5, the edge controller adopts a hierarchical collaborative deployment strategy based on the collaborative gain score, specifically including: S51. Sort all candidate items from high to low according to their synergistic gain scores and generate a priority queue; S52. Perform the first level of local filling: Store high-scoring content to the local edge server in sequence until the local storage space is exhausted; S53. Perform the second level of overflow coordination: For content that is not stored locally, if there is no neighbor cache copy within the coordination radius R, push it to the free neighbor node with the largest remaining space within the radius R for storage.

7. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the dynamic collaborative awareness content deployment optimization method as described in any one of claims 1 to 6.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the dynamic collaborative awareness content deployment optimization method as described in any one of claims 1 to 6.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the dynamic collaborative awareness content deployment optimization method as described in any one of claims 1 to 6.

Images