A data storage method based on network technology

Abstract

This invention relates to the field of distributed storage technology, specifically a data storage method based on network technology. The method includes: collecting multi-dimensional vehicle data and constructing a T-pattern prediction tree historical trajectory database; using the source edge node currently providing services to the vehicle as the center, predicting the access probability of the vehicle's next target node using a pre-trained trajectory prediction model, and determining the preferred target edge node and candidate target edge nodes; calculating the pre-migration time window, and initiating data pre-transmission from the source node to the target edge node when the triggering conditions are met; using probability-driven erasure coding technology to allocate data and check blocks to each target node to achieve low-overhead fault-tolerant storage; and using a distributed consistency protocol based on multi-version concurrency control for incremental data synchronization, and performing consistency comparison when the vehicle accesses a new node. This invention eliminates cold start latency during node switching, ensuring efficient and seamless data flow.

Description

Technical Field

[0001] This invention relates to the field of distributed storage technology, specifically a data storage method based on network technology. Background Technology

[0002] Vehicle-to-everything (V2X) technology has become a key infrastructure for realizing autonomous driving, augmented reality navigation, vehicle cooperative control, and real-time processing of massive amounts of sensor data. Modern intelligent connected vehicles are equipped with a large number of environmental perception devices, such as high-resolution cameras, LiDAR, and millimeter-wave radar, which continuously generate and consume massive amounts of data during operation. This data includes not only in-vehicle entertainment content and high-definition video streams, but also real-time road condition perception information, vehicle control signaling, and high-precision map data that are crucial for driving safety.

[0003] Traditional centralized cloud computing models, due to long physical transmission distances and numerous network routing nodes, often result in high round-trip latency, failing to meet the reliable, low-latency communication requirements of tasks such as autonomous driving in vehicle-to-everything (V2X) applications. To overcome the inherent latency bottleneck of cloud computing, mobile edge computing architecture has emerged. By deploying edge servers on the access network side (such as Roadside Units, RSUs) close to the vehicle terminal, vehicles can offload computationally intensive or latency-sensitive tasks to network edge nodes, significantly reducing end-to-end communication latency and effectively alleviating bandwidth pressure on the core backbone network. However, the high-speed mobility of vehicles means their dwell time within the coverage area of ​​a single RSU is often extremely short (typically only tens of seconds to a few minutes). As vehicles move, they inevitably shuttle frequently between the coverage areas of different RSUs, leading to extremely frequent edge node switching or task migration. Existing edge node switching mechanisms mainly suffer from the following problems:

[0004] First, traditional edge data and task migration solutions typically only begin retrieving vehicle status data and computational context from the original node or cloud center after the vehicle has completely moved out of the original RSU's coverage area, disconnected, and completed a handshake with the new RSU. This approach leads to severe "cold start" delays for the new access node. During the cold start, because the data is not yet available, the vehicle's service requests must be suspended or rolled back to the cloud for processing, exposing the vehicle to the risk of service interruption, loss of critical data, and control task failure during the switchover.

[0005] Second, the storage and computing resources of edge nodes are typically extremely limited. Blindly multicasting or fully replicating data to all potentially adjacent nodes would rapidly deplete the already scarce storage resources of the edge network, while simultaneously causing severe link congestion. Balancing data migration redundancy with storage costs based on high-precision trajectory prediction is a highly challenging multidimensional optimization problem. Summary of the Invention

[0006] To address the problems in existing technologies, such as frequent interruptions of edge computing services due to vehicle movement, high startup latency and data loss caused by passive data migration, and low utilization of collaborative storage resources and lack of state consistency of edge nodes, this invention provides a data storage method based on network technology.

[0007] A data storage method based on network technology includes the following steps;

[0008] S1. Collect multi-dimensional perception and positioning information of connected vehicles in real time through edge nodes, and build a vehicle driving history trajectory database based on historical vehicle trajectory data during the system initialization phase;

[0009] S2. Centered on the source edge node currently providing services to the vehicle, use a pre-trained trajectory prediction model to predict the vehicle's direction of movement and trajectory in real time, output the probability distribution of the target edge nodes that the vehicle may connect to in the future, and determine the preferred target edge node and candidate target edge nodes accordingly.

[0010] S3. Based on the trajectory prediction results, the current task data volume of the vehicle, the queue length, and the available bandwidth of the backhaul network, dynamically calculate the pre-migration time window, and send the data pre-transmission from the origin node to the target edge node when the triggering conditions are met.

[0011] S4. Encode the full dataset of vehicles to be pre-migrated into several data blocks and verification blocks, and distribute the data blocks and verification blocks to the preferred target edge node and the secondary target edge node according to the probability distribution of the target edge node.

[0012] S5. During the window period for performing data pre-migration and vehicle cross-node physical switching, if the vehicle continuously generates incremental data at the source edge node, a distributed consistency protocol based on multi-version concurrency control is used to synchronize the incremental data, and a consistency comparison is performed when the vehicle accesses a new node.

[0013] Preferably, the construction of the vehicle driving history trajectory database specifically includes: extracting vehicle local frequent itemsets and vehicle frequent trajectory patterns from historical vehicle trajectory data, and constructing them into a multi-branch state transition T-pattern prediction tree; each edge node only stores the local subtree related to its own physical coverage area and adjacent areas, and performs dynamic incremental updates based on the feedback from the actual entering node.

[0014] Preferably, the step of using a pre-trained trajectory prediction model to predict the vehicle's direction of movement and trajectory in real time, and outputting the probability distribution of target edge nodes that the vehicle may connect to in the future, specifically includes:

[0015] The system acquires real-time bird's-eye view images captured by the vehicle's onboard camera, processes them using a semantic segmentation model, and obtains the bounding box coordinate features and semantic segmentation masks of entity targets within the scene.

[0016] The processing results are processed using a deep neural network to extract features, and a spatial graph representation of the scene is constructed based on the K-nearest neighbor algorithm. Nodes are the extracted entity features, and edges represent the spatial adjacency relationships between entities. Subsequently, a graph convolutional network is used to perform node state aggregation calculations on the spatial graph representation to extract high-order spatial features containing road network physical topology constraints.

[0017] The historical time step state sequence of the vehicle is obtained, and the sequence is filtered and preprocessed using state smoothing and data augmentation techniques.

[0018] The extracted spatial features and the processed historical time step state sequence are input into a long short-term memory network to analyze and process the time evolution sequence features of vehicle dynamic high-frequency displacement.

[0019] The spatiotemporal attention mechanism is used to learn the importance weights of the interaction between vehicles and the traffic environment in different spatiotemporal dimensions. The above features are combined and the normalized probability distribution vector of each target edge node that the vehicle may connect to in the future is output by a classifier.

[0020] Preferably, determining the preferred target edge node and candidate target edge nodes specifically includes:

[0021] Obtain the normalized probability distribution vector of the trajectory prediction model, which covers all adjacent candidate edge nodes;

[0022] Select the candidate edge node with the maximum predicted probability value from the probability distribution vector and determine it as the preferred target edge node;

[0023] Among the remaining candidate edge nodes other than the primary target edge node, the candidate edge nodes with a predicted probability greater than a preset probability truncation threshold are included in the secondary target edge node set.

[0024] Candidate edge nodes whose predicted probability is lower than the preset probability truncation threshold are removed from the candidate set.

[0025] Preferably, the step of dynamically calculating the pre-migration time window based on trajectory prediction results, the current task data volume of the vehicle, queue length, and available bandwidth of the backhaul network, and initiating data pre-transmission from the origin node to the target edge node when the triggering conditions are met, specifically includes:

[0026] The source node obtains vehicle location information, instantaneous speed, and road constraint based on the trajectory prediction model, and calculates the estimated remaining time for the vehicle to reach the wireless signal coverage boundary of the target edge node, which is denoted as the remaining dwell time. ;

[0027] Calculate the total latency required to completely deliver the vehicle's mission data to the target edge node, the total latency including communication link transmission latency. target node cache warm-up latency Superposition; wherein the transmission delay satisfies:

[0028] ;

[0029] In the formula, The total capacity of the data to be migrated. This represents the real-time available bandwidth of the current communication link. For routing queuing and physical propagation delay;

[0030] Introducing a safety redundancy time factor to resist network jitter and queuing delays And continuously evaluate the inequalities that trigger the judgment:

[0031] ;

[0032] Once the aforementioned trigger inequality is met, the source node initiates the data pre-migration and push process to the target edge node.

[0033] Preferably, the step of distributing the data blocks and verification blocks to the preferred target edge node and the secondary target edge node according to the probability distribution of the target edge nodes specifically includes: sending all native direct data blocks generated based on the erasure coding algorithm to the preferred target edge node corresponding to the highest probability; distributing the generated redundant verification blocks to the set of secondary target edge nodes whose probabilities are greater than a set threshold according to their relative probability ratios; if a vehicle deviates from the prediction and connects to any secondary target edge node, the node uses the locally stored redundant verification blocks to request the missing fragments from neighboring nodes and performs matrix inverse operation to restore the full original data.

[0034] Preferably, the distributed consistency protocol based on multi-version concurrency control specifically includes: the source node multicasts a lightweight incremental patch carrying incremental data to the target node cluster that has received the pre-migration data; the target node performs incremental aggregation and splicing application operations; after the vehicle completes node switching, the source node sends the maintained session state data hash digest to the new node, and the new node performs cryptographic confirmation comparison with the hash value of its local data. If they match, the new node seamlessly takes over; if there is a conflict, it triggers interruption protection and differential synchronization retrieval.

[0035] Compared with the prior art, the advantages of this invention are:

[0036] This invention constructs a high-precision spatiotemporal trajectory prediction model by integrating TPPT frequent pattern mining and LSTM-GCN deep learning network. Based on this, the implemented "storage pre-migration" can accurately pinpoint the next edge node connected to the vehicle, effectively eliminating the "cold start" latency during edge service switching and achieving seamless and imperceptible switching of vehicle tasks between edge servers.

[0037] To address the inherent uncertainties and prediction errors in vehicle trajectories, the costly full-scale multi-point replication method was abandoned. By implementing reasonable data sharding and redundant configuration of check blocks between primary and secondary nodes, the risks of long-tail latency and data failure caused by prediction failures were prevented, while minimizing unnecessary storage consumption and backhaul bandwidth compression in the edge network. This achieved optimal edge collaborative fault tolerance and resource orchestration.

[0038] By using multi-version concurrency control based on vector clocks and hash consistency verification, the technical blind spot of asynchronous pre-migration data states between new and old edge nodes is solved when vehicles are moving at high speeds and on-board sensors are continuously collecting and updating data. Attached Figure Description

[0039] Figure 1 This is a flowchart of a data storage method based on network technology proposed in this invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0041] refer to Figure 1 This scheme is a data storage method based on network technology, which includes the following steps;

[0042] S1. Collect multi-dimensional perception and positioning information of connected vehicles in real time through edge nodes, and build a vehicle driving history trajectory database based on historical vehicle trajectory data during the system initialization phase.

[0043] S2. Centered on the source edge node currently providing services to the vehicle, use a pre-trained trajectory prediction model to predict the vehicle's direction of movement and trajectory in real time, output the probability distribution of target edge nodes that the vehicle may connect to in the future, and determine the preferred target edge node and candidate target edge nodes accordingly.

[0044] S3. Based on the trajectory prediction results, the current task data volume of the vehicle, the queue length, and the available bandwidth of the backhaul network, dynamically calculate the pre-migration time window, and initiate data pre-transmission from the origin node to the target edge node when the triggering conditions are met.

[0045] S4. Encode the full dataset of vehicles to be pre-migrated into several data blocks and verification blocks, and distribute the data blocks and verification blocks to the preferred target edge node and the secondary target edge node according to the probability distribution of the target edge node, so as to realize distributed collaborative fault-tolerant storage.

[0046] S5. During the window period for performing data pre-migration and vehicle cross-node physical switching, if the vehicle continuously generates incremental data at the source edge node, a distributed consistency protocol based on multi-version concurrency control is used to synchronize the incremental data, and a consistency comparison is performed when the vehicle accesses a new node.

[0047] Example 1:

[0048] Traditional vehicle-to-everything (V2X) networks are often limited by flat or centralized network architectures, resulting in low data collection efficiency and a lack of structured management of historical trajectories. This embodiment builds a three-layer architecture of "perception-edge-global" to provide real-time and highly structured high-value historical data sources for subsequent spatiotemporal trajectory prediction under limited edge storage conditions. The specific system architecture includes:

[0049] The perception and terminal execution layer primarily consists of a massive number of connected vehicles equipped with advanced onboard units and various sensors (LiDAR, cameras, millimeter-wave radar, etc.). Vehicles act not only as consumers of data but also as producers of large-scale spatiotemporal data. Vehicles communicate with surrounding RSUs via V2I (Vehicle-to-Infrastructure) protocols or 5G NR-V2X interfaces, periodically reporting their basic kinematic state set. This set of data includes geographic coordinate sequences, velocity vectors, instantaneous acceleration, heading angles, and timestamps accurate to milliseconds.

[0050] Distributed Edge Collaboration Layer: This layer consists of multiple RSUs (Roadside Units) and their mounted Mobile Edge Computing (MEC) servers, widely deployed at road intersections, straight sections, and along highways. The MEC servers are interconnected via high-bandwidth, low-latency fiber optic backhaul links, supporting high-throughput, high-speed collaborative data transmission between nodes.

[0051] Global Control Layer: The SDN (Software Defined Networking) controller is deployed in the core data center or regional control room. The SDN controller is responsible for the real-time maintenance of the entire regional road network topology, monitoring the system load status of edge nodes, unified allocation of backhaul link bandwidth, and global routing coordination and scheduling for cross-regional, large-scale pre-migration strategies.

[0052] During the system initialization phase supported by the above architecture, the SDN controller establishes a physical and logical adjacency topology graph of all edge nodes within the coverage area through a probing protocol. Based on this, a structured historical trajectory database is constructed.

[0053] S101: The vehicle periodically collects multi-dimensional data of the vehicle through onboard sensors. When the vehicle enters the wireless coverage area of ​​the edge node, the data is uploaded to the local edge server in real time through the onboard communication device.

[0054] S102: The local edge server cleans the received raw vehicle trajectory dataset, removes jumps or outliers, and mines and extracts "vehicle local frequent itemsets" and "vehicle frequent trajectory patterns" from the processed raw vehicle trajectory set.

[0055] S103: Using the prefixes of the extracted frequent trajectory patterns as nodes, construct a multi-branch state transition T-pattern prediction tree (TPPT). The TPPT consists of three data structures: vehicle T-pattern tree, vehicle frequent itemsets, and vehicle frequent patterns. Specifically, given a vehicle historical trajectory dataset D collected locally by the edge server, the triples of this T-pattern prediction tree can be represented as:

[0056] ;

[0057] in Represents the vehicle T-pattern tree, Represents frequent itemsets of vehicles, This represents a frequent vehicle trajectory pattern. When actually executing the T-pattern tree construction algorithm, the system extracts all trajectory prefixes from the frequent vehicle patterns according to the time series and uses them as branch nodes to construct a top-down multi-branch tree transition topology.

[0058] S104: Considering the limited storage resources of edge nodes, global TPPT performs spatial partitioning. Each local edge node retains and maintains only a local subtree strongly correlated with its own physical coverage area and its immediate one-hop area. During operation, based on the handshake feedback results returned by the vehicle when it actually enters the target node, the weights of frequent patterns in the local subtree are updated in real time to maintain the timeliness of the model prediction.

[0059] Example 2:

[0060] This embodiment provides a detailed explanation of the real-time prediction of vehicles in this scheme. By exploring the spatiotemporal relationship between the dynamic displacement behavior of vehicles and the surrounding environment and physical road network, it enables the probabilistic extrapolation of the next possible edge node to be connected, providing a basis for decision-making on "where to migrate to" and "how to allocate data redundancy".

[0061] S201: First, a real-time bird's-eye view image stream captured by an onboard camera is introduced, and the bounding box coordinate features and semantic masks representing vehicles, obstacles, and lane boundaries are extracted. Then, to extract the high-order spatial constraint features inherent in the road network topology, the above graph structure is input into a graph convolutional network (GCN). After multiple layers of GCN graph convolution, the complex spatial constraints of the urban road network are transformed into dense mathematical vector representations.

[0062] S202: Before inputting into the neural network, state smoothing and data augmentation techniques are used to pre-filter the historical trajectory sequence to remove outliers. The smoothed time series is then fed into a Long Short-Term Memory (LSTM) network. The system concatenates the GCN road network spatial feature representation extracted in step S201 into the input gate of the LSTM, enabling the hidden state update function inside the LSTM to simultaneously digest temporal continuity and spatial constraints.

[0063] S203: Following the LSTM decoding output layer, a spatiotemporal attention mechanism is further introduced. Ultimately, the prediction model outputs a probability distribution vector for the current vehicle, encompassing all possible neighboring candidate edge nodes for the next step. The system categorizes candidate nodes based on this distribution vector: the edge node with the highest predicted probability is defined as the primary target node; and the remaining nodes whose probabilities exceed a preset truncation threshold (e.g., 0.15) are defined as the secondary target node set, serving as alternative nodes.

[0064] Example 3:

[0065] After obtaining the predicted route, it is necessary to determine the timing of the pre-migration. If the migration is triggered too early, the new data continuously generated by the vehicle at the source node will cause the data at the target node to become severely outdated, resulting in high costs for subsequent differential synchronization. If the migration is triggered too late in pursuit of freshness, the pre-migration may not be completed before the vehicle arrives due to transmission latency and disk loading latency. This embodiment provides a detailed description of the migration scheme in this approach.

[0066] S301: Within each extremely short decision cycle, the source node calculates the estimated remaining time for the vehicle to reach the wireless signal coverage boundary of the target edge node based on the location information fed back by the trajectory prediction model, the vehicle's instantaneous speed, and road constraint conditions. This remaining time is denoted as the remaining dwell time. .

[0067] S302: Multidimensional migration total latency calculation. Calculate the total latency required to completely deliver the vehicle's mission data to the target edge node, including communication link transmission latency. target node cache warm-up latency Superposition; wherein the transmission delay satisfies:

[0068] ;

[0069] In the formula, The total capacity of the data to be migrated. This represents the real-time available bandwidth of the current communication link. For routing queuing and physical propagation delay;

[0070] S303: The system continuously monitors the above variables and repeatedly evaluates the following trigger decision inequalities:

[0071] ;

[0072] in This is a safety redundancy time factor introduced to combat network jitter and queuing delays.

[0073] Once the aforementioned trigger inequality is met, the source node initiates the data pre-migration and push process to the target edge node.

[0074] Example 4: Erasure coding collaborative fault-tolerant storage distribution.

[0075] Due to the complexity of road traffic environments, prediction models may make misjudgments. Using only full multi-copy would severely deplete the already scarce storage and bandwidth resources on the edge; without any fault tolerance measures, prediction failures would lead to cache misses and service suspension at newly accessed nodes. This embodiment introduces erasure coding technology to strike a balance between saving storage costs and system fault tolerance. It ensures that even if a vehicle accidentally enters a low-probability secondary node, the system can recover data via the local area network using a very low proportion of checksum block resources, thus preventing long-tail latency.

[0076] S401: The encoder module of the source node reads the vehicle context dataset D that needs to be pre-migrated and performs a mathematical encoding matrix transformation on it using the Reed-Solomon forward error correction coding algorithm based on finite field polynomial interpolation. Specifically, the original dataset D is first uniformly divided into N equal basic data blocks; then, using a specific encoding generation matrix (usually composed of an N×N identity matrix and an M×N Vandermonde or Cauchy matrix), these N data blocks are mapped and expanded to generate an additional M redundant check blocks. The final generated encoding set contains a total of N+M data fragments.

[0077] S402: Based on the candidate edge node probability distribution output in Example 2, the generated N+M blocks are targeted and allocated:

[0078] Since the predicted access probability of the preferred target node is absolutely dominant, the source node directly pushes the complete N raw data blocks to the preferred target node first. Because it receives pure data blocks, if the vehicle arrives as scheduled, there is no need to consume CPU computing power for decoding; the plaintext data can be directly read for business use.

[0079] For the set of secondary target nodes, the source node distributes the M smaller parity blocks into slices according to their relative probability of exceeding the truncation threshold and sends them to these secondary nodes.

[0080] S403: After the vehicle finally completes the physical-level radio frequency switching operation:

[0081] If a vehicle enters the primary target node, it can be quickly matched with the target node since the node already holds all N data blocks in its memory. At this point, the system releases the check blocks distributed on the surrounding secondary target nodes to free up space.

[0082] If a vehicle deviates from the highest probability prediction and connects to a secondary target node, this secondary node uses the connections between edge nodes to send parallel "pull requests" to the primary node or other secondary sibling nodes. Based on the mathematical properties of RS erasure coding, this secondary node only needs to collect any N independent fragments from the aforementioned N+M fragments (regardless of whether it holds raw data blocks or parity blocks). Once all N fragments are collected, the node can construct a decoding matrix by extracting the corresponding N rows from the generator matrix and perform matrix inversion and multiplication decoding operations over a finite field, thereby quickly reconstructing the original vehicle dataset D. This recovery process only requires pulling a small number of fragments from neighbors, avoiding the latency of re-downloading the entire data over long distances.

[0083] Example 5:

[0084] In the dynamic environment of vehicle-to-everything (V2X) communication, pre-migration is not a static operation. During the time window between the full data packet transmission and the vehicle's complete departure from the old node's wireless coverage area, the onboard sensing device continues to operate and may frequently update critical state variables. Without intervention, when the vehicle retrieves pre-migration data at the new node, it will read incomplete, outdated, and potentially dangerous data, posing a significant risk to autonomous driving safety decisions. This embodiment ensures consistency between the target node's data state and the vehicle's actual physical state through lightweight micro-incremental synchronization and end-point hash handshake.

[0085] S501: The system abstracts and packages each vehicle state object that can be independently addressed and cached into a distributed data tuple, which includes a vector clock array V that records the trajectory modification of each participating node and an identity tamper-proof fingerprint digest H.

[0086] S502: When a vehicle generates a valuable update mutation at the end of the pre-migration window, the source node captures the difference variable, encapsulates it into a lightweight incremental patch packet, and delivers it to all involved target node clusters using a low-latency multicast protocol.

[0087] S503: The target edge node compares the two versions of the vector clock array V according to the preset partial order rules, performs incremental aggregation and splicing application operations, and thus smoothly advances the old state to the latest state.

[0088] S504: When a vehicle completes RF re-establishment with a new edge node and issues a connection command, this command includes a data check hash digest representing the complete session state at the last moment before leaving the previous node. Upon receiving this digest, the new edge node immediately performs a fast bit-by-bit comparison with its local data hash. If the two match, the new node releases resources and takes over the vehicle; if a conflict occurs, the system triggers a long-delay emergency interruption protection until an emergency differential synchronization comparison and retrieval process is triggered, ultimately achieving state unification.

[0089] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0090] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A data storage method based on network technology, characterized in that, Includes the following steps; S1. Collect multi-dimensional perception and positioning information of connected vehicles in real time through edge nodes, and build a vehicle driving history trajectory database based on historical vehicle trajectory data during the system initialization phase; S2. Centered on the source edge node currently providing services to the vehicle, use a pre-trained trajectory prediction model to predict the vehicle's direction of movement and trajectory in real time, output the probability distribution of the target edge nodes that the vehicle may connect to in the future, and determine the preferred target edge node and candidate target edge nodes accordingly. S3. Based on the trajectory prediction results, the current task data volume of the vehicle, the queue length, and the available bandwidth of the backhaul network, dynamically calculate the pre-migration time window, and send the data pre-transmission from the origin node to the target edge node when the triggering conditions are met. S4. Encode the full dataset of vehicles to be pre-migrated into several data blocks and verification blocks, and distribute the data blocks and verification blocks to the preferred target edge node and the secondary target edge node according to the probability distribution of the target edge node. S5. During the window period for performing data pre-migration and vehicle cross-node physical switching, if the vehicle continuously generates incremental data at the source edge node, a distributed consistency protocol based on multi-version concurrency control is used to synchronize the incremental data, and a consistency comparison is performed when the vehicle accesses a new node.

2. The data storage method based on network technology according to claim 1, characterized in that, The construction of the vehicle driving history trajectory database specifically includes: extracting vehicle local frequent itemsets and vehicle frequent trajectory patterns from historical vehicle trajectory data, and constructing them into a multi-branch state transition T-pattern prediction tree; each edge node only stores the local subtree related to its own physical coverage area and adjacent areas, and performs dynamic incremental updates based on the feedback from the actual entering node.

3. The data storage method based on network technology according to claim 1, characterized in that, The method of using a pre-trained trajectory prediction model to predict the vehicle's direction of movement and trajectory in real time, and outputting the probability distribution of target edge nodes that the vehicle may connect to in the future, specifically includes: The system acquires real-time bird's-eye view images captured by the vehicle's onboard camera, processes them using a semantic segmentation model, and obtains the bounding box coordinate features and semantic segmentation masks of entity targets within the scene. The processing results are processed using a deep neural network to extract features, and a spatial graph representation of the scene is constructed based on the K-nearest neighbor algorithm. Nodes are the extracted entity features, and edges represent the spatial adjacency relationships between entities. Subsequently, a graph convolutional network is used to perform node state aggregation calculations on the spatial graph representation to extract high-order spatial features containing road network physical topology constraints. The historical time step state sequence of the vehicle is obtained, and the sequence is filtered and preprocessed using state smoothing and data augmentation techniques. The extracted spatial features and the processed historical time step state sequence are input into a long short-term memory network to analyze and process the time evolution sequence features of vehicle dynamic high-frequency displacement. The spatiotemporal attention mechanism is used to learn the importance weights of the interaction between vehicles and the traffic environment in different spatiotemporal dimensions. The above features are combined and the normalized probability distribution vector of each target edge node that the vehicle may connect to in the future is output by a classifier.

4. The data storage method based on network technology according to claim 1, characterized in that, The determination of the preferred target edge node and candidate target edge nodes specifically includes: Obtain the normalized probability distribution vector of the trajectory prediction model, which covers all adjacent candidate edge nodes; Select the candidate edge node with the maximum predicted probability value from the probability distribution vector and determine it as the preferred target edge node; Among the remaining candidate edge nodes other than the primary target edge node, the candidate edge nodes with a predicted probability greater than a preset probability truncation threshold are included in the secondary target edge node set. Candidate edge nodes whose predicted probability is lower than the preset probability truncation threshold are removed from the candidate set.

5. A data storage method based on network technology according to claim 1, characterized in that, The process of dynamically calculating the pre-migration time window based on trajectory prediction results, the current task data volume of the vehicle, queue length, and available bandwidth of the backhaul network, and initiating data pre-transmission from the origin node to the target edge node when triggering conditions are met, specifically includes: The source node obtains vehicle location information, instantaneous speed, and road constraint based on the trajectory prediction model, and calculates the estimated remaining time for the vehicle to reach the wireless signal coverage boundary of the target edge node, which is denoted as the remaining dwell time. ; Calculate the total latency required to completely deliver the vehicle's mission data to the target edge node, the total latency including communication link transmission latency. target node cache warm-up latency Superposition; wherein the transmission delay satisfies: ； In the formula, The total capacity of the data to be migrated. This represents the real-time available bandwidth of the current communication link. For routing queuing and physical propagation delay; Introducing a safety redundancy time factor to resist network jitter and queuing delays And continuously evaluate the inequalities that trigger the judgment: ； Once the aforementioned trigger inequality is met, the source node initiates the data pre-migration and push process to the target edge node.

6. The data storage method based on network technology according to claim 1, characterized in that, The distribution of the data blocks and verification blocks to the primary target edge node and the secondary target edge node based on the probability distribution of the target edge nodes specifically includes: sending all native direct data blocks generated based on the erasure coding algorithm to the primary target edge node corresponding to the highest probability; distributing the generated redundant verification blocks to the set of secondary target edge nodes whose probabilities are greater than a set threshold according to their relative probability ratios; if a vehicle deviates from the prediction and connects to any secondary target edge node, the node uses the locally stored redundant verification blocks to request the missing fragments from neighboring nodes and performs matrix inverse operation to restore the full original data.

7. The data storage method based on network technology according to claim 1, characterized in that, The distributed consensus protocol based on multi-version concurrency control specifically includes: the source node multicasts a lightweight incremental patch carrying incremental data to the target node cluster that has received the pre-migration data; the target node performs incremental aggregation and splicing application operations; after the vehicle completes node switching, the source node sends the maintained session state data hash digest to the new node, and the new node performs cryptographic confirmation comparison with the hash value of the local data. If they match, the new node seamlessly takes over; if they conflict, the new node triggers interruption protection and differential synchronization retrieval.

Images