A multi-agent distributed optimization system for vehicle networking edge computing

Abstract

This application relates to the field of vehicle-to-everything (V2X) and edge computing communication technology, and discloses a multi-agent distributed optimization system for V2X edge computing, including a physical perception interception module, a state and conservation management module, a ledger and recovery module, and a barrier and scheduling module. The physical perception interception module updates the coherent epoch number based on the underlying channel state to intercept timeout anomaly variable packets; the state and conservation management module divides the topology edge states according to state scores, and absorbs penalty parameters into near-end regularization parameters and establishes a conservation ledger when a link is abnormal; the ledger and recovery module monitors abnormal indicators to determine whether to perform homotopy recovery or topology truncation; the barrier and scheduling module issues a computation task unloading instruction after verifying residual convergence and resource feasibility. This invention ensures the mathematical convergence of distributed optimization in a communication disconnection environment and ensures the physical security and reliable execution of edge computing instructions.

Description

Technical Field

[0001] This invention relates to the field of vehicle networking and edge computing communication technology, specifically to a multi-agent distributed optimization system for vehicle networking edge computing. Background Technology

[0002] In the context of vehicle-to-everything (V2X) edge computing, roadside units and vehicle terminals typically need to collaborate on resource allocation and task scheduling using multi-agent distributed optimization algorithms. However, V2X operates in a highly dynamic physical environment, where vehicle movement and multipath effects cause frequent fading of underlying communication links, resulting in a continuous state of disconnection and reconstruction in the communication topology between nodes.

[0003] Traditional distributed optimization algorithms rely on stable and reliable communication links and synchronous data interaction. When communication between network nodes is delayed or interrupted, lagging variable data packets continuing to participate in iterations can violate the algorithm's consistency constraints, leading to divergence in the overall minimization solution. Existing technologies often handle communication link disconnections and restorations by directly discarding constraints or crudely resetting optimization parameters. This approach disrupts historical optimization states and can easily cause severe oscillations in the computational trajectory or even system crashes when the link is reconnected. Furthermore, traditional algorithms lack physical feasibility management of intermediate iteration variables. If unconverged intermediate variables are directly issued as control commands during network transients, the actual allocation of edge computing tasks may exceed the physical capacity limits of the system's computing power and bandwidth, leading to resource conflicts and execution failures.

[0004] Therefore, this invention proposes a multi-agent distributed optimization system for edge computing in the Internet of Vehicles (IoV) to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a multi-agent distributed optimization system for edge computing in the Internet of Vehicles (IoV). This system solves the problems of distributed optimization algorithm divergence caused by frequent communication link disconnections and data lag in IoV edge computing scenarios, as well as physical resource allocation conflicts caused by the direct distribution of unconverged intermediate variables.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multi-agent distributed optimization system for edge computing in the Internet of Vehicles, comprising:

[0007] The physical sensing interception module is used to update the coherent epoch number based on the channel state of the underlying physical layer; when receiving the alternating direction multiplier variable packet, it compares the coherent epoch number carried by the alternating direction multiplier variable packet with the local coherent epoch number, marks the alternating direction multiplier variable packet with a coherent epoch number difference greater than the epoch tolerance sliding window threshold as an abnormal variable packet, and extracts metadata digest for the abnormal variable packet.

[0008] The state and conservation management module is used to calculate the comprehensive state score of the communication topology edges, and classify the communication topology edges into normal state, frozen state or shadow state based on the comprehensive state score; when the communication topology edge is in the frozen state or shadow state, the edge-level Lagrange penalty parameter associated with the communication topology edge is released and absorbed into the local near-end regularization parameter, and a conservation ledger is established for the communication topology edge in the shadow state to record state deviation.

[0009] The ledger and recovery module is used to monitor the number of epochs spanned by abnormal states and the cumulative size of the conservation ledger to determine whether the communication topology edge enters the recovery state or the truncated state; when the communication topology edge is in the recovery state, the conservation ledger is rewritten based on the homotopy recovery phase; when the communication topology edge is in the truncated state, the communication topology edge is cut off.

[0010] The barrier and scheduling module is used to search and construct a scheduling rule set among adjacent nodes that meet connectivity requirements. When the verification simultaneously meets the residual convergence condition and the resource feasibility condition, the committed barrier is released, and the local master variable is converted into an edge computing task unloading instruction and issued.

[0011] Preferably, the physical sensing interception module is specifically used to extract the Doppler frequency shift of the underlying communication link to calculate the smooth coherence time; increment the coherent epoch number when the system hardware timestamp crosses the current epoch boundary; calculate the epoch deviation between the coherent epoch number carried by the alternating direction multiplier variable packet and the local coherent epoch number; and mark the alternating direction multiplier variable packet as an abnormal variable packet when the epoch deviation is greater than the epoch tolerance sliding window threshold.

[0012] Preferably, the state and conservation management module is specifically used to extract variable freshness factor, physical link confidence factor and residual direction consistency factor for weighted calculation to obtain the comprehensive state score of the communication topology edge; to migrate communication topology edges whose comprehensive state score is less than the normal judgment threshold and are identified as key edges for maintaining local topology connectivity by the depth-first search connectivity detection algorithm to the shadow state; and to migrate communication topology edges whose comprehensive state score is less than the normal judgment threshold and are not key edges for maintaining local topology connectivity to the frozen state.

[0013] Preferably, the state and conservation management module is specifically used to assign the edge-level Lagrange penalty parameter associated with the communication topology edge in the frozen state or shadow state to zero; to superimpose the released edge-level Lagrange penalty parameter into the local near-end regularization parameter according to the transformation coefficient; and to record the remaining near-end absorption amount in the penalty quality absorption ledger, which is a component of the conservation ledger.

[0014] Preferably, the conservation ledger includes an original deviation ledger and a dual drift ledger; the state and conservation management module is specifically used to freeze the variable data received in the previous round when it was in the normal state as shadow relaxation variables; calculate the integral distance between the local master variable and the shadow relaxation variable and record it in the original deviation ledger; calculate the unexecuted dual update increment and record it in the dual drift ledger.

[0015] Preferably, the ledger and recovery module is specifically used to obtain the number of epochs spanned by the abnormal state and the cumulative scale of the original deviation ledger; when the comprehensive state score is greater than or equal to the normal judgment threshold, and the number of epochs spanned by the abnormal state is less than the maximum tolerable epoch span threshold, and the cumulative scale of the original deviation ledger is less than the maximum tolerable deviation threshold, the communication topology edge is migrated to the recovery state; when the number of epochs spanned by the abnormal state is greater than or equal to the maximum tolerable epoch span threshold, or the cumulative scale of the original deviation ledger is greater than or equal to the maximum tolerable deviation threshold, the communication topology edge is migrated to the truncated state.

[0016] Preferably, the ledger and recovery module is specifically used to exchange ledger metadata digests with neighboring nodes and perform temporal alignment based on the iteration sequence number to generate a merged ledger when the communication topology edge enters the recovery state; based on the increasing homotopy recovery phase, the merged original deviation ledger and dual drift ledger are transformed into consistency compensation vector and dual compensation vector for round-by-round cancellation; and the edge-level Lagrange penalty parameters are synchronously returned according to the total amount of penalty parameters recorded in the penalty quality absorption ledger.

[0017] Preferably, the ledger and recovery module is specifically used to remove the communication topology edge from the locally maintained adjacency list and stop including the conservation ledger in the iterative calculation when the communication topology edge enters the truncated state; start the proximal damping exponential decay mechanism; perform exponential decay calculation on the unreleased proximal absorption residual amount recorded in the penalty mass absorption ledger according to the proximal forgetting factor; and stop the exponential decay calculation when the proximal absorption residual amount is less than or equal to the proximal absorption residual amount closing threshold.

[0018] Preferably, the barrier and scheduling module is specifically used to traverse the adjacency list using a breadth-first search algorithm; to include adjacent nodes that are in the normal state or have completed ledger reconciliation and meet the maximum search hop limit into the scheduling rule set; and to numerically accumulate the available physical computing power and available channel bandwidth of all nodes included in the scheduling rule set to generate a local capacity limit for verifying the resource feasibility conditions.

[0019] Preferably, the barrier and scheduling module is specifically used to initiate a downgraded executable scheduling mechanism after the communication topology edge enters the truncated state and updates the local topology version number; redefine the scheduling rule set based on the updated local topology version number; issue edge computing task unloading instructions for high-priority services according to the service priority label; and perform local suspension or direct discard operations on low-priority services whose resource allocation requirements exceed the local capacity limit.

[0020] This invention provides a multi-agent distributed optimization system for edge computing in the Internet of Vehicles (IoV). It offers the following advantages:

[0021] 1. This invention solves the algorithm divergence problem caused by link disconnection and data lag in dynamic networks through the overall coordination of epoch interception, topology state hierarchical management, ledger recovery, and scheduling barriers. This architecture ensures mathematical convergence of multi-node distributed optimization while strictly guaranteeing the physical execution resource feasibility of the final converted edge computing instructions.

[0022] 2. When communication edges are abnormal, this invention absorbs the Lagrange penalty parameter into the local near-end regularization parameter and establishes a conservation ledger, thereby transforming external cooperative constraints into local smoothing constraints and preventing deviation of the calculation trajectory during the disconnection period; when the link is restored, the ledger is rewritten round by round through homotopy recovery phase, eliminating system oscillations caused by parameter concentration compensation.

[0023] 3. This invention constructs a scheduling rule set based on connected nodes and releases the commitment barrier after forcibly verifying the residual convergence condition and resource feasibility condition. This mechanism effectively intercepts decision variables that fail to converge or are overloaded in a transient state, ensuring the safe and reliable execution of computation task offloading instructions in a real physical environment. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0025] Figure 2 This is a schematic diagram of the method flow of the present invention;

[0026] Figure 3 This is a schematic diagram illustrating the internal working principle of the physical sensing interception module of the present invention.

[0027] Figure 4 This is a schematic diagram of the control principle of the state and conservation management module of the present invention;

[0028] Figure 5 This is a schematic diagram illustrating the internal execution flow of the ledger and recovery module and the barrier and scheduling module of the present invention.

[0029] Figure 6This is a schematic diagram showing the convergence residual comparison curves of different algorithms of the present invention under dynamic topology;

[0030] Figure 7 This is a time-series diagram illustrating the ledger size and state evolution of a single communication topology edge during disconnection and recovery cycles according to the present invention.

[0031] Figure 8 This is a bar chart comparing the success rates of unloading tasks with different service priorities under extreme operating conditions according to the present invention.

[0032] Among them, 100 is the physical perception and interception module; 200 is the state and conservation management module; 300 is the ledger and recovery module; and 400 is the barrier and scheduling module. Detailed Implementation

[0033] The technical solutions in 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Reference Figure 1 This invention provides a multi-agent distributed optimization system for edge computing in the Internet of Vehicles (IoV), deployed in an in-vehicle communication terminal and a roadside unit constituting a dynamically connected topology, comprising:

[0035] The physical sensing interception module 100 is coupled between the underlying baseband and the kernel network protocol stack of the communication terminal. It is used to extract underlying physical parameters to generate coherent epochs and intercept abnormal data packets at the receiving end based on epoch comparison to extract metadata digests.

[0036] The State and Conservation Management Module 200 is used to receive metadata digests to control the state transitions of communication topology edges, absorb penalty quality into local near-end regularization parameters when the edge state is abnormal, and build and update the original dual ledger.

[0037] The ledger and recovery module 300 is used to monitor the duration of abnormal states and perform double-end ledger merging and reconciliation operations or topology truncation and damping degradation operations based on the time threshold determination results.

[0038] The barrier and scheduling module 400 is used to evaluate the resource conditions of the scheduling rule set, perform state flipping on the main variable when the resource threshold and convergence conditions are met, and issue scheduling instructions to the lower-level physical controller.

[0039] Reference Figure 2 This invention provides a multi-agent distributed optimization method for edge computing in the Internet of Vehicles, comprising the following steps:

[0040] S100 extracts the underlying physical parameters to calculate the channel coherence time, generates a coherent epoch number and encapsulates it into an alternating direction multiplier variable packet, and intercepts abnormal data packets with epoch mismatch or timeout at the receiving end and extracts metadata digests.

[0041] S200 evaluates the state of communication topology edges based on metadata digests. When a topology edge transitions to an abnormal state, the edge-level penalty parameter is set to zero, the released penalty quality is extracted and transferred to the local near-end regularization parameter, and the original deviation ledger and dual drift ledger are constructed for the abnormal edge to solve the main variables.

[0042] S300 monitors the duration of abnormal states. When physical connectivity is restored within the time threshold, it performs double-end ledger merging and write-off. When the time threshold is exceeded, it performs a topology truncation operation, marks the corresponding ledger as truncation closed, stops the iterative calculation of the current local topology version, and performs exponential decay on the near-end regularization parameter.

[0043] S400 calculates the total physical computing power and bandwidth aggregated within the scheduling set. When the total physical resources reach the task distribution resource threshold, the master variable residuals meet the convergence conditions, and the relevant ledgers are rewritten, the master variable is flipped. Based on the flipped master variable, edge computing task unloading instructions and wireless resource block configuration instructions are generated and distributed.

[0044] To further clarify the implementation of each technical aspect of the present invention, the following will provide a detailed description of the implementation of each functional module involved above and its internal processing flow.

[0045] Reference Figure 3 In the context of vehicle-to-everything (V2X) edge computing, inter-node communication is highly susceptible to interference from high-speed vehicle movement and multipath effects, causing variable data packets in traditional distributed optimization algorithms to frequently experience delays or failures. To accurately capture the time boundaries of channel changes from the underlying physical link and isolate expired optimization variables, this embodiment introduces a physical perception interception module 100. This module 100, through a cross-layer architecture design, maps the coherent time characteristics of the physical layer to coherent epoch numbers recognizable by the application layer. It then utilizes an epoch-tolerant sliding window to mark, isolate, and extract metadata summaries from abnormal variable packets. The specific steps are as follows:

[0046] S101, in a vehicle-mounted dynamic networking environment, rapid changes in channel state directly affect the physical effectiveness of distributed optimization variables. To capture this physical-level change, the physical sensing interception module 100 extracts the Doppler frequency shift of the communication link from the underlying physical layer, thereby calculating the smoothed coherence time. For the process of calculating the coherence time from the Doppler frequency shift, those skilled in the art can refer to the classic wireless communication fading channel model for deduction; this is a well-known technique in the field and will not be elaborated upon here. The smoothed coherence time is a statistical measure of the period of change in the underlying channel state, used to reflect the time range within which the communication link remains relatively stable.

[0047] In this embodiment, the physical sensing interception module 100 maintains the link-level epoch boundary based on hardware timestamps and smoothed coherence time. When the hardware timestamp crosses the current link-level epoch boundary, the physical sensing interception module 100 increments the link-level physical coherence epoch number and updates the next epoch boundary according to the new smoothed coherence time. Considering that a unified time base is needed among distributed nodes to ensure logical consistency, the system's hardware timestamp is aligned at the underlying level using a Global Positioning System clock or a precise time protocol. Specifically, the coherence epoch number update formula is defined as:

[0048] ;

[0049] In the formula, For nodes With nodes The coherent epoch number of the communication link between them, whose value is a non-negative integer; The current system's aligned hardware timestamp, in milliseconds; For nodes With nodes The smooth coherence time of the communication link between them, in milliseconds; The minimum safe coherence time limit set for the system, in milliseconds. This parameter is used to prevent division calculation anomalies or state update stagnation caused by the coherence time approaching infinity when the vehicle is relatively stationary, causing the Doppler frequency shift to approach 0. This indicates a floor operation. This mathematical operation enables the monotonically increasing maintenance of coherent epoch numbers, discretizing continuous physical time into epoch states recognizable by optimization algorithms. In this embodiment, the aforementioned formula is used to determine the epoch interval to which the current hardware timestamp belongs. The link-level physical coherent epoch number is maintained incrementally as it crosses the current epoch boundary, based on the previous epoch count value, and does not regress due to a single smooth coherent time update.

[0050] In S102, as a preferred approach, when a node sends an alternating direction multiplier variable packet, in addition to carrying the alternating direction multiplier variable itself, the system performs epoch metadata encapsulation through the protocol stack adaptation layer; the system initializes the local topology version number to a preset initial value and increments it after the communication topology edge enters the truncated state. The specific encapsulation structure includes the main variable, dual variable, iteration sequence number, coherent epoch number, and local topology version number.

[0051] In this process, the main variables and dual variables constitute the basic data for optimization. The coherent epoch number and the local topology version number are used together at the receiver to determine whether the alternating direction multiplier method variable packet meets the conditions to be a valid input for the current alternating direction multiplier method iteration. Through the above encapsulation mechanism, each data packet carries the time identifier of its physical channel period.

[0052] S103, the physical perception interception module 100 reads the metadata of the alternating direction multiplier method variable packet in the driver receiving path of the communication device. The driver receiving path can use a kernel-mode packet receiving path or a vehicle network protocol stack adaptation layer to achieve cross-layer interception. The physical perception interception module 100 determines whether the coherent epoch number in the packet header of the alternating direction multiplier method variable packet lags behind the local current coherent epoch number, whether the iteration sequence number carried in the variable packet is consistent with the local current iteration sequence number, and whether the local topology version number is lower than the local processed local topology version number.

[0053] The system pre-sets an epoch tolerance sliding window threshold. This threshold is a boundary value for the epoch difference used to define the tolerance range of communication delays. Its value is typically set to 1 to 3 epochs, and the specific value is determined based on the average end-to-end delay distribution of the vehicular network and the mathematical robustness of the alternating direction multiplier method to the lag gradient. The physical sensing interception module 100 performs epoch deviation calculation, and the epoch deviation calculation formula is:

[0054] ;

[0055] In the formula, For the first Epochal bias under alternating direction multiplier method iteration; A positive integer, representing the current iteration number of the optimization; For sending node The coherent epoch number carried in the header of the variable packet for the alternating direction multiplier method, with superscript... Characterizes the source of the data sender; For receiving nodes The currently maintained coherent epoch number, superscript Characterizes the observation perspective of the local receiver.

[0056] When the arrival time of the variable packet in the alternating direction multiplier method exceeds the maximum waiting time set for the current iteration cycle, the maximum waiting time is the preset reception deadline within the current alternating direction multiplier method iteration cycle, or the epoch deviation. When the value exceeds the epoch-tolerance sliding window threshold, or when the iteration number carried in the variable packet is inconsistent with the current local iteration number, or when the local topology version number is lower than the local processed local topology version number, the physics-aware interception module 100 marks the alternating direction multiplier method variable packet as an anomalous variable packet. For alternating direction multiplier method variable packets falling within the epoch-tolerance sliding window threshold, the physics-aware interception module 100 allows them to enter the application layer optimization engine to participate in iterative calculations. Anomalous variable packets are not used as inputs for differentiating the current alternating direction multiplier method consistency constraints, but the metadata of the anomalous variable packets is intercepted and retained to isolate the direct damage of physical layer latency to the mathematical convergence of the algorithm.

[0057] S104, after the abnormal variable packet is marked, the kernel-mode program of the physical perception interception module 100 extracts the metadata digest of the abnormal variable packet. The metadata digest includes the sending node identifier, receiving node identifier, iteration sequence number, coherent epoch number, local topology version number, arrival timestamp, and abnormal type marker.

[0058] The exception type marker is mainly used to distinguish the specific reasons for exceptions such as packet timeout, epoch mismatch, or version lag. After extraction, the physical perception interception module 100 reports the metadata digest to the application layer optimization engine through a lock-free digest reporting mechanism. The specific implementation of lock-free digest reporting uses a lock-free circular buffer or shared memory mapping area to complete data transmission. This mechanism effectively avoids the memory copy overhead between kernel mode and user mode and the latency jitter caused by lock contention, ensuring the efficiency, real-time performance, and reliability of cross-layer communication.

[0059] Reference Figure 4 In highly dynamic distributed computing environments such as vehicle-to-everything (V2X) networks, the communication topology between nodes is often unstable. When a link suffers severe degradation or interruption, conventional distributed consistency constraint mechanisms can cause local nodes to continuously iterate on outdated data, leading to overall algorithm divergence. To address this issue, this embodiment constructs a state and conservation management module 200. This module assesses the health of communication edges using multi-dimensional factors, rationally classifying link states into different migration levels. Based on this, by transferring the penalty parameters on failed edges to a local near-end regularization term, a multi-dimensional conservation ledger is established while maintaining computational stability. This quantifies and records state deviations during disconnection, ensuring a smooth transition when the topology is restored. The specific steps are as follows:

[0060] S201. In multi-agent distributed optimization scenarios, relying solely on a single network latency determination often fails to accurately measure the actual availability of communication edges. Therefore, the state and conservation management module 200 comprehensively extracts variable freshness factors, physical link confidence factors, and residual direction consistency factors to calculate the comprehensive state score of the communication topology edges.

[0061] Specifically, the variable freshness factor is jointly quantized based on the epoch deviation, the difference between the iteration number carried in the variable packet and the local current iteration number, and the comparison result of the local topology version number, reflecting the validity of the received variable in the time dimension. When the local topology version number is lower than the local processed local topology version number or the iteration number carried in the variable packet is inconsistent with the local current iteration number, the variable freshness factor is directly set to 0. When step S103 has marked the alternating direction multiplier method variable packet as an abnormal variable packet, the corresponding communication topology edge must not migrate to the normal state in the current iteration round. In other cases, the variable freshness factor decreases monotonically with the increase of the epoch deviation. The physical link confidence factor is generated according to the joint mapping of the signal-to-interference-plus-noise ratio (SINR) and smoothed coherence time of the underlying physical layer, characterizing the reliability of the physical channel transmission quality. Its value ranges from 0 to 1. In this embodiment, the physical link confidence factor is obtained by combining the normalized value of the SINR and the normalized value of the smoothed coherence time according to a preset weighting coefficient, and decreases monotonically with the decrease of the SINR or the smoothed coherence time. The residual direction consistency factor is calculated using the angle between the original residual vector from the previous iteration and the predicted residual vector constructed based on the currently available variables. It measures the stability of the convergence direction of the optimization algorithm at the current state evaluation moment. The predicted residual vector is constructed from the effective neighbor variables from the previous iteration and their differences from the two most recent iterations. When historical differences are unavailable, the most recent effective neighbor variable is used as the prediction benchmark. In this embodiment, the formula for calculating the comprehensive state score is defined as:

[0062] ;

[0063] In the formula, For the first Node at the next iteration With nodes The comprehensive state score of the communication topology edges between them, with a value ranging from 0 to 1; It is a positive integer, representing the current iteration number of the alternating direction multiplier method; , and These are the weight coefficients of the variable freshness factor, physical link confidence factor, and residual directional consistency factor, respectively, and they satisfy the following conditions: .

[0064] The specific values ​​of the weight coefficients are predetermined based on the network's emphasis on latency sensitivity, packet loss rate, and the algorithm's own convergence characteristics. As a variable, the freshness factor For physical link confidence factor, The residual direction consistency factor is used. To ensure dimensional uniformity and computational validity, all three factors are mapped to a range of 0 to 1 using an extremum normalization function before being input into the formula. The upper and lower bounds used for extremum normalization are preset based on the protocol index range during system initialization or updated statistically based on a fixed-length historical window, and remain unchanged within the same local topology version. For the calculation of the residual direction consistency factor, if the norm of the residual vector approaches 0, the system defaults to setting its factor value to 1 to avoid division anomalies with a denominator of 0. In the implementation, the residual direction consistency factor is taken as the value of the cosine similarity between the original residual vector and the predicted residual vector from the previous iteration, after linear shift.

[0065] The State and Conservation Management Module 200 pre-sets a set of states, including normal state, frozen state, shadow state, recovery state, and truncated state. The recovery state and truncated state are not directly triggered by the comprehensive state score, but are triggered jointly by the duration of the abnormal state and the cumulative scale of the ledger in step S301. The system pre-sets a normal state determination threshold and a frozen state determination threshold, with the normal state determination threshold being greater than the frozen state determination threshold. These two thresholds are determined based on the statistical distribution of connectivity probability during historical network operation, and their values ​​range from 0 to 1.

[0066] When the overall status score When the value is greater than or equal to the normal judgment threshold, the state and conservation management module 200 migrates the communication topology edge to the normal state; when the comprehensive state score is... If the value is below the normal judgment threshold but greater than or equal to the freeze judgment threshold, it is migrated to the frozen state; when the comprehensive state score... If the edge is less than the freezing threshold and is identified by a depth-first search connectivity detection algorithm as a key edge for maintaining the connectivity of the currently visible topology using the node's locally maintained adjacency list or the regional adjacency list of roadside units as input, it is migrated to the shadow state; when the comprehensive state score is... If the edge is less than the freezing threshold and is identified by the depth-first search connectivity detection algorithm as not being a key edge for maintaining the connectivity of the currently visible topology, it is moved to the frozen state.

[0067] S202, when a communication topology edge migrates into a frozen or shadow state, the real data interaction between nodes becomes unreliable or completely interrupted. If the original edge-level Lagrangian consistency constraints are maintained, nodes may use severely outdated data for updates. Therefore, the state and conservation management module 200 triggers a penalty quality release and near-end absorption mechanism on such abnormal edges.

[0068] As a preferred approach, the state and conservation management module 200 forcibly sets the edge-level Lagrange penalty parameter corresponding to the topological edge in the frozen or shadow state to 0, interrupting the mathematical tracking of the lagged variables of adjacent nodes. To prevent the loss of constraints from causing the node's local principal variables to diverge during the minimization solution, the system absorbs the released edge-level Lagrange penalty parameter into the node's local proximal regularization parameter according to the transformation coefficient. The proximal regularization parameter update formula is defined as:

[0069] ;

[0070] In the formula, For nodes In the The proximal regularization parameter in the next iteration has the physical meaning of a penalty stiffness that limits the step size of the local variable in two adjacent iterations, and its dimension is consistent with the dimension of the second derivative of the objective function. For nodes The proximal regularization parameter in the previous iteration.

[0071] For nodes The set of abnormal edge neighbor nodes that are in a frozen or shadow state in the current iteration round; This is a conversion coefficient, ranging from 0 to 1, used to adjust the proportion of edge penalties transferred to local near-end penalties. In this embodiment, Based on the set of abnormal edge neighbor nodes The node degree of each node in the algorithm is normalized to obtain the result, and satisfies the following conditions: ; For nodes With nodes The boundary Lagrange penalty parameter between the two in the previous iteration round.

[0072] The upper limit of the system's pre-set regularization parameters is designed to prevent the penalty parameters from accumulating indefinitely when nodes are in a disconnected state for extended periods, which could lead to overly conservative or stagnant updates of local variables. Simultaneously, the system records the corresponding near-end absorption surplus for each communication topology edge that enters a frozen or shadow state. This near-end absorption surplus is written into the penalty quality absorption ledger and, together with the baseline near-end regularization parameters, constitutes the node's total near-end regularization parameters. Through this mathematical process, the system transforms the external consistency constraint pressure of the network space into the smoothing constraint pressure of the local temporal space.

[0073] S203. For critical edges in the shadow state, because they play an irreplaceable crucial role in the global convergence of the network, directly discarding their constraints would cause the original dual algorithm to lose the conditions for finding the optimal solution. Therefore, the state and conservation management module 200 constructs shadow consistency constraints for the topological edges in the shadow state and establishes a multi-dimensional conservation ledger recording system for these abnormal edges.

[0074] In this embodiment, the system freezes the variable data received in the previous round when it was in the normal communication state as shadow relaxation variables, replacing the neighbor node variables that are currently unavailable in the real network, thereby maintaining the original connected graph structure at the mathematical level. If the communication topology edge has not received valid neighbor variables before entering the shadow state, the shadow relaxation variable is initialized with the system initialization master variable or the most recent available local estimate, and the shadow relaxation variable is marked as uncalibrated. At the same time, the state and conservation management module 200 establishes a record construction mechanism locally on the node, including the original deviation ledger, the dual drift ledger, and the penalty quality absorption ledger.

[0075] The original deviation ledger is used to accumulate the integral distance between the local master variable and the shadow relaxation variable during the communication interruption period, measured using the Euclidean norm. In this embodiment, the vector difference between the local master variable and the shadow relaxation variable is recorded according to the iteration round, and the cumulative Euclidean norm value of the vector difference in each round is used as the cumulative size of the original deviation ledger. The dual drift ledger is used to store the missing amount of dual gradients caused by the cessation of updates of the edge-level Lagrange multipliers. In this embodiment, the unexecuted dual update increment vector is recorded according to the iteration round, and the cumulative Euclidean norm value of the unexecuted dual update increment vector in each round is used as the cumulative size of the dual drift ledger. The penalized quality absorption ledger accurately records the cumulative parameter values ​​transferred to the proximal regularization parameters through step S202, as well as their corresponding iteration sequence number, coherent epoch number, and local topology version number. The above ledger system converts the disconnected state in the asynchronous network into a locally quantifiable state offset, providing data support for state reconstruction during subsequent network recovery.

[0076] In step S204, after completing the state isolation and near-end absorption of penalty quality for abnormal edges, the node enters the iterative calculation phase of the local main variable. When constructing the node minimization subproblem, the state and conservation management module 200 actively excludes the true consistency constraints of frozen-state communication topology edges and introduces an enhanced near-end regularization term and constraints to maintain normal-state edges. For shadow-state edges, a consistency penalty term referencing the shadow relaxation variable is added to the local main variable minimization objective function. This consistency penalty term only participates in the solution during the shadow state's existence and exits the solution after the communication topology edge migrates to the restored or truncated state. The corresponding ledger reconciliation is still completed by the restored or truncated state process. Furthermore, the local main variable minimization objective function also includes a shadow consistency penalty term for the set of shadow-state communication topology edges. This shadow consistency penalty term references the shadow relaxation variable and is weighted by the shadow penalty coefficient corresponding to the shadow-state communication topology edge.

[0077] Specifically, the objective function corresponding to the computation process of minimizing the local host variable independently executed by each node is defined as:

[0078] ;

[0079] In the formula, For nodes In the The local master variable vector obtained in the next iteration represents the optimization decision data carried by the node itself; For nodes The local objective function is used to reflect the local data fitting or resource allocation cost of a node; For nodes The set of neighboring nodes that are in a normal state in the current iteration round.

[0080] For the first In the next iteration, the dual variables (i.e., the Lagrange multiplier vectors) on the normal topological edges are indicated by their superscripts. This represents the vector transpose operation; Neighboring nodes in a normal state The effective master variable vector passed in the previous iteration; For the first Node at the next iteration With nodes The boundary Lagrange penalty parameter between them; Represents Euclidean norm operations; This indicates the search for the independent variable that minimizes the objective function within the brackets. Minimization operation.

[0081] In this calculation process, the first half of the formula retains the effective communication and cooperation constraints on normal edges, while the second half of the formula adds a proximal regularization term. Effectively limited The single-step update range. Even if most of the neighbor links of a node are abnormally disconnected, the near-end regularization term can stabilize the local iteration trajectory and constrain the offset range of the node's main variable in the island state, thereby ensuring the smooth progress of the distributed optimization algorithm in a variable communication environment.

[0082] Reference Figure 5 In the dynamic edge computing scenario of vehicle-to-everything (V2X) networks, the communication link status between nodes frequently switches between disconnection and recovery. If the link is reconnected after a brief interruption, directly resetting the optimization parameters will cause system computational divergence; if the link is in a long-term failed state, continuously maintaining historical state data will increase the memory overhead of the nodes and delay the overall algorithm progress. To cope with the above complex working conditions, this embodiment constructs a ledger and recovery module 300. This ledger and recovery module 300 dynamically determines whether the communication topology edge enters the recovery route or the truncated route by continuously monitoring the duration of the abnormal state and the cumulative scale of the ledger. For recoverable links, the system smooths the cancellation of the ledger and returns the penalty parameters through a homotopy phase mechanism; for links that have been in a long-term failed state, the system performs topology pruning and guides the local near-end parameters back to the baseline through an exponential decay mechanism, thereby ensuring the adaptability and robustness of the distributed optimization algorithm. The specific steps are as follows:

[0083] S301, during the iterative computation process, the ledger and recovery module 300 continuously monitors the communication topology edges that enter the frozen state or shadow state. The core monitoring indicators include the number of epochs spanned by the abnormal state and the cumulative scale of the local original deviation from the ledger.

[0084] The number of epochs spanned by the anomalous state is obtained by subtracting the current local coherent epoch number from the coherent epoch number at the time of the initial link disconnection. The cumulative size of the original deviation ledger is obtained by calculating the cumulative Euclidean distance between the local master variable and the shadow slack variable during the disconnection period.

[0085] In this embodiment, the system pre-sets a maximum tolerance threshold spanning epochs and a maximum tolerance threshold for the two monitoring indicators mentioned above. The maximum tolerance threshold spanning epochs is used to define the time span limit of link failure, and its value range is typically set to 5 to 15 epochs; the maximum tolerance threshold for deviation is used to define the spatial tolerance limit of the local optimized trajectory deviating from the network consensus direction. The specific settings of these two thresholds are determined based on a comprehensive assessment of the vehicle speed distribution and the margin of the node's own computing power.

[0086] The ledger and recovery module 300 performs branch routing determination accordingly. When the comprehensive state score of the communication topology edge reaches the normal determination threshold again, and the number of epochs traversed by the abnormal state is less than the maximum tolerable epoch traversal threshold, while the cumulative scale of the original deviation from the ledger is less than the maximum tolerable deviation threshold, the ledger and recovery module 300 transitions the link state to the recovery state. When the number of epochs traversed by the abnormal state is greater than or equal to the maximum tolerable epoch traversal threshold, or the cumulative scale of the original deviation from the ledger is greater than or equal to the maximum tolerable deviation threshold, the link state is directly transitioned to the truncated state.

[0087] S302, when a communication topology edge enters the recovery state, it means that a reliable connection has been re-established between nodes, requiring the elimination of state deviations generated during the previous disconnection. During this process, nodes and their neighbors exchange their extracted ledger metadata summaries to generate a merged ledger.

[0088] Due to the asynchronous nature of network transmission, the ledger summaries exchanged between the two nodes often exhibit time-series misalignment. As a preferred approach, before performing the merge operation, the system performs time-series alignment based on the iteration numbers carried in the alternating direction multiplier method variable package. It takes the arithmetic mean of the original offset ledger vector records belonging to the same iteration number, and the arithmetic mean of the dual drift ledger vector records belonging to the same iteration number. The penalized quality absorption ledger uses the near-end absorption amount recorded by the local node as the recovery benchmark and performs consistency verification with the peer's summary. For ledger entries that fail to complete iteration number pairing, they are retained in the local ledger and continue to participate in alignment in subsequent recovery rounds.

[0089] After completing ledger alignment and merging, the ledger and recovery module 300 does not directly restore the parameters to their pre-disconnection state in a single step. Instead, it introduces an incremental homotopy recovery phase to synchronously execute a closed-loop recovery operation. The merged original offset ledger and dual drift ledger are converted into consistency compensation vectors and dual compensation vectors in batches according to the same recovery phase, and write-offs are performed round by round. The consistency compensation vector is added to the local master variable minimization subproblem, and the dual compensation vector is added to the dual variable update process. The homotopy recovery mechanism can distribute the difference in parameters to be recovered across multiple subsequent iterations, avoiding divergence in the minimization solution caused by abrupt changes in the compensation amount. The formula for calculating the homotopy recovery of the penalty parameter and near-end quality is defined as follows:

[0090] ;

[0091] ;

[0092] In the formula, For the first Node at the next iteration With nodes The edge-level Lagrange penalty parameters for topological recovery of communication between edges; For nodes In the The proximal regularization parameter at the next iteration; It is a positive integer, representing the current iteration round of optimization.

[0093] The total amount of penalty parameters to be restored recorded in the penalty absorption ledger; The conversion factor is used to match the equivalent substitution relationship of the penalty mass during the absorption and release process.

[0094] In this embodiment, For the first The homotopy recovery phase step size allocation ratio in the next iteration is a dimensionless real number, and the system sets the cumulative sum of this ratio over the entire recovery cycle to be strictly equal to 1. This incremental, step-by-step allocation mechanism avoids abrupt changes in system state caused by excessive single compensation. The ledger and recovery module 300 thereby achieves the step-by-step reconstruction of the boundary penalty parameters and the equivalent reduction of local near-end quality, ensuring the conservation of the overall system constraint energy before and after the transfer.

[0095] S303, for communication topology edges that have been determined to have migrated to a truncated state, it indicates that the link has experienced long-term physical isolation or severe fading, and continuing to maintain the mathematical connection would severely slow down the convergence speed of normal nodes. At this time, the ledger and recovery module 300 performs topology truncation and ledger closure procedures on the persistently abnormal edges.

[0096] As a preferred approach, the ledger and recovery module 300 removes the truncated edge from the adjacency list maintained locally by the node and the set of active neighbor nodes corresponding to the current local topology version, and stops including the corresponding original deviation ledger and dual drift ledger in the iterative calculation of the current local topology version. At the same time, it marks the two types of ledgers as truncated and closed. The near-end absorption residual amount corresponding to the penalty quality absorption ledger is transferred to the local decay list. Under the premise of not participating in the communication topology edge consistency constraint calculation, the exponential decay in step S304 continues to be executed until the closing condition is met and then it is marked as decay closed.

[0097] Since the local network structure has undergone substantial changes, nodes need to generate a new local topology version number and append it to the subsequent alternating direction multiplier variable packets. The new version of the topology information is used to notify relevant nodes in the currently visible topology that the communication topology edge has been physically pruned, preventing other nodes from continuing to assign computational weights to the failed communication topology edge.

[0098] S304, after a communication topology edge is truncated, the penalty mass transferred to the local near-end regularization parameter through the near-end absorption mechanism in the early stages of the edge will lose its channel to be released back to the original dual link. If the status quo is maintained, the local near-end regularization parameter of the node will remain at a high stiffness for a long time, causing the node variables to reduce their response sensitivity to new data in subsequent normal iterations. To address this, the ledger and recovery module 300 activates the near-end damping exponential decay and baseline regression mechanism for truncated edges. In this embodiment, the corresponding near-end absorption residual amount is maintained separately for each communication topology edge entering the truncated state. When the communication topology edge enters the truncated state, the unreleased absorption amount recorded in its penalty mass absorption ledger is taken as the initial near-end absorption residual amount. The near-end damping exponential decay is performed only on the near-end absorption margin corresponding to the edge of the truncated communication topology; the total near-end regularization parameter of a node is composed of the baseline near-end regularization parameter and the set. The near-end absorption residual is obtained by superimposing the residual values ​​of all communication topology edges that have not yet been released or decayed and closed. The formula for calculating the near-end damping exponential decay is defined as follows:

[0099] ;

[0100] ;

[0101] In the formula, For nodes In the During the next iteration, the communication topology edges are considered. The amount of proximal absorption residue maintained; For nodes In the The set of neighbor nodes corresponding to the communication topology edge that is in the truncated state at the next iteration; For nodes In the In the next iteration, the set of neighboring nodes corresponding to all the remaining amount of near-end absorber that has not yet been released or decayed off; For nodes In the The total proximal regularization parameter of the node is obtained by superimposing the baseline proximal regularization parameter with the residual amount of each proximal absorption in the next iteration; The baseline near-end regularization parameter is preset for the system. Its value represents the basic penalty stiffness that nodes need to maintain locally in a completely healthy and abnormal state.

[0102] The proximal forgetting factor is set to a value between 0.9 and 0.99, with the specific value depending on the system's requirement for a smooth degradation timeframe. The system further presets a threshold for closing the remaining proximal absorption. When a certain communication topology edge corresponds to When the near-end absorbance remainder corresponding to the communication topology edge is set to 0, it is removed from the set. Removed from the middle.

[0103] As a preferred implementation, the proximal absorption residual amount shut-off threshold This is a preset positive real number that approaches zero; the value range of this threshold is usually set to 10. -5 Up to 10 -3 Between, or set as the baseline proximal regularization parameter The threshold is 1% to 5%. This threshold serves to truncate the exponentially decaying mathematical tail, prevent under-float overflow in the underlying hardware, and promptly release the memory space occupied by invalid neighbor nodes locally.

[0104] In this embodiment, through the above calculation process, the excess penalty stiffness of the local drive node of the ledger and recovery module 300 is slowly lost according to an exponential law, so that it gradually and smoothly returns to the baseline state, thereby restoring the computational flexibility of the node in the current latest topology environment and avoiding overly conservative or stagnant updates due to stiffness overload.

[0105] Reference Figure 5 In the edge computing environment of the Internet of Vehicles (IoV), although the alternating direction multiplier method can achieve distributed optimization among multiple nodes, the variables in the intermediate iteration process often cannot be directly used as control commands for the physical system. Especially during transient periods of network link disconnection and ledger compensation, directly issuing unconverged or resource-infeasible intermediate states can lead to physical control failure or resource allocation conflicts. To address this issue, this embodiment constructs a barrier and scheduling module 400. By introducing a scheduling rule set and a commitment barrier mechanism, it achieves a reliable conversion of optimization variables into physical control commands while ensuring both mathematical convergence and physical resource feasibility. The specific steps are as follows:

[0106] S401, During the iterative calculation process, the barrier and scheduling module 400 searches the network topology at the end of each iteration to construct a scheduling rule set. The scheduling rule set consists of nodes whose communication topology edges participating in the scheduling are all in a normal state or whose related ledgers have all been cleared.

[0107] Specifically, the barrier and scheduling module 400 employs a breadth-first search algorithm, traversing the adjacency list maintained by the current node. To avoid uncontrolled expansion of the search scale leading to signaling overhead, the system sets a preset maximum hop limit for this breadth-first search. The hop limit is set to a range of 2 to 5 hops, with the specific value determined based on the average communication latency requirements of the edge computing cluster.

[0108] During the traversal, the system actively skips adjacent nodes that are in the frozen state, shadow state, or recovery state where the cumulative homotopy recovery phase step size allocation ratio has not reached 1, and includes available nodes that meet the connectivity requirements and are within the hop count limit into the scheduling rule set.

[0109] In this embodiment, after determining the scheduling rule set, the barrier and scheduling module 400 aggregates and statistically analyzes the physical resources within the scheduling rule set. The system reads the resource status data reported by each node and sums the available physical computing power and available channel bandwidth of all nodes within the scheduling rule set to obtain the aggregated available physical computing power and aggregated available bandwidth. These two values ​​characterize the upper limit of the overall resource carrying capacity within the local connectivity graph.

[0110] S402, after completing resource aggregation and statistics, the barrier and scheduling module 400 performs constraint condition verification. The system stipulates that the release of a commitment barrier must simultaneously meet resource feasibility conditions, variable residual convergence conditions, and ledger write-off conditions.

[0111] Regarding resource feasibility conditions, the system determines whether the resource allocation requirements represented by the main variable generated in the current iteration are less than or equal to the previously calculated aggregate available physical computing power and aggregate available bandwidth in independent dimensions. As a preferred approach, the system constructs a multi-dimensional resource request vector and compares it element-by-element with a local capacity upper limit vector containing aggregate computing power and aggregate bandwidth to ensure that no out-of-bounds conflicts occur in the allocation of physical resources in each dimension.

[0112] For ledger reconciliation conditions, the system checks that there are no communication topology edges in the current local topology version that are in a frozen or shadow state, and that all communication topology edges that have entered the recovery state have completed the homotopy recovery phase and ledger reconciliation. Communication topology edges that have migrated to the truncated state and are marked as truncated are no longer included in the ledger reconciliation judgment of the current local topology version. For the variable residual convergence condition, the determination formula is defined as:

[0113] ;

[0114] ;

[0115] In the formula, For the first The original residual vector of the next iteration is used to reflect the consistency deviation between the local master variables of a node and the variables of its neighboring nodes. Its internal elements are the absolute values ​​of the differences between the master variables of two adjacent nodes. For the first The dual residual vector of the next iteration is used to reflect the change in the consistency constraint shrinkage between two adjacent iterations. Its internal elements are the changes in the node's own principal variables in two adjacent iterations, weighted by the edge-level Lagrange penalty parameter of the corresponding communication topology edge.

[0116] A positive integer, representing the current iteration number of the optimization; This indicates the calculation of the Euclidean norm of a vector.

[0117] The preset original residual convergence threshold; These are the preset dual residual convergence thresholds. The dimensions of these two convergence thresholds are consistent with the dimensions of their corresponding residual vectors, and their value range is set to 10. -5 Up to 10 -3 The specific value is determined based on the edge computing system's control requirements for resource allocation accuracy.

[0118] The barrier and scheduling module 400 releases the suspension restriction of the current iteration master variable and triggers the release logic of the committed barrier only when all three conditions above are met.

[0119] S403, after the commitment barrier is released, the barrier and scheduling module 400 flips the state label of the master variable in the current iteration from the temporary state to the commitment state.

[0120] In this embodiment, the temporary state indicates that the main variable is only used for convergence iteration in the mathematical space and has not yet been confirmed for feasibility in the physical system; the committed state confirms that the main variable satisfies network consensus and resource physical limits and is qualified to be executed securely in the physical system.

[0121] Based on the main variable flipped to the committed state and the local topology version number after the cancellation and update completed in the aforementioned steps, the barrier and scheduling module 400 generates an edge computing task unloading instruction and a wireless resource block configuration instruction, and sends the instruction to the lower-level physical controller.

[0122] The specific encoding, modulation, and underlying radio frequency transmission processes of edge computing task offloading instructions and wireless resource block configuration instructions can be implemented by those skilled in the art using standard mobile communication protocol specifications. These are well-known technologies in the field and will not be elaborated upon here.

[0123] S404 In a vehicle-to-everything (V2X) communication environment with severe degradation, if some communication topology edges experience continuous anomalies and have already completed topology truncation and ledger closure, the original global connectivity structure of the system will experience local breaks. To prevent nodes from waiting for global convergence for a long time in such local disconnection scenarios, the barrier and scheduling module 400 triggers a degraded executable scheduling mechanism.

[0124] When a persistent abnormal edge is detected and truncated, and the local topology version number of the node has been generated and updated, the barrier and scheduling module 400 redefines the scheduling rule set based on the updated local topology version number. Within this scheduling rule set, the system masks data dependencies on the truncated abnormal node and outputs downgraded executable scheduling instructions only based on the convergent master variables within the current local network.

[0125] As a preferred approach, the downgraded executable scheduling instructions are based on the preset priority tags of the vehicle-to-everything (V2X) services. Priority is given to issuing edge computing task offloading instructions for high-priority services, while low-priority tasks that exceed the available physical computing power and bandwidth in the local aggregation are suspended locally or directly discarded.

[0126] To prevent long-term accumulation of suspended tasks leading to node memory overflow, the barrier and scheduling module 400 adds a timestamp to each task in the suspension queue. When the suspension duration exceeds the preset task latency tolerance limit, the system forcibly discards the task and sends a resource rejection signal back to the task initiator. Through the above-mentioned degraded execution process, the system ensures that even in a fragmented network topology, edge computing nodes can still maintain the operation of critical tasks based on locally available resources, avoiding long-term blocking of algorithm iterations.

[0127] This paper describes the specific implementation and verification process of the present invention in the context of a vehicle-to-everything (V2X) collaborative perception and computation offloading scenario at a complex urban intersection. The system deploys three roadside units to serve 20 dynamically moving vehicles. This cluster shares sensor data through distributed computing and jointly optimizes global trajectory and edge computing power allocation.

[0128] refer to Figures 6 to 8 High-speed vehicle weaving causes multipath fading in the physical link. The physical sensing interception module 100 reads the underlying baseband Doppler frequency shift, calculates the smoothed coherence time, and increments the coherence epoch number. When vehicle A sends an optimization variable packet to vehicle B, if spatial lag occurs due to interference, vehicle B's physical sensing interception module 100 compares the metadata at the receiving end and determines that its coherence epoch number lags behind the local epoch and exceeds the sliding window threshold. This data packet is intercepted and isolated; the lagging variables are directly intercepted and do not enter the consistency gradient calculation, and the metadata digest is reported to the application layer.

[0129] When vehicle A enters the building's obstruction zone, the link's signal-to-interference-plus-noise ratio (SNR) drops sharply. The state and conservation management module 200 calculates a comprehensive state score that falls below the freezing threshold. Because this edge is a critical edge for maintaining local topological connectivity, the state transitions to a shadow state. Vehicle B disconnects from external Lagrangian consistent pursuit, absorbing the released penalty parameter stiffness into the local near-end regularization parameter, and iterates independently using the shadow relaxation variable as an anchor point. Figure 7As shown, in the 40th to 70th iterations (shadow state interval), the system establishes the original deviation ledger, whose cumulative size increases linearly and is constrained within the maximum tolerance deviation threshold.

[0130] After vehicle A leaves the obstructed area, the link is restored. The state and conservation management module 200 determines that the restoration conditions are met, and the link switches to the restored state. Both ends of the node perform ledger timing alignment by exchanging over-the-air digests. For example... Figure 7 As shown in rounds 70 to 90, the deviation from the ledger size did not suddenly become zero; instead, a homotopy phase recovery mechanism was activated, resulting in a smooth, step-like decrease. Edge-level penalty parameters were returned in batches, and local near-end quality was synchronously and equivalently reduced, smoothing out the collaborative deviation through multiple rounds.

[0131] For vehicle C, which leaves the coverage area, the number of epochs it loses connection for reaches the upper limit. The ledger and recovery module 300 switches the corresponding link into a truncated state, updates the local topology version number, and broadcasts it. For vehicle C's legacy near-end penalty stiffness, the damping exponential decay is activated, the local constraint force is lost exponentially according to the forgetting factor, and the node weights are calculated to smoothly revert to the baseline.

[0132] After the local topology is truncated and restored, the barrier and scheduling module 400 searches for normal or ledger-verified nodes among adjacent nodes to construct a scheduling rule set, aggregating computing power and bandwidth. When the original residual and dual residual are verified to be below the convergence threshold, and the resource request is not out of bounds, the committed barrier is released. The temporary master variable is flipped to the committed state, and converted into radio frequency resource block allocation signaling and edge computing offloading instructions for execution.

[0133] To verify the effectiveness of the mechanism, an equivalent road network topology model was constructed, and experiments were conducted using the standard alternating direction multiplier method as a control benchmark.

[0134] like Figure 6 As shown, during the testing phase (after round 50), accompanied by high-frequency packet loss and topology breakage, the residual curve of the benchmark algorithm oscillated violently, and the minimization solution diverged. This invention maintains the downward trend of the residual during the link degradation interval through state isolation and a penalized near-end absorption mechanism; during the topology recovery period, the homotopy cancellation mechanism only causes minor perturbations, and the residual eventually converges smoothly to the preset accuracy (10). -4 This avoids the problem of invalid data disrupting global convergence.

[0135] like Figure 8As shown, during the tail-end testing phase of network fragmentation, the baseline solution experienced numerous control command timeouts due to global convergence failure, causing the success rate of offloading both high and low priority tasks to drop to around 40%. This invention triggers a degraded executable scheduling mechanism, reducing the scheduling rule set based on the latest topology version number. The success rate of offloading high-priority services remains above 95%, while the success rate of low-priority services is approximately 60%. This mechanism ensures the smooth operation of the core service computing power allocation channel even in scenarios with partial network disconnection.

[0136] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-agent distributed optimization system for edge computing in the Internet of Vehicles (IoV), characterized in that, include: The physical sensing interception module is used to update the coherent epoch number based on the underlying physical layer channel state; When receiving the alternating direction multiplier variable packet, the coherent epoch number carried by the alternating direction multiplier variable packet is compared with the local coherent epoch number. The alternating direction multiplier variable packets with a coherent epoch number difference greater than the epoch tolerance sliding window threshold are marked as abnormal variable packets, and metadata digests are extracted for the abnormal variable packets. The state and conservation management module is used to calculate the comprehensive state score of the communication topology edges, and classify the communication topology edges into normal state, frozen state or shadow state based on the comprehensive state score; when the communication topology edge is in the frozen state or shadow state, the edge-level Lagrange penalty parameter associated with the communication topology edge is released and absorbed into the local near-end regularization parameter, and a conservation ledger is established for the communication topology edge in the shadow state to record state deviation. The ledger and recovery module is used to monitor the number of epochs spanned by abnormal states and the cumulative size of the conservation ledger to determine whether the communication topology edge enters the recovery state or the truncated state; when the communication topology edge is in the recovery state, the conservation ledger is rewritten based on the homotopy recovery phase; when the communication topology edge is in the truncated state, the communication topology edge is cut off. The barrier and scheduling module is used to search and construct a scheduling rule set among adjacent nodes that meet connectivity requirements. When the verification simultaneously meets the residual convergence condition and the resource feasibility condition, the committed barrier is released, and the local master variable is converted into an edge computing task unloading instruction and issued.

2. The multi-agent distributed optimization system for edge computing in vehicle networking according to claim 1, characterized in that, The physical sensing interception module is specifically used to extract the Doppler frequency shift of the underlying communication link to calculate the smooth coherence time; increment the coherent epoch number when the system hardware timestamp crosses the current epoch boundary; calculate the epoch deviation between the coherent epoch number carried by the alternating direction multiplier variable packet and the local coherent epoch number; and mark the alternating direction multiplier variable packet as an abnormal variable packet when the epoch deviation is greater than the epoch tolerance sliding window threshold.

3. The multi-agent distributed optimization system for edge computing in vehicle networking according to claim 1, characterized in that, The state and conservation management module is specifically used to extract variable freshness factor, physical link confidence factor and residual direction consistency factor, perform weighted calculation to obtain the comprehensive state score of the communication topology edge; migrate the communication topology edge whose comprehensive state score is less than the normal judgment threshold and is identified as a key edge for maintaining local topology connectivity by the depth-first search connectivity detection algorithm to the shadow state; migrate the communication topology edge whose comprehensive state score is less than the normal judgment threshold and is not a key edge for maintaining local topology connectivity to the frozen state.

4. The multi-agent distributed optimization system for edge computing in vehicle networking according to claim 1, characterized in that, The state and conservation management module is specifically used to assign the edge-level Lagrange penalty parameter associated with the communication topology edge in the frozen state or shadow state to zero; to superimpose the released edge-level Lagrange penalty parameter into the local near-end regularization parameter according to the transformation coefficient; and to record the remaining near-end absorption amount in the penalty quality absorption ledger, which is a component of the conservation ledger.

5. The multi-agent distributed optimization system for edge computing in vehicle networking according to claim 4, characterized in that, The conservation ledger includes the original deviation ledger and the dual drift ledger; the state and conservation management module is specifically used to freeze the variable data received in the previous round when it was in the normal state as shadow relaxation variables; calculate the integral distance between the local master variable and the shadow relaxation variable and record it in the original deviation ledger; calculate the unexecuted dual update increment and record it in the dual drift ledger.

6. The multi-agent distributed optimization system for edge computing in vehicle networking according to claim 5, characterized in that, The ledger and recovery module is specifically used to obtain the number of epochs spanned by the abnormal state and the cumulative scale of the original deviation ledger; when the comprehensive state score is greater than or equal to the normal judgment threshold, and the number of epochs spanned by the abnormal state is less than the maximum tolerable epoch span threshold, and the cumulative scale of the original deviation ledger is less than the maximum tolerable deviation threshold, the communication topology edge is migrated to the recovery state. When the number of epochs spanned by the abnormal state is greater than or equal to the maximum tolerable epoch span threshold, or when the cumulative size of the original deviation ledger is greater than or equal to the maximum tolerable deviation threshold, the communication topology edge is migrated to the truncated state.

7. The multi-agent distributed optimization system for edge computing in vehicle networking according to claim 6, characterized in that, The ledger and recovery module is specifically used to exchange ledger metadata summaries with neighboring nodes when the communication topology edge enters the recovery state, and to perform time-series alignment based on the iteration sequence number to generate a merged ledger; based on the increasing homotopy recovery phase, the merged original deviation ledger and dual drift ledger are transformed into consistency compensation vector and dual compensation vector for round-by-round cancellation; and the edge-level Lagrange penalty parameters are synchronously returned according to the total amount of penalty parameters recorded in the penalty quality absorption ledger.

8. The multi-agent distributed optimization system for edge computing in vehicle networking according to claim 4, characterized in that, The ledger and recovery module is specifically used to remove the communication topology edge from the locally maintained adjacency list and stop including the conservation ledger in the iterative calculation when the communication topology edge enters the truncated state; Initiate the proximal damping exponential decay mechanism; perform exponential decay calculation on the unreleased proximal absorption residual amount recorded in the penalty mass absorption ledger according to the proximal forgetting factor; stop the exponential decay calculation when the proximal absorption residual amount is less than or equal to the proximal absorption residual amount closing threshold.

9. The multi-agent distributed optimization system for edge computing in vehicle networking according to claim 1, characterized in that, The barrier and scheduling module is specifically used to traverse the adjacency list using a breadth-first search algorithm; to include adjacent nodes that are in the normal state or have completed ledger reconciliation and meet the maximum search hop limit into the scheduling rule set; and to accumulate the available physical computing power and available channel bandwidth of all nodes included in the scheduling rule set to generate a local capacity limit for verifying the resource feasibility conditions.

10. The multi-agent distributed optimization system for edge computing in vehicle networking according to claim 9, characterized in that, The barrier and scheduling module is specifically used to initiate a downgraded executable scheduling mechanism after the communication topology edge enters the truncated state and updates the local topology version number; redefine the scheduling rule set based on the updated local topology version number; issue edge computing task unloading instructions for high-priority services according to the service priority label; and perform local suspension or direct discard operations for low-priority services whose resource allocation requirements exceed the local capacity limit.

Images