Networking charging data processing load balancing method based on multi-objective collaborative optimization

Abstract

The invention discloses a networking charging data processing load balancing method based on multi-target collaborative optimization, and the method comprises the following steps: S1, collecting the operation data of networking charging nodes, and carrying out the preprocessing; s2, constructing a node operation state matrix, inputting the matrix into the feature extraction network, and extracting association state features between nodes; s3, inputting the state feature vector set into a graph attention neural network, and outputting a globally dependent node representation matrix; s4, constructing a task scheduling vector set, and setting three optimization objective functions; s5, constructing a multi-objective evolutionary algorithm, and performing multi-generation iterative solution on the set objective function; s6, recording task completion time, node resource occupancy rate and actual response time, and updating network parameters; and S7, repeatedly executing the steps S2-S6 to realize dynamic evolution and load balancing control of the networking charging data processing task. According to the method, feature extraction and multi-objective optimization are fused, and efficient scheduling and stable operation of networking charging tasks are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of multi-objective optimization and load balancing, and particularly relates to a networking toll data processing load balancing method based on multi-objective collaborative optimization. BACKGROUND

[0002] With the rapid development of intelligent transportation systems, networking toll systems (such as ETC, MTC and hybrid toll architectures) as the core component to achieve fair and efficient allocation of road resources, are constantly evolving towards automation, intelligence and distribution. The networking toll system, through the front-end collection equipment deployed at each toll node, the back-end computing platform and the data transmission network, identifies the information of passing vehicles, collects data, settles fees and controls traffic. Under the trend of continuous growth of traffic flow and increasing complexity of toll business, the processing load of networking toll data faces high concurrency, high dynamics and real-time requirements. Especially during the holiday peak period, at the highway hubs and entrance and exit sections of large cities, the computing resources of toll nodes may be bottlenecked, communication delays may be increased, and service responses may be uneven.

[0003] The existing networking toll data processing method mainly relies on traditional static task allocation, polling scheduling or simple load balancing strategies (such as the minimum connection number method, the minimum load method, the weighted round robin method, etc.). Most of these methods are based on a single indicator (such as CPU usage, current connection number, recent response time) for scheduling judgment, lack joint consideration of multi-dimensional performance indicators, cannot adaptively respond to complex changes in node operating status, and are prone to performance degradation problems under high system load. In addition, although some schemes introduce machine learning or prediction mechanisms for load guidance, they generally ignore the explicit and implicit state association between toll nodes, resulting in a lack of global perspective in task scheduling and frequent local optimal problems.

[0004] From the perspective of architecture design, most current networking toll systems are based on centralized or semi-centralized platforms, which have the problems of lack of flexibility in node scheduling and inability to dynamically optimize resource allocation. Some research attempts to introduce dynamic scheduling methods based on genetic algorithms and particle swarm optimization algorithms in distributed frameworks, which have improved scheduling efficiency to some extent, but face technical bottlenecks such as difficulty in collaborative optimization of multiple objective functions, slow convergence speed of algorithms, and serious population degradation. In particular, in complex scenarios where multiple performance objectives (such as response time, node load balancing, throughput capacity, etc.) need to be optimized simultaneously, traditional evolutionary algorithms lack dynamic guidance mechanisms and cannot adaptively adjust parameters based on real-time system feedback, making the scheduling result prone to local optimum.

[0005] On the other hand, the existing research on the modeling of the state relationship between nodes is relatively simple, most of which regards the nodes as independent individuals, and fails to consider the potential coupling relationship between the charging nodes in terms of geographical distribution, data flow correlation and task collaborative processing. Once this kind of coupling relationship is not effectively modeled and quantified, it will affect the global effectiveness and generalization ability of the whole load scheduling strategy. In addition, the modeling of the node state is mostly based on a static feature extraction method, which is difficult to capture the changing trend of the node running state over time, and is not conducive to the formation of a long-term, stable and efficient scheduling model. Especially in the face of new hybrid charging architecture (such as manual and automatic charging in parallel, edge and center computing collaboration), the system task scheduling needs to have high-precision state perception and global optimization control ability.

[0006] In addition, there is currently a lack of an effective method to integrate multi-objective collaborative optimization and graph neural modeling to comprehensively handle the load scheduling problem of networked charging nodes. Traditional optimization strategies mostly use linear weighting methods to handle multiple objectives, which is difficult to solve the essential problem of conflict between objectives. For example, minimizing response time and maximizing task throughput may conflict in some scheduling decisions, and it is impossible to achieve overall optimization without sacrificing part of the target performance. Single-objective optimization methods not only constrain the richness of the solution space, but also limit the adaptability of the system to high-dynamic and complex scenarios.

[0007] In the above background, although artificial intelligence methods such as graph attention mechanism and deep representation learning have been widely applied in recommendation systems, intelligent manufacturing and industrial internet, etc., there is still no standard method system in the field of networked charging data processing scheduling. Since there is a natural network structure and data flow dependency relationship between the charging nodes, using graph neural networks to model the node relationship, extracting multi-scale semantic information based on attention mechanism, and combining evolutionary algorithms to iteratively optimize the scheduling scheme, it is expected to break through the technical difficulties of traditional methods, such as the inability to model node states comprehensively, uncontrollable target conflicts, and lack of convergence guarantee for algorithms.

[0008] Therefore, how to provide a multi-objective collaborative optimization-based load balancing method for networked charging data processing is a problem that those skilled in the art need to solve. SUMMARY

[0009] One object of the present application is to provide a multi-objective collaborative optimization-based load balancing method for networked charging data processing, which fully utilizes graph attention neural networks, feature extraction networks and multi-objective evolutionary algorithms, and describes in detail the whole process of modeling, scheduling optimization and adaptive evolution of networked charging nodes in complex traffic scenarios, realizes multi-objective synchronous optimization of task response time, node load balancing and system throughput capacity, and has the advantages of high scheduling efficiency, strong system robustness and strong adaptability to dynamic changes.

[0010] The multi-target cooperative optimization-based networking charging data processing load balancing method according to the embodiment of the application comprises the following steps:

[0011] S1, collecting running data of the networking charging nodes and performing preprocessing;

[0012] S2, constructing a node running state matrix and inputting the node running state matrix into a feature extraction network to extract explicit and implicit associated state features of the nodes and generate a state feature vector set;

[0013] S3, inputting the state feature vector set into a graph attention neural network to generate a weighted graph structure representing the state correlation between the nodes and output a node representation matrix containing global dependency relationships;

[0014] S4, constructing a task scheduling vector set according to the node representation matrix and setting three optimization objective functions, the objective functions being minimizing the average response time of the tasks, minimizing the standard deviation of the node loads and maximizing the number of tasks completed per unit time;

[0015] S5, constructing a multi-objective evolutionary algorithm, taking the task scheduling vector set as an initial population, performing multi-generation iterative solving on the set objective functions and outputting an optimal task allocation matrix;

[0016] S6, dispatching the current task to the corresponding networking charging node, recording the task completion time, node resource occupancy rate and actual response time and updating the edge weight coefficients in the weighted graph structure and the evolutionary algorithm parameters;

[0017] S7, repeatedly executing steps S2-S6 to realize dynamic evolution and load balancing control of the networking charging data processing tasks.

[0018] Optionally, the running data comprises the number of node tasks, processing delay, CPU utilization rate, memory occupancy rate and network transmission rate.

[0019] Optionally, the preprocessing comprises data deduplication, missing value filling, outlier removal, uniform timestamp and data standardization.

[0020] Optionally, the feature extraction network is composed of a multi-layer residual convolution structure and a self-attention mechanism, the residual connection is used to retain the original information of the node running state matrix, the self-attention mechanism is used to enhance the feature capturing capability for the explicit and implicit associations between the nodes, the state feature vector set output by the feature extraction network is processed by batch normalization and a nonlinear activation function after each convolution output, and a fully connected layer is connected to the last layer to generate a fixed-dimensional state feature representation.

[0021] Optionally, the graph attention neural network takes a set of state feature vectors as input, constructs a multi-head attention mechanism to capture multi-scale interaction relationships between nodes, generates an initial weighted graph structure by calculating the attention weight coefficients between each node and its adjacent nodes, introduces a Laplace regularization term to constrain the smoothness of the attention distribution, enhances the propagation consistency of node features in the graph structure, and finally outputs a node representation matrix containing global dependency relationships and context semantics through a graph convolution layer.

[0022] Optionally, the multi-objective evolutionary algorithm is improved by adopting a two-stage guidance mechanism and an adaptive population compression strategy. In the first stage, a guided mutation operation based on node state gradient is introduced during the evolution of the initial population. In the second stage, a local disturbance mechanism based on task scheduling bias degree is introduced when approaching convergence. At the same time, the number of low-contribution individuals is adaptively compressed by calculating the convergence density of population individuals, thereby improving the convergence accuracy of scheduling solutions and the ability to maintain population diversity.

[0023] Optionally, the S2 specifically includes:

[0024] S21, frame processing is performed on the preprocessed operation data to construct a node operation state matrix wherein R i,j represents the operation value of the i-th node in the j-th feature dimension, N represents the total number of nodes, and M represents the number of operation features of each node;

[0025] S22, inputting the node operation state matrix R into a feature extraction network to perform multi-layer residual convolution operation and self-attention fusion processing to generate an intermediate feature tensor T wherein d is the number of extracted feature dimensions, the convolution kernel size is set to k x k, and each layer of convolution output is processed through batch normalization and ReLU activation function, and the residual connection adopts an "input-output addition" method to complete dimension alignment;

[0026] S23, obtaining a state feature vector set V = {v1, v2, …, v N} by performing feature channel compression operation on the intermediate feature tensor T, the channel compression adopts one-dimensional fully connected layer transformation operation, and the state feature vector set satisfies the following optimization relationship:

[0027]

[0028] wherein min(·) represents the minimum value function, A i,j represents the explicit association measure value between node i and node j, v i , v j represents the state feature vector of the i, j-th node, is v iThe transpose of , σ(·) represents the Sigmoid function, ||·|| 2 represents the square of the Euclidean norm, W represents the feature interaction weight matrix, R i Represents the row vector of node i in the original running state matrix, W r represents the state mapping weight matrix, and λ is the regularization factor, which is used to control the degree of preservation between node features and the original state.

[0029] Optionally, the S3 specifically includes:

[0030] S31. Input the state feature vector set V into the graph attention neural network to construct a graph structure G = (U, E), where U represents the node set and E represents the edge set. The initial weight of each edge is set to a constant α.

[0031] S32, for each node u i ∈U constructs H attention heads, each of which uses a different linear mapping matrix The input feature v i Mapping to subspace vector in

[0032] S33. For each attention head h∈{1,2,…,H}, calculate the node u i and adjacent node u j The attention weight Use a scoring mechanism with a Laplace regularization term:

[0033]

[0034] Among them, || represents the vector splicing operation, is the weight vector of the h-th attention head, for a h The transpose of , μ is the regularization coefficient, represents the square of the Euclidean distance between node i and node j in the graph Laplace embedding space, represents the square of the Euclidean distance between node i and node k in the graph Laplace embedding space, is the graph Laplacian embedding matrix obtained by spectral decomposition, 1 represents the embedding dimension, Represents node u i The set of adjacent nodes, exp(·) represents the natural exponential function with e as the base, LeakyReLU(·) represents the activation function, denote subspace vectors respectively;

[0035] S34. Concatenate the output results of all attention heads to obtain the multi-head fusion representation vector of each node

[0036] S35, fusing the node representation set The input graph convolution operation layer aggregates features and outputs a node representation matrix with global dependency and context semantic fused node state expression:

[0037]

[0038] wherein Z represents the node representation matrix, represents an adjacency matrix plus a self-loop, A is a weighted adjacency matrix constructed by attention weights, I N is an N*N identity matrix, is a degree matrix of W z is a weight matrix of the graph convolution layer, and σ(·) is an activation function. The finally output node representation matrix Z represents a node state expression with global dependency and context semantic fusion.

[0039] Optionally, the S4 specifically includes:

[0040] S41, according to the node representation matrix Z and the current task set to be scheduled wherein K represents the number of tasks to be scheduled, t k represents the kth task to be scheduled, and a task scheduling vector set is constructed wherein x k,i ∈{0,1} represents whether the task t k is assigned to the node u i , satisfying the constraint condition

[0041] S42, three optimization objective functions in the task scheduling process are set, respectively being the minimum of task average response time, the minimum of node load standard deviation, and the maximum of the number of tasks completed per unit time, and a joint optimization objective function is constructed:

[0042]

[0043] wherein C k represents the complexity constant of the task t k , P i represents the processing capacity per unit time of the node u i , represents the total task load of the node u i , represents the average load of all nodes, δ k,i ∈{0,1} represents whether the task t kwhether the task t i is able to be completed by the node u N in the time period τ, 1 if able to be completed, otherwise 0, τ represents the length of a unit scheduling period, ω1, ω2, ω3 are weighting factors of each objective function;

[0044] S43, obtaining a task scheduling vector set by using a task allocation result expressed in the objective function, and taking the task scheduling vector set as an input of a multi-objective evolutionary solving process.

[0045] Optionally, the S5 specifically includes:

[0046] S51, taking the task scheduling vector set as an initial population of a multi-objective evolutionary algorithm, each individual representing a task allocation scheme, assuming that a population size is Q, and an individual set is represented as An iteration number is g = 0;

[0047] S52, in the gth iteration evolution, for each individual calculating a node mutation probability matrix by using a characteristic gradient vector of each node in a node representation matrix Z = {z1, z2, …, zn}: N

[0048]

[0049] wherein, represents whether the task t k is allocated to the node u i as a mutation candidate, ε is a constant for preventing a denominator from being zero, respectively represent different characteristic gradient vectors, represents whether the task t k is allocated to the node u i , represents whether the task t k is allocated to the node u j ;

[0050] S53, in the evolution process of each iteration population, according to individual non-inferior sorting results in a joint optimization objective function space, recording sorting positions in the objective functions F1, F2, F3, and constructing a scheduling deviation function:

[0051]

[0052] wherein, respectively represent optimal values of three optimization objective functions in a current population, respectively represent maximum or minimum boundary values of the three optimization objective functions, represents an individual a scheduling bias value in the joint optimization objective function space;

[0053] S54, before the evolution algebra reaches G, the guiding crossover, bias disturbance and environmental selection operations are continuously carried out, and after each iteration, according to the population density function, the low-quality individuals with high density and large scheduling bias value are adaptively compressed:

[0054]

[0055] wherein, represents the aggregation density between the qth individual and other individuals, the distribution represents different individuals of the gth generation, epsilon represents a density perception adjustment parameter, exp(·) represents a natural exponential function with e as the base, and G represents the maximum iteration number in the multi-objective evolutionary algorithm;

[0056] S55, when the algebra g=G, the non-dominated solution set on the Pareto optimal frontier is selected to construct an optimal task allocation matrix, and the optimal task allocation matrix is output.

[0057] The beneficial effects of the present application are:

[0058] Firstly, the present application can fully extract the explicit and implicit state association between the networking charging nodes by constructing the node running state matrix and combining the multi-layer residual convolution structure and the self-attention mechanism, generate the node state feature vector with context semantics and global dependence features, effectively enhance the description ability of the dynamic changes of the node running state, and provide high-dimensional and high-precision input expression for subsequent scheduling optimization, which fundamentally overcomes the problems of single state modeling and information missing in traditional methods.

[0059] Secondly, the present application introduces a graph attention neural network to construct a weighted graph structure and fuse a multi-head attention mechanism and a Laplacian regularization term, realizes modeling of multi-scale interaction relationship between nodes and consistency enhancement of propagation, combines a graph convolution operation to output a node representation matrix, so that the task scheduling strategy has a global optimization vision. At the same time, three optimization objectives of minimizing the average response time, minimizing the load standard deviation and maximizing the task throughput are set, and a multi-objective evolutionary algorithm is used for efficient solution, which effectively solves the technical problems of multi-objective conflict and difficult coordination of multi-index performance in traditional methods.

[0060] Finally, the multi-objective evolutionary algorithm proposed in the application fuses the guided mutation mechanism based on the node state gradient and the local disturbance mechanism based on the task scheduling deviation degree in the evolution process, combines the adaptive population compression strategy, maintains the population diversity while ensuring the scheduling solution convergence accuracy, enables the system to realize stable and efficient task scheduling when facing complex traffic states and sudden load fluctuations, and greatly improves the real-time performance, balance and resource utilization of the networked toll data processing process, and has significant engineering promotion value. BRIEF DESCRIPTION OF DRAWINGS

[0061] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application, and do not limit the application. In the drawings:

[0062] Fig. 1 The flow chart of the networked toll data processing load balancing method based on multi-objective collaborative optimization proposed in the application;

[0063] Fig. 2 The weighted graph structure and node representation matrix construction flow chart of the networked toll data processing load balancing method based on multi-objective collaborative optimization proposed in the application;

[0064] Fig. 3 The data flow chart of the networked toll data processing load balancing method based on multi-objective collaborative optimization proposed in the application. DETAILED DESCRIPTION

[0065] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and therefore only show the components related to the application.

[0066] Reference Figs. 1-3 The networked toll data processing load balancing method based on multi-objective collaborative optimization comprises the following steps:

[0067] S1, collecting the running data of the networked toll nodes and pre-processing;

[0068] S2, constructing a node running state matrix, inputting the node running state matrix into a feature extraction network, extracting explicit and implicit associated state features between nodes, and generating a state feature vector set;

[0069] S3, inputting the state feature vector set into a graph attention neural network, generating a weighted graph structure representing the state correlation between nodes, and outputting a node representation matrix containing global dependency relationships;

[0070] S4, construct a task scheduling vector set according to the node representation matrix, and set three optimization objective functions, the objective functions are respectively minimizing the task average response time, minimizing the node load standard deviation and maximizing the number of tasks completed per unit time;

[0071] S5, construct a multi-objective evolutionary algorithm, take the task scheduling vector set as the initial population, perform multi-generation iteration solving on the set objective functions, and output an optimal task allocation matrix;

[0072] S6, dispatch the current task to the corresponding networking charging node, record the task completion time, node resource occupation rate and actual response time, and update the edge weight coefficient in the weighted graph structure and the evolutionary algorithm parameters;

[0073] S7, repeat steps S2-S6 to realize dynamic evolution and load balancing control of the networking charging data processing task.

[0074] The present application proposes a complete networking charging data processing load balancing method based on multi-objective collaborative optimization, which integrates feature extraction, graph neural modeling and evolutionary scheduling optimization process, and can realize efficient resource scheduling and task allocation of networking charging nodes in a dynamic traffic environment, improve the overall response performance and processing throughput of the system, and effectively alleviate the data processing bottleneck in high concurrency scenarios.

[0075] In the embodiment, the running data includes node task quantity, processing delay, CPU utilization, memory occupation rate and network transmission rate.

[0076] The present application provides a comprehensive and accurate input basis for subsequent state modeling and scheduling optimization by clearly defining the composition of running data, including node task quantity, processing delay, CPU utilization, memory occupation rate and network transmission rate, and enhances the real perception ability of the system to the node running state.

[0077] In the embodiment, the preprocessing includes data deduplication, missing value filling, outlier removal, unified timestamp and data standardization.

[0078] The present application performs data deduplication, missing value filling, outlier removal, unified timestamp and data standardization operations on the original running data in the preprocessing stage, significantly improves the data quality and consistency, reduces the interference of data abnormalities on the performance of the scheduling algorithm, and enhances the stability and robustness of the overall system.

[0079] In this embodiment, the feature extraction network is composed of a multi-layer residual convolution structure and a self-attention mechanism, the original information of the node operation state matrix is preserved through residual connection, and the feature capturing ability of explicit and implicit correlations between nodes is enhanced through the self-attention mechanism, the state feature vector set output by the feature extraction network is processed through batch normalization and a nonlinear activation function after each convolution output, and a fixed-dimensional state feature representation is generated through a fully connected layer in the last layer.

[0080] The feature extraction network designed in the application combines a multi-layer residual convolution structure and a self-attention mechanism, so that the explicit and implicit state relationships between nodes can be effectively extracted, the original information structure is preserved, the integrity and representativeness of high-dimensional state feature expression are enhanced, and strong support is provided for constructing an accurate scheduling basis.

[0081] In this embodiment, the graph attention neural network takes the state feature vector set as input, constructs a multi-head attention mechanism to capture multi-scale interaction relationships between nodes, generates an initial weighted graph structure by calculating the attention weight coefficients between each node and adjacent nodes, introduces a Laplace regularization term to constrain the smoothness of the attention distribution, enhances the propagation consistency of node features in the graph structure, and performs feature aggregation operation through a graph convolution layer, and finally outputs a node representation matrix containing global dependency relationships and context semantics.

[0082] The application introduces a graph attention neural network to construct a state-related graph structure between nodes, uses a multi-head attention mechanism and a graph convolution aggregation feature, so that node features can be dynamically captured and propagated under multi-scale semantics, effectively model the complex interaction relationships between nodes, and improve the global perception ability in the scheduling process.

[0083] In this embodiment, the multi-objective evolutionary algorithm is improved by using a two-stage guidance mechanism and an adaptive population compression strategy, the guidance mutation operation based on the node state gradient is introduced in the initial population evolution process in the first stage, the local disturbance mechanism based on the task scheduling deviation degree is introduced when approaching convergence in the second stage, the number of low-contribution individuals is adaptively compressed through the convergence density of population individuals, and the convergence accuracy of scheduling solutions and the diversity maintenance ability of the population are improved.

[0084] The multi-objective evolutionary algorithm proposed in the application integrates a two-stage guidance mechanism and an adaptive population compression strategy, strengthens the node state guidance in the early stage of scheduling evolution, strengthens the diversity control of solutions in the convergence stage, realizes the dual optimization of accuracy and efficiency, and improves the quality and controllability of the final scheduling result.

[0085] In this embodiment, the S2 specifically comprises:

[0086] S21, frame processing is performed on the pretreated operation data, and a node operation state matrix is constructed wherein R i,j represents the operation value of the i th node in the j th feature dimension, N represents the total number of nodes, and M represents the number of operation features of each node;

[0087] S22, input the node operation state matrix R into a feature extraction network, perform multi-layer residual convolution operation and self-attention fusion processing, and generate an intermediate feature tensor wherein d is the number of extracted feature dimensions, the convolution kernel size is set to k×k, each layer of convolution output is subjected to batch normalization and ReLU activation function processing, and the residual connection adopts an "input-output addition" mode to complete dimension alignment;

[0088] S23, obtain a state feature vector set V={v1, v2,…, v N} by performing a feature channel compression operation on the intermediate feature tensor T, the channel compression adopts a one-dimensional full connection layer transformation operation, and the state feature vector set satisfies the following optimization relationship:

[0089]

[0090] wherein min(·) represents a minimum value function, A i,j represents the explicit association degree value between node i and node j, v i ,v j represents the state feature vector of the i th and j th node, is the transpose of v i , σ(·) represents a Sigmoid function, and ||·|| 2 represents the square of the Euclidean norm. 2 i represents the row vector of node i in the original operation state matrix, W r represents a state mapping weight matrix, and λ is a regularization factor for controlling the preservation degree between node features and original states.

[0091] The node operation state matrix and the state feature vector set are constructed in detail, the high-dimensional feature extraction and specific optimization relationship constraints are ensured, the extracted features have both discriminability and consistency of original state mapping, and an accurate, compact and semantically rich representation form is provided for dispatching optimization.

[0092] In the embodiment, the S3 specifically comprises:

[0093] ​S31, input the state feature vector set V into a graph attention neural network, construct a graph structure G=(U, E), wherein U represents a node set, E represents an edge set, the weight of each edge is initialized, and is set as a constant a;

[0094] S32, for each node u i ∈U, construct H attention heads, each attention head uses a different linear mapping matrix maps the input feature v i to a subspace vector wherein

[0095] S33, for each attention head h∈{1, 2, …, H}, calculate the attention weight i of the adjacent node u j of the node u using a scoring mechanism with a Laplacian regularization term:

[0096]

[0097] wherein || represents a vector concatenation operation, is the weight vector of the hth attention head, is the transpose of a h , μ is a regularization coefficient, represents the Euclidean distance square of node i and node j in the graph Laplacian embedding space, represents the Euclidean distance square of node i and node k in the graph Laplacian embedding space, is a graph Laplacian embedding matrix obtained by spectral decomposition, 1 represents the embedding dimension, represents the adjacent node set of the node u i , exp(·) represents a natural exponential function with base e, LeakyReLU(·) represents an activation function, respectively represent subspace vectors;

[0098] S34, concatenate the output results of all attention heads to obtain a multi-head fusion representation vector of each node

[0099] S35, input the fused node representation set into a graph convolution operation layer, aggregate features, and output a node representation matrix, wherein the node representation matrix has a global dependency relationship and a context semantic fusion node state expression:

[0100]

[0101] wherein Z represents a node representation matrix, denotes the adjacency matrix plus self-loop, A is a weighted adjacency matrix constructed by attention weights, I N is an N*N identity matrix, is the degree matrix of , W z is the weight matrix of the graph convolution layer, and sigma (·) is an activation function.

[0102] The application realizes dynamic modeling of the state relationship between nodes by constructing node attention weights, graph convolution fusion and Laplace regularization term smoothing mechanism, improves the information propagation consistency and global expression ability of the network in the node task allocation process, and provides a stable and accurate support structure for scheduling optimization.

[0103] In the embodiment, the S4 specifically includes:

[0104] S41, according to the node representation matrix Z and the current to-be-scheduled task set wherein K represents the number of to-be-scheduled tasks, t k represents the kth to-be-scheduled task, and a to-be-scheduled task vector set is constructed wherein x k,i ∈{0,1} represents whether the task t k is allocated to the node u i , and satisfies the constraint condition 1;

[0105] S42, three optimization objective functions in the task scheduling process are set, which are respectively the minimum of the task average response time, the minimum of the node load standard deviation and the maximum of the number of completed tasks per unit time, and a joint optimization objective function is constructed:

[0106]

[0107] wherein C k represents the complexity constant of the task t k , P i represents the processing capacity per unit time of the node u i , L represents the total task load of the node u i , L represents the average load of all nodes, δ k,i ∈{0,1} represents whether the task t k can be completed by the node u i in the time period τ, and if it can be completed, it is 1, otherwise it is 0, τ represents the length of the unit scheduling period, and ω1, ω2 and ω3 are the weighting factors of the objective functions.

[0108] S43, obtaining a task scheduling vector set by using the task allocation result expressed in the objective function, and taking the task scheduling vector set as an input of a multi-objective evolutionary solving process.

[0109] The application constructs a task scheduling vector based on a node representation matrix, and designs three-objective optimization functions of minimum average response time, minimum load balancing degree and maximum throughput, so that the task scheduling achieves multi-objective coordination while meeting performance indicators, and comprehensively improves scheduling efficiency and resource utilization.

[0110] In the embodiment, the S5 specifically includes:

[0111] S51, obtaining a task scheduling vector set As an initial population of a multi-objective evolutionary algorithm, each individual represents a task allocation scheme, and the population size is Q, and the individual set is represented as The generation number is g=0;

[0112] S52, in the gth generation evolution, for each individual Using the characteristic gradient vector of each node in the node representation matrix Z={z1, z2,…, z N} to calculate the node mutation probability matrix:

[0113]

[0114] Wherein, Indicates whether the task t k is allocated to the node u i as a mutation candidate, and ε is a constant to prevent the denominator from being zero, Indicate different characteristic gradient vectors, Indicates whether the task t k is allocated to the node u i in the gth generation, Indicates whether the task t k is allocated to the node u j in the gth generation;

[0115] S53, in the evolution process of each generation population, according to the individual The non-inferior sorting result in the joint optimization objective function space is recorded in the sorting position in the objective functions F1, F2 and F3, and a scheduling deviation function is constructed:

[0116]

[0117] Wherein, Indicate the optimal values of the three optimization objective functions in the current population, respectively represent the maximum or minimum boundary value of three optimization objective functions, representing an individual The scheduling deviation value in the joint optimization objective function space is adjusted.

[0118] S54, before the evolution algebra reaches G, the guided crossover, deviation disturbance and environmental selection operations are continuously performed, and after each iteration, the low-quality individuals with high density and large scheduling deviation value are adaptively compressed according to the population density function:

[0119]

[0120] wherein, representing the aggregation density between the qth individual and other individuals, The distribution represents different individuals of the gth generation, epsilon represents the density perception adjustment parameter, exp(·) represents the natural exponential function with e as the base, and G represents the maximum iteration number in the multi-objective evolutionary algorithm.

[0121] S55, at the algebra g=G, the non-dominated solution set on the Pareto optimal frontier is selected to construct an optimal task allocation matrix, and the optimal task allocation matrix is output.

[0122] The multi-objective evolutionary scheduling algorithm constructed by the state gradient guidance, scheduling deviation disturbance, adaptive density compression and Pareto solution set extraction mechanism effectively improves the stability, adaptability and optimization quality of the evolutionary process, and ensures that the final task scheduling scheme has strong global optimization ability and practicality.

[0123] Example 1:

[0124] In order to verify the feasibility of the application in the implementation, the application is applied to

[0125] In order to verify the feasibility of the application in the implementation, the application is applied to the real running environment of a county-level highway networked toll collection system, covering a high-traffic peak frequency, dense node deployment, and complex task scheduling of a highway trunk road. There are 48 networked toll nodes in this area, and the system needs to process tasks including license plate recognition image uploading and analysis, billing data generation, blacklist comparison, and real-time transaction data synchronization. Due to the frequent occurrence of increased processing pressure, significantly increased response delay, and partial task lag accumulation in some nodes during holidays and peak travel periods, the overall network scheduling imbalance problem is becoming increasingly serious, which has become a key bottleneck affecting the stable operation of the toll collection system.

[0126] In the actual deployment process, the charging system originally adopts a static allocation method based on the minimum CPU load strategy, which is easy to form local resource overload under high concurrency conditions, and part of the nodes have low resource utilization, which cannot realize the optimal matching of resources and tasks. The method of the application constructs a node running state matrix, obtains the running data of each node in real time, including the current task number, processing delay, CPU utilization, memory occupancy and network transmission rate, inputs the multi-layer residual convolution and self-attention feature extraction network after preprocessing, extracts explicit and implicit state features, and combines the graph attention neural network to model the state correlation between nodes. Then, according to the extracted node features, a task scheduling vector set is constructed, and combined with a three-dimensional optimization objective function (task response time, load standard deviation, task completion rate) into a two-stage guided multi-objective evolutionary algorithm.

[0127] In the test process, two typical time periods of a working day and a weekend are selected for comparison experiment. The morning peak period of the working day is 7:00-9:00, and the weekend return peak is 16:00-18:00. The system generates about 1200-1500 tasks in each 1-minute time window, and the task complexity is randomly distributed in the range of [1.0, 3.5] (unit calculation consumption standardization). The CPU performance of the node is distributed between 2.0GHz and 3.8GHz, the memory is 4GB to 32GB, and the network bandwidth difference is up to 40%. The average response time of the original system in the peak period is as high as 420ms, the response delay of part of the nodes is more than 700ms, the utilization rate of the lowest node is less than 18%, the resource occupancy rate of the highest node is up to 93%, and the task processing accumulation occurs frequently.

[0128] After running the method of the application, the average response time is significantly reduced to 245ms, the overall node load standard deviation is reduced from 26.4% to 11.2%, the task completion amount per unit time is increased by 23.7%, and the resource utilization distribution is more balanced. In the two-stage evolutionary scheduling, through the guiding mutation and density compression mechanism, the allocation scheme with poor performance in the optimization objective is effectively eliminated, and after the 12th iteration, the scheduling solution tends to be stable, and the average value of the scheduling deviation function is reduced by about 52%. In terms of task scheduling success rate, the method increases the high complexity task completion rate from 85.3% to 96.4%, and the medium and low complexity task completion rate always remains above 98%.

[0129] In addition, through the dynamic monitoring feedback mechanism in the deployment period, the method can automatically adapt to special load fluctuation conditions such as holidays, and the task response dynamic adjustment of the node has an average time delay of not more than 150ms, which ensures the real-time and continuity of the charging service. In a sudden traffic surge (vehicle detention batch release due to traffic accident) scene, the system completes the load rebalancing scheduling within two minutes, avoiding data accumulation and task retransmission phenomenon, and ensuring the stable operation of the networked charging system.

[0130] The following is a test data summary table showing the core performance comparison of the method of the present application and the original method in the actual operation scenario:

[0131] Table 1 Comparison of dispatching effect data summary table of networked toll collection system

[0132]

[0133]

[0134] In summary, the present application shows significant performance advantages in the real networked toll environment, not only solves the problem of imbalance of node resource scheduling, but also effectively improves the intelligent scheduling level, robustness and resource utilization efficiency of the whole system, and has good popularization and application value.

[0135] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this, any person skilled in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A load balancing method for network charging data processing based on multi-objective collaborative optimization, characterized in that: The steps include: S1. Collecting and preprocessing the operating data of networked charging nodes; S2. Constructing a node operation state matrix and inputting the node operation state matrix into a feature extraction network to extract explicit and implicit association state features between nodes and generate a state feature vector set; S3. Input the state feature vector set into the graph attention neural network to generate a weighted graph structure representing the state correlation between nodes, and output a node representation matrix containing global dependencies; S4. Construct a set of task scheduling vectors based on the node representation matrix and set three optimization objective functions, which are minimizing the average task response time, minimizing the standard deviation of node load, and maximizing the number of tasks completed per unit time; S5. Construct a multi-objective evolutionary algorithm, use the task scheduling vector set as the initial population, perform multiple generations of iterative solutions to the set objective function, and output the optimal task allocation matrix; S6. Dispatching the current task to the corresponding network charging node, recording the task completion time, node resource occupancy rate and actual response time, and updating the edge weight coefficients and evolutionary algorithm parameters in the weighted graph structure; S7. Repeat steps S2-S6 to achieve dynamic evolution and load balancing control of the network charging data processing task.

2. The method for load balancing of network charging data processing based on multi-objective collaborative optimization according to claim 1 is characterized in that: The operation data includes the number of node tasks, processing delay, CPU utilization, memory occupancy and network transmission rate.

3. The method for load balancing of network charging data processing based on multi-objective collaborative optimization according to claim 1, characterized in that: The preprocessing includes data deduplication, missing value filling, outlier removal, unified timestamp and data standardization.

4. The method for network charging data processing load balancing based on multi-objective collaborative optimization according to claim 1 is characterized in that: The feature extraction network is composed of a multi-layer residual convolution structure and a self-attention mechanism. The original information of the node operation state matrix is ​​retained through residual connections, and the self-attention mechanism is used to enhance the ability to capture the features of explicit and implicit associations between nodes. The state feature vector set output by the feature extraction network is batch normalized and processed with a nonlinear activation function after each layer of convolution output, and is connected to a fully connected layer in the last layer to generate a fixed-dimensional state feature representation.

5. The method for load balancing of network charging data processing based on multi-objective collaborative optimization according to claim 1, characterized in that: The graph attention neural network takes a set of state feature vectors as input, constructs a multi-head attention mechanism to capture the multi-scale interaction relationship between nodes, generates an initial weighted graph structure by calculating the attention weight coefficient between each node and the adjacent nodes, introduces the Laplace regularization term to constrain the smoothness of the attention distribution, enhances the propagation consistency of node features in the graph structure, and performs feature aggregation operations through the graph convolution layer, and finally outputs a node representation matrix containing global dependencies and contextual semantics.

6. The method for network charging data processing load balancing based on multi-objective collaborative optimization according to claim 1 is characterized in that: The multi-objective evolutionary algorithm is improved by adopting a two-stage guidance mechanism and an adaptive population compression strategy. In the first stage, a guided mutation operation based on the node state gradient is introduced during the initial population evolution process. In the second stage, a local perturbation mechanism based on the task scheduling deviation degree is introduced when the convergence is close. At the same time, the number of low-contributing individuals is adaptively compressed by calculating the individual convergence density of the population, thereby improving the convergence accuracy of the scheduling solution and the ability to maintain population diversity.

7. The method for load balancing of network charging data processing based on multi-objective collaborative optimization according to claim 1, characterized in that: The S2 specifically includes: S21, perform frame processing on the pre-processed operation data and construct a node operation status matrix where R i,j represents the running value of the i-th node on the j-th feature dimension, N represents the total number of nodes, and M represents the number of running features of each node; S22, input the node operation state matrix R into the feature extraction network, perform multi-layer residual convolution operation and self-attention fusion processing, and generate an intermediate feature tensor Where d is the number of dimensions of the extracted features, the convolution kernel size is set to k×k, the output of each convolution layer is processed by batch normalization and ReLU activation function, and the residual connection uses the "input-output addition" method to complete dimension alignment; S23, the intermediate feature tensor T is compressed by the feature channel to obtain the state feature vector set V = {v1, v2, ..., v N }, the channel compression adopts a one-dimensional fully connected layer transformation operation, and the state feature vector set satisfies the following optimization relationship: Among them, min(·) represents the minimum function, A i,j represents the explicit association metric between node i and node j, v i ,v j represents the state feature vector of the i,jth node, v i The transpose of , σ(·) represents the Sigmoid function, ||·|| 2 represents the square of the Euclidean norm, W represents the feature interaction weight matrix, R i Represents the row vector of node i in the original running state matrix, W r represents the state mapping weight matrix, and λ is the regularization factor, which is used to control the degree of preservation between node features and the original state.

8. The method for network charging data processing load balancing based on multi-objective collaborative optimization according to claim 1 is characterized in that: The S3 specifically includes: S31. Input the state feature vector set V into the graph attention neural network to construct a graph structure G = (U, E), where U represents the node set and E represents the edge set. The initial weight of each edge is set to a constant α. S32, for each node u i ∈U constructs H attention heads, each of which uses a different linear mapping matrix The input feature v i Mapping to subspace vector in S33. For each attention head h∈{1,2,…,H}, calculate the node u i and adjacent node u j The attention weight Use a scoring mechanism with a Laplace regularization term: Among them, || represents the vector splicing operation, is the weight vector of the h-th attention head, for a h The transpose of , μ is the regularization coefficient, represents the square of the Euclidean distance between node i and node j in the graph Laplace embedding space, represents the square of the Euclidean distance between node i and node k in the graph Laplace embedding space, is the graph Laplacian embedding matrix obtained by spectral decomposition, 1 represents the embedding dimension, Represents node u i The set of adjacent nodes, exp(·) represents the natural exponential function with e as the base, LeakyReLU(·) represents the activation function, denote subspace vectors respectively; S34. Concatenate the output results of all attention heads to obtain the multi-head fusion representation vector of each node S35. The fused nodes are represented as a set Input the graph convolution operation layer, perform feature aggregation, and output a node representation matrix. The node representation matrix has a node state expression that integrates global dependencies and contextual semantics: Among them, Z represents the node representation matrix, represents the adjacency matrix with self-loop, A is the weighted adjacency matrix constructed by attention weights, I N is the N×N identity matrix, for The degree matrix, W z is the weight matrix of the graph convolutional layer, σ(·) is the activation function, and the final output node representation matrix Z represents the node state expression with global dependency and contextual semantic fusion.

9. The method for network charging data processing load balancing based on multi-objective collaborative optimization according to claim 1, characterized in that: The S4 specifically includes: S41, according to the node representation matrix Z and the current set of tasks to be scheduled Where K represents the number of tasks to be scheduled, t k Represents the kth task to be scheduled and constructs a task scheduling vector set where x k,i ∈{0,1} represents task t k Is it assigned to node u? i , satisfying the constraints S42. Set three optimization objective functions in the task scheduling process, namely, minimizing the average task response time, minimizing the standard deviation of node load, and maximizing the number of tasks completed per unit time, and construct a joint optimization objective function: Among them, C k Represents task t k The complexity constant, P i Represents node u i The processing capacity per unit time, Represents node u i The total task load, represents the average load of all nodes, δ k,i ∈{0,1} represents task t k Can node u i If it can be completed, it is 1, otherwise it is 0. τ represents the length of the unit scheduling cycle, and ω1, ω2, and ω3 are the weighting factors of each objective function; S43. Using the task allocation results expressed in the objective function, a set of task scheduling vectors is obtained and used as input for the multi-objective evolutionary solution process.

10. The method for network charging data processing load balancing based on multi-objective collaborative optimization according to claim 1, characterized in that: The S5 specifically includes: S51, set the task scheduling vector As the initial population of the multi-objective evolutionary algorithm, each individual represents a task allocation scheme. Let the population size be Q, and the individual set is expressed as The algebraic number is g=0; S52. In the g-th generation of evolution, for each individual Using the node representation matrix Z={z1,z2,…,z N }, calculate the node mutation probability matrix: in, Indicates that task t in generation g k Select node u i The probability of being a candidate for mutation, ε is a constant to prevent the denominator from being zero, Represent different feature gradient vectors, Denotes that task t in generation g k Is it assigned to node u? i , Denotes that task t in generation g k Is it assigned to node u? j ; S53. In the process of population evolution in each generation, according to individual The non-inferior sorting results in the joint optimization objective function space are recorded in the sorting positions of the objective functions F1, F2, and F3, and the scheduling deviation function is constructed: in, Represent the optimal values ​​of the three optimization objective functions in the current population, Represent the maximum or minimum boundary values ​​of the three optimization objective functions, Represents an individual The scheduling deviation value in the joint optimization objective function space; S54. Before the evolutionary generation reaches G, continue to perform guided crossover, deviation perturbation, and environmental selection operations. After each round of iteration, according to the population density function, adaptively compress low-quality individuals with high density and large scheduling deviation values: in, represents the aggregation density between the qth individual and other individuals, distribution represents the different individuals of the gth generation, ε represents the density-aware adjustment parameter, exp(·) represents the natural exponential function with e as the base, and G represents the maximum number of iterations in the multi-objective evolutionary algorithm; S55. When algebra g=G, select a non-dominated solution set on the Pareto optimal front to construct an optimal task allocation matrix, and output the optimal task allocation matrix.

Images