Networking charging data cross-platform fusion and transmission method based on intelligent scheduling algorithm

Abstract

The invention discloses a networking charging data cross-platform fusion and transmission method based on an intelligent scheduling algorithm. The method comprises the following steps: S1, collecting and preprocessing networking platform charging data; s2, feature similarity is calculated, and an undirected graph with samples as nodes and similarity as edge weight is constructed; s3, calculating a standardized Laplacian matrix, extracting feature vectors, and clustering to generate feature cluster tags; s4, constructing a modal mapping matrix, mapping the feature vectors, and inputting the mapped feature vectors into a first convolution restricted Boltzmann machine to extract fusion features; s5, inputting the fusion features into a second convolution restricted Boltzmann machine, and combining graph structure sorting and constructing a weighted directed graph; s6, initializing a virtual scheduling factor in the graph, and searching an optimal transmission path by an ant colony algorithm; and S7, distributing fusion features according to paths, and realizing cross-platform efficient transmission of charging data. According to the method, spectral clustering, modal mapping, convolution extraction and intelligent scheduling are fused, and the multi-platform charging data fusion accuracy and transmission efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of cross-platform data fusion and transmission, and particularly to a method for cross-platform fusion and transmission of networked toll data based on an intelligent scheduling algorithm. Background Art

[0002] In modern traffic management systems, with the continuous development of intelligent technologies, the networked toll system has become a core technology widely used in traffic facilities such as highways, urban roads, and parking lots. The networked toll system can achieve efficient toll management, reduce manual intervention, and improve traffic efficiency by collecting and transmitting traffic toll data. However, with the increase in traffic volume, the intelligence of the toll system and the complexity of data management have also increased. How to efficiently process, fuse, and transmit data from different platforms and terminals has become an urgent problem to be solved.

[0003] Currently, there are multiple problems in the existing networked toll data processing technologies. First of all, the existing networked toll systems are usually isolated and adopt a toll collection method based on a single platform, making it difficult to achieve cross-platform data sharing and integration. The data structures, formats, protocols, and storage methods of different toll platforms are different, resulting in a lack of unified standards and effective compatibility between data. This situation makes the cross-platform fusion and efficient transmission of data between multiple toll platforms extremely complex, unable to achieve real-time updates and unified management, bringing great work pressure to traffic management departments.

[0004] Secondly, the existing toll data transmission technologies have significant performance bottlenecks when dealing with large-scale data. Especially in the face of real-time data transmission and large amounts of data, traditional transmission methods are difficult to meet the requirements in terms of transmission speed and data integrity. Although common transmission protocols can ensure the reliability of data, they cannot provide sufficient flexibility and efficiency under different network conditions. Especially in the case of network congestion or bandwidth limitations, the existing transmission methods often cannot maintain efficient data flow, resulting in data loss or delay, seriously affecting the stability and real-time performance of the system.

[0005] In addition, the existing networked toll systems often rely on traditional scheduling algorithms for data processing and transmission, but these scheduling algorithms are usually too simple and lack the ability to adapt to complex situations. For example, when dealing with data transmission between multiple platforms, traditional scheduling algorithms do not fully consider factors such as network status, device load, and data priority, resulting in low scheduling efficiency and unable to ensure optimal resource allocation in a changing environment. This limitation makes the existing system show low efficiency and large delays when facing the fusion and data transmission of different toll platforms, and cannot meet the requirements of modern intelligent traffic systems for efficient and real-time data processing.

[0006] Therefore, how to provide a cross-platform fusion and transmission method for networked toll data based on an intelligent scheduling algorithm to solve problems such as difficult data fusion, low transmission efficiency, and insufficient intelligent scheduling in the prior art is the core problem that those skilled in the art urgently need to solve. Summary of the Invention

[0007] An object of the present invention is to propose a cross-platform fusion and transmission method for networked toll data based on an intelligent scheduling algorithm. The present invention fully integrates algorithms such as spectral clustering analysis, modal mapping modeling, convolutional restricted Boltzmann machine feature extraction, and ant colony optimization path search, and details how to achieve efficient and standardized data fusion and intelligent path selection under the conditions of complex multi-platform toll data structures and dynamically changing network states. This method has the advantages of high fusion accuracy, strong scheduling intelligence, high cross-platform transmission efficiency, and strong system scalability.

[0008] The cross-platform fusion and transmission method for networked toll data based on an intelligent scheduling algorithm according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect toll data from the networked platform and perform preprocessing;

[0010] S2. Calculate the feature similarity of the preprocessed toll data, and construct an undirected graph structure based on the feature similarity. Nodes in the graph represent toll data samples, and edge weights represent the feature similarity between samples;

[0011] S3. Calculate the normalized Laplacian matrix according to the undirected graph structure, extract the low-order eigenvectors corresponding to the first k smallest eigenvalues, and perform clustering on the low-order eigenvectors using the k-means algorithm to obtain feature cluster labels representing different data modalities;

[0012] S4. Construct a modal mapping matrix based on the feature cluster labels, map the low-order eigenvectors to standard modal vectors, and input them into the first convolutional restricted Boltzmann machine to perform feature extraction and deep fusion operations, and output the fused high-order eigenvectors;

[0013] S5. Input the high-order eigenvectors into the second convolutional restricted Boltzmann machine, and combine with the normalized Laplacian matrix to perform data scheduling priority ranking, and construct a weighted directed graph according to the ranking results;

[0014] S6. Initialize multiple virtual scheduling factors in the weighted directed graph, and use the ant colony optimization algorithm to iteratively search for the path sequence with the minimum cost to determine the optimal cross-platform transmission path;

[0015] S7. Perform real-time distribution of the fused high-order eigenvectors according to the optimal cross-platform transmission path to achieve efficient and stable transmission of networked toll data among multiple platforms.

[0016] Optionally, the toll data includes vehicle information, payment records, toll timestamps, lane numbers, and corresponding image recognition results, and the preprocessing includes missing value filling, field standardization, time format unification, and noise data elimination operations.

[0017] Optionally, S2 specifically includes:

[0018] S21. Map each piece of preprocessed toll data into a feature vector where x i is the feature vector of the i-th piece of toll data, n is the total number of fields included in each piece of toll data, and x ij is the normalized value of the i-th piece of toll data in the j-th field, and all data samples form a feature matrix m is the total number of toll data samples;

[0019] S22. Based on the feature vector and the feature matrix, calculate the comprehensive feature distance between two samples based on the weighted feature kernel function:

[0020]

[0021] where d ij is the comprehensive feature distance between the i-th piece of toll data and the j-th piece of toll data, and α k is the feature weight coefficient of the k-th field, satisfying x ik , x jk respectively represent the normalized feature values of the i-th and j-th data in the k-th field, is the mean value of the k-th field in all toll data samples, that is ε is a positive real number used to avoid the denominator being zero;

[0022] S23. Construct a feature similarity matrix W = [w ij according to the comprehensive feature distance. Taking all toll data as nodes and the non-zero elements in the feature similarity matrix as edge weights, construct an undirected graph G = (V, E), where:

[0023] V = {v1, v2,..., v m} represents the node set composed of toll data samples;

[0024] represents the edge set between nodes;

[0025] For any v i , v j ∈ V, if w ij > 0, then there exists an edge e ij ∈ E, and the weight value corresponding to this edge is w ij .

[0026] Optionally, the feature similarity matrix W = [w ij , and the similarity weights are generated using a Gaussian kernel function and normalized:

[0027]

[0028] where exp(·) is the natural exponential function with base e, d ij is the comprehensive feature distance between the i-th charging data and the j-th charging data, σ is a hyperparameter for adjusting the smoothness of the kernel function, d ij is the comprehensive feature distance between the i-th charging data and the j-th charging data, w ij is the similarity weight between the charging data x i and x j , ensuring

[0029] Optionally, the S3 specifically includes:

[0030] S31. Based on the feature similarity matrix W = [w ij , calculate the degree matrix D = diag(d1, d2, …, d m ), where m is the total number of charging data samples, w ij is the similarity weight between the charging data x i and x j , diag(·) is a diagonal matrix, the i-th diagonal element is d i , and represents the sum of the connection weights of the node v i ;

[0031] S32. Based on the feature similarity matrix W and the degree matrix D, calculate the normalized graph Laplacian matrix:

[0032] L sym = I - D -1 / 2 WD -1 / 2 ;

[0033] where L sym is the normalized graph Laplacian matrix, I is the identity matrix, D -1 / 2 is the element-wise negative square root diagonal matrix of the degree matrix D, used to balance the weight influence between different nodes;

[0034] S33. Perform eigen decomposition on the Laplacian matrix, extract the eigenvectors corresponding to the smallest first k eigenvalues, and construct the eigenvector matrix U = [u1, u2, …, u k , where k is the preset number of clusters and also the dimension of the low-dimensional subspace, u lis the l-th eigenvector, where l ∈ {1, 2,..., k};

[0035] Normalize the eigenvector matrix U row by row to obtain the normalized eigenmatrix where z i is the representation vector of the i-th sample in the low-dimensional feature subspace;

[0036] S34. Based on the normalized eigenmatrix Z, cluster all node representation vectors, and use the k-means algorithm to minimize the following objective function:

[0037]

[0038] where J is the clustering loss function, c l(i) is the cluster center vector to which the i-th sample belongs, is the sample set of the p-th cluster, m is the total number of toll data samples, and |C p | is the number of samples in the cluster C p and λ is the balance hyperparameter that adjusts the clustering compactness and separation;

[0039] S35. Use the clustering result of each sample z i as the feature cluster label for the subsequent modal mapping process.

[0040] Optionally, the S4 specifically includes:

[0041] S41. Based on the feature cluster labels of each toll data sample, construct the modal mapping matrix M = [m ij , where m ij = 1 indicates that the i-th toll data belongs to the j-th feature cluster, otherwise m ij = 0;

[0042] Transform the normalized eigenmatrix into the standard modal vector matrix through the mapping relationship:

[0043]

[0044] where m is the total number of toll data samples, is the standard modal vector matrix, R is the feature cluster weight mapping matrix, Q is the feature compression matrix, γ is the modal enhancement coefficient, and d is the standard modal dimension;

[0045] S42. Reconstruct the standard modal vector matrix into the tensor representation as the input of the first convolutional restricted Boltzmann machine, and define the set of learnable convolutional kernels as and the hidden layer activation output of the first convolutional restricted Boltzmann machine as:

[0046]

[0047] Among them, h is the number of convolution kernels, and r is the spatial size of each convolution kernel. is the value of the hidden activation unit at the (j, t) position under the action of the i-th sample on the convolution kernel K t σ(·) is the Sigmoid activation function. is the double summation within the two-dimensional convolution window, and K t (p, q) is the weight coefficient of the t-th convolution kernel at the position (p, q). is the standard modal vector matrix The set of sample tensors obtained by reconstruction. is the i-th sample tensor The element value at the position (i + p - 1, j + q - 1), and b t is the convolution kernel bias term;

[0048] S43. Aggregate the activation outputs of all samples under the convolution kernel to form a three-dimensional tensor Perform a fully connected compression on the convolution output of each sample to generate a set of high-order feature vectors The set of high-order feature vectors is used as the input of the subsequent second convolutional restricted Boltzmann machine.

[0049] Optionally, the first convolutional restricted Boltzmann machine adopts a shared weight structure, including no less than 3 groups of convolution kernels. The size of each group of convolution kernels is r×r, where r is the spatial size of each convolution kernel and r≥5. The convolution operation is a sliding convolution with a step size of 1 on the standard modal vector tensor representation, and the activation function is The convolution output is processed by a pooling layer for max pooling to generate local invariant features. The local invariant features are converted into a one-dimensional vector through a flattening operation and used as part of the high-order feature vectors.

[0050] Optionally, the specific steps of S5 are as follows:

[0051] S51. Use the set of high-order feature vectors as the input of the second convolutional restricted Boltzmann machine. First, reconstruct each high-order feature vector v i into a tensor and combine it with the set of convolution kernels of the second convolutional restricted Boltzmann machine to obtain Define the convolution activation output as:

[0052]

[0053] Among them, q is the number of convolution kernels, and r ′ is the convolution kernel size. is the activation value at the (u, v) position under the action of the j-th convolution kernel for the i-th sample, is the i-th sample tensor The element value at the position (u + a - 1, v + b - 1), G j (a, b) is the weight coefficient of the j-th convolution kernel at the position (a, b), β j is the bias term of the j-th convolution kernel, and σ(·) is the Sigmoid activation function;

[0054] S52. Expand all convolution outputs of each sample into a one-dimensional vector r i , and for the vector r i Perform priority score calculation:

[0055]

[0056] where s i is the scheduling priority score of the i-th sample, g is the length of the one-dimensional vector after convolution expansion, f is the dimension of the high-order feature vector, δ1, δ2, δ3 are the weighted coefficients of different scoring factors, and r it is the t-th dimension component of the i-th sample in the convolution-expanded vector r i , v ik is the k-th component value of the high-order feature vector v i , log2(·) is the logarithmic function, and max is the maximum value function;

[0057] S53. Construct a weighted directed graph G′ = (V′, E′, W′) according to the priority score of each sample and the real-time network state, where the node set V′ = {v′1, v′2, …, v′ m} is the toll data sample, the edge set E′ is the connection relationship between samples with a transmission path, and the edge weight matrix W′ = [w′ ij , and the edge weight calculation formula is:

[0058]

[0059] where w′ ij is the transmission priority weight, φ ij is the bandwidth utilization rate, λ ij is the observed real-time network delay, and μ j is the current load index.

[0060] Optionally, it is characterized in that during the training of the second convolutional restricted Boltzmann machine, the convolution kernel weight G j needs to be updated:

[0061]

[0062] where ΔGj is the updated value of the weight matrix of the j-th convolutional kernel, η is the learning rate, is the expected value of the product of the input and output during the forward propagation process, is the reconstruction error, that is, the expected value of the product of the output and input of the reconstructed data through backpropagation.

[0063] Optionally, the S6 specifically includes:

[0064] S61. In the weighted directed graph G′=(V′, E′, W′), initialize the set of ant individuals of the ant colony optimization algorithm. Each ant individual starts from the starting node in the graph for path search. Define the transition probability of the d-th ant from node i to node j as:

[0065]

[0066] where represents the probability that the d-th ant transfers from node i to node j at time t, τ ij (t) is the current pheromone concentration on edge e′ ij η ij (t) = 1 / w′ ij is the path heuristic function value, w′ ij is the transmission priority weight, is the set of adjacent nodes that the d-th ant can reach from node i at the current moment, α and β are respectively the pheromone intensity and heuristic factor weight parameters;

[0067] S62. After all ants complete path traversal, define the total cost function of each path as:

[0068]

[0069] where, C d is the total path cost corresponding to the d-th ant, is the ordered edge set in the path traversed by the d-th ant, s i is the scheduling priority score, w′ ij is the transmission priority weight, ∈ ij is the instantaneous packet loss rate on edge e′ ij ψ j is the network congestion degree of the target node j, ρ is the path cost smoothing factor, δ is the regularization term balance factor;

[0070] S63. According to the total cost function C n sort each path, select the path with the minimum total cost as the optimal path, and update the pheromone concentration on each path edge according to the following pheromone update rule:

[0071]

[0072] Among them, τ ij (t + 1) is the pheromone concentration on edge e′ at time t + 1 ij τ ij (t) is the pheromone concentration on edge e′ at time t ij ω is the pheromone evaporation coefficient, is the contribution of the ant to the pheromone on edge e′ ij Q is the pheromone release constant, and N is the number of ant individuals.

[0073] The beneficial effects of the present invention are as follows:

[0074] First of all, by introducing the spectral clustering algorithm and the modal mapping mechanism, the present invention solves the fusion problem caused by large differences in data structures and inconsistent field types between different networked toll platforms. By constructing a standard modal mapping matrix and combining with a convolutional restricted Boltzmann machine for deep feature extraction, the unified modeling and semantic alignment of multi-source heterogeneous toll data are realized, fundamentally improving the accuracy and robustness of cross-platform data fusion, and avoiding the low efficiency and high error rate caused by relying on manual configuration or static mapping in the prior art.

[0075] Secondly, the present invention introduces a two-stage convolutional feature fusion and priority scoring mechanism at the scheduling layer, incorporating the semantic importance of the data itself, the historical transmission adaptability, and the platform load status into the priority calculation process, enabling each toll data to be dynamically assigned a reasonable transmission scheduling level. Cooperating with the weighted directed graph structure constructed based on the real-time network status, it can effectively reflect the reachability and transmission cost of the current network topology, providing true and accurate weight support for subsequent path optimization, and significantly improving the intelligent level of scheduling decisions.

[0076] Finally, the present invention uses an improved ant colony optimization algorithm for path search and pheromone update, considering multiple factors such as delay, congestion, and priority during the path selection process, being able to adaptively adjust the transmission path, avoiding network bottlenecks and faulty nodes, and realizing the efficient, stable, and low-latency transmission of toll data between different platforms. The entire method constructs a complete intelligent scheduling closed-loop from data fusion to scheduling scoring and then to path optimization, significantly improving the data processing efficiency and system response ability of the networked toll system in a multi-platform environment, and having good practicality and promotion value. Description of the Drawings

[0077] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, but do not constitute a limitation to the present invention. In the drawings:

[0078] Figure 1Flow chart of the cross-platform fusion and transmission method of networked toll collection data based on intelligent scheduling algorithm proposed by the present invention;

[0079] Figure 2 Flow chart of feature similarity calculation and undirected graph construction of the cross-platform fusion and transmission method of networked toll collection data based on intelligent scheduling algorithm proposed by the present invention;

[0080] Figure 3 Flow chart of Laplacian matrix and spectral clustering of the cross-platform fusion and transmission method of networked toll collection data based on intelligent scheduling algorithm proposed by the present invention. Detailed implementation manners

[0081] The present invention will be further described in detail below with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0082] Refer to Figures 1 - 3 , the cross-platform fusion and transmission method of networked toll collection data based on intelligent scheduling algorithm includes the following steps:

[0083] S1. Collect toll collection data of the networked platform and perform preprocessing;

[0084] S2. Calculate the feature similarity of the preprocessed toll collection data, and construct an undirected graph structure based on the feature similarity. The nodes in the graph represent toll collection data samples, and the edge weights represent the feature similarity between samples;

[0085] S3. Calculate the normalized Laplacian matrix according to the undirected graph structure, extract the low-order eigenvectors corresponding to the first k smallest eigenvalues, and perform clustering on the low-order eigenvectors using the k-means algorithm to obtain feature cluster labels representing different data modalities;

[0086] S4. Construct a modal mapping matrix based on the feature cluster labels, map the low-order eigenvectors to standard modal vectors, and input them into the first convolutional restricted Boltzmann machine to perform feature extraction and depth fusion operations, and output the fused high-order eigenvectors;

[0087] S5. Input the high-order eigenvectors into the second convolutional restricted Boltzmann machine, and combine with the normalized Laplacian matrix to perform data scheduling priority sorting, and construct a weighted directed graph according to the sorting result;

[0088] S6. Initialize multiple virtual scheduling factors in the weighted directed graph, and use the ant colony optimization algorithm to iteratively search for the path sequence with the minimum cost to determine the optimal cross-platform transmission path;

[0089] S7. Perform real-time distribution of the fused high-order eigenvectors according to the optimal cross-platform transmission path to achieve efficient and stable transmission of networked toll collection data among multiple platforms.

[0090] The present invention realizes the efficient fusion and intelligent transmission of networked toll data among multiple platforms by constructing a complete processing flow from toll data collection, feature similarity calculation, spectral clustering dimensionality reduction, modal standardization fusion, two-stage convolutional feature extraction, priority scoring, weighted graph construction to path scheduling optimization. This method takes into account the complexity of data structure, the dynamics of network environment and platform heterogeneity, improves the efficiency of data collaborative processing and the stability of system operation, and is particularly suitable for the data fusion and transmission scheduling requirements in multi-source toll scenarios.

[0091] In this embodiment, the toll data includes vehicle information, payment records, toll timestamps, lane numbers and corresponding image recognition results, and the preprocessing includes missing value filling, field standardization, time format unification and noise data elimination operations.

[0092] The present invention effectively ensures the structural consistency and quality controllability of the input of raw data by clearly defining the keyword field information included in the toll data, such as vehicle information, payment records, image recognition results, etc., and standardizing the preprocessing process, including operations such as field standardization, time format unification and noise elimination, provides a stable and reliable input basis for subsequent modeling and fusion, and improves the preprocessing efficiency and compatibility of the system in a large-scale data environment.

[0093] In this embodiment, the specific steps of S2 include:

[0094] S21. Map each piece of preprocessed toll data into a feature vector where x i is the feature vector of the i-th piece of toll data, n is the total number of fields included in each piece of toll data, x ij is the normalized value of the i-th piece of toll data in the j-th field, and all data samples form a feature matrix X = m is the total number of toll data samples;

[0095] S22. Based on the feature vectors and the feature matrix, calculate the comprehensive feature distance between two samples based on the weighted feature kernel function:

[0096]

[0097] where d ij is the comprehensive feature distance between the i-th piece of toll data and the j-th piece of toll data, α k is the feature weight coefficient of the k-th field, satisfying x ik , x jk respectively represent the normalized feature values of the i-th and j-th data in the k-th field, is the mean of the k-th field in all toll data samples, i.e., ε is a positive real number used to avoid a zero denominator;

[0098] S23. Construct a feature similarity matrix W = [w ij based on the comprehensive feature distance. Using all toll data as nodes and the non-zero elements in the feature similarity matrix as edge weights, construct an undirected graph G = (V, E), where:

[0099] V = {v1, v2, …, v m}, representing the set of nodes composed of toll data samples;

[0100] represents the set of edges between nodes;

[0101] For any v i , v j ∈V, if w ij > 0, then there exists an edge e ij ∈E, and the weight corresponding to this edge is w ij .

[0102] The present invention defines a method for constructing a feature vector and introduces a weighted feature kernel function to calculate the comprehensive distance between samples, enabling the system to more accurately express the different impacts of each field on data similarity and improving the accuracy and robustness of feature measurement. At the same time, by constructing a feature similarity matrix based on the Gaussian kernel function and converting it into an undirected graph structure, an explicit representation of the data space structure is realized, providing a solid foundation for subsequent graph structure processing and clustering analysis.

[0103] In this embodiment, the feature similarity matrix W = [w ij generates similarity weights using the Gaussian kernel function and performs normalization processing:

[0104]

[0105] where exp(·) is the natural exponential function with base e, d ij is the comprehensive feature distance between the i-th toll data and the j-th toll data, σ is a hyperparameter for adjusting the smoothness of the kernel function, d ij is the comprehensive feature distance between the i-th toll data and the j-th toll data, w ij is the similarity weight between toll data x i and x j , ensuring

[0106] In the present invention, by introducing the Gaussian kernel function in the calculation of feature similarity, the weighted Euclidean distance between samples is mapped into a similarity weight in the form of exponential decay, realizing the accurate characterization of the non-linear correlation between toll data samples. In addition, the feature similarity matrix generated after normalization has the global balance property that the sum of the numerical values of the row vectors and column vectors is consistent, ensuring the stability and clustering robustness of the subsequent spectral clustering process in the similarity space, thereby effectively improving the accuracy of modal recognition and the overall feature expression ability of the system.

[0107] In this embodiment, step S3 specifically includes:

[0108] S31. Based on the feature similarity matrix W = [w ij , calculate the degree matrix D = diag(d1, d2,..., d m ), where m is the total number of toll data samples, w ij is the similarity weight between toll data x i and x j , diag(·) is a diagonal matrix, the i-th diagonal element is d i , and represents the sum of the connection weights of node v i ;

[0109] S32. Based on the feature similarity matrix W and the degree matrix D, calculate the normalized graph Laplacian matrix:

[0110] L sym = I - D -1 / 2 WD -1 / 2 ;

[0111] Among them, L sym is the standardized graph Laplacian matrix, I is the identity matrix, D -1 / 2 is the element-wise negative square root diagonal matrix of the degree matrix D, used to balance the weight influence between different nodes;

[0112] S33. Perform eigenvalue decomposition on the Laplacian matrix, extract the eigenvectors corresponding to the smallest first k eigenvalues, and construct the eigenvector matrix U = [u1, u2,..., u k , where k is the preset number of clusters and also the dimension of the low-dimensional subspace, u l is the l-th eigenvector, l ∈ {1, 2,..., k};

[0113] Normalize the eigenvector matrix U row by row to obtain the normalized eigenmatrix where z i is the representation vector of the i-th sample in the low-dimensional feature subspace;

[0114] S34. Based on the normalized feature matrix Z, cluster all node representation vectors, and use the k-means algorithm to minimize the following objective function:

[0115]

[0116] where J is the clustering loss function, and c l(i) is the cluster center vector to which the i-th sample belongs, is the set of samples in the p-th cluster, m is the total number of toll data samples, and |C p | is the number of samples in the cluster C p , and λ is the balance hyperparameter that adjusts the clustering compactness and separation;

[0117] S35. Use the clustering result of each sample z i as the feature cluster label for the subsequent modal mapping process.

[0118] The present invention uses a standardized Laplacian matrix for graph modeling, combined with eigen-decomposition and the k-means clustering algorithm, to effectively achieve modal partitioning of toll data in a low-dimensional feature subspace. By designing a clustering loss function and introducing a balance mechanism for clustering compactness and separation, the modal recognition accuracy is improved, making the clustering results more discriminative, which is beneficial to the accurate execution of subsequent modal mapping and feature fusion, and reduces the interference of cross-platform data structure differences on the fusion effect.

[0119] In this embodiment, the S4 specifically includes:

[0120] S41. Based on the feature cluster labels of each toll data sample, construct a modal mapping matrix M = [m ij , where m ij = 1 indicates that the i-th toll data belongs to the j-th feature cluster, otherwise m ij = 0;

[0121] Transform the normalized feature matrix into a standard modal vector matrix through the mapping relationship:

[0122]

[0123] where m is the total number of toll data samples, is the standard modal vector matrix, R is the feature cluster weight mapping matrix, Q is the feature compression matrix, γ is the modal enhancement coefficient, and d is the standard modal dimension;

[0124] S42. Reconstruct the standard modal vector matrix into a tensor representation as the input of the first convolutional restricted Boltzmann machine, and define the set of learnable convolutional kernels as and the hidden layer activation output of the first convolutional restricted Boltzmann machine is:

[0125]

[0126] where h is the number of convolutional kernels, r is the spatial size of each convolutional kernel, is the value of the hidden activation unit at the (j, t) position under the action of the i-th sample on the convolutional kernel K t σ(·) is the Sigmoid activation function, is the double summation within the two-dimensional convolutional window, K t (p, q) are the weight coefficients of the t-th convolutional kernel at the position (p, q), is the standard modal vector matrix the set of sample tensors obtained by reconstruction, is the i-th sample tensor the element value at the position (i + p - 1, j + q - 1), b t is the convolutional kernel bias term;

[0127] S43. Aggregate the activation outputs of all samples under the convolutional kernel to form a three-dimensional tensor Fully connect and compress the convolutional output of each sample to generate a set of high-order feature vectors The set of high-order feature vectors is used as the input of the subsequent second convolutional restricted Boltzmann machine.

[0128] In the modal fusion stage of the present invention, by constructing a modal mapping matrix, the clustering result is mapped to the standard modal space, and combined with the first convolutional restricted Boltzmann machine for multi-channel convolutional feature extraction and deep fusion, which improves the semantic alignment ability of the data at the feature layer. This method enables the system to adaptively extract discriminative fusion features when processing data sources with structural heterogeneity and semantic inconsistency, improving the accuracy and robustness of cross-platform data integration.

[0129] In this embodiment, the first convolutional restricted Boltzmann machine adopts a shared weight structure, includes no less than 3 groups of convolutional kernels, the size of each group of convolutional kernels is r×r, where r is the spatial size of each convolutional kernel, and r≥5. The convolutional operation is a sliding convolution with a step size of 1 on the standard modal vector tensor representation, and the activation function is After the convolutional output is subjected to max-pooling processing by the pooling layer, local invariant features are generated. The local invariant features are converted into one-dimensional vectors through a flattening operation and used as part of the high-order feature vectors.

[0130] The present invention adopts a multi-core convolution mechanism with shared weights in the convolution extraction structure, and performs max pooling and local invariant feature extraction operations on the standard modal tensor, significantly improving the generalization ability and representation ability of the system in the high-dimensional feature space. The combination of multi-scale feature fusion and local stability enables the finally output high-order feature vector to have stronger abstract expression ability, effectively supporting the high-precision input requirements for subsequent scheduling scoring and path selection.

[0131] In this embodiment, the S5 specifically includes:

[0132] S51. Take the high-order feature vector set as the input of the second convolutional restricted Boltzmann machine. First, reconstruct each high-order feature vector v i into a tensor and combine it with the convolution kernel set of the second convolutional restricted Boltzmann machine as Define the convolution activation output as:

[0133]

[0134] where q is the number of convolution kernels, r′ is the convolution kernel size, is the activation value at the (u, v) position under the action of the jth convolution kernel for the ith sample, is the element value of the ith sample tensor at the position (u + a - 1, v + b - 1), G j (a, b) is the weight coefficient of the jth convolution kernel at the position (a, b), β j is the bias term of the jth convolution kernel, and σ(·) is the Sigmoid activation function;

[0135] S52. Expand all convolution outputs of each sample into a one-dimensional vector r i , and calculate the priority score for the vector r i :

[0136]

[0137] where s i is the scheduling priority score of the ith sample, g is the length of the one-dimensional vector after convolution, f is the dimension of the high-order feature vector, δ1, δ2, δ3 are the weighted coefficients of different scoring factors, r it is the tth dimension component of the ith sample in the convolution-expanded vector r i , v ik is the kth component value of the high-order feature vector v i , log2(·) is the logarithmic function, and max is the maximum value function;

[0138] S53. Construct a weighted directed graph G′=(V′, E′, W′) according to the priority scores of each sample and the real-time network status, where the node set V′={v′1, v′2, …, v′ m} is the toll data sample, the edge set E′ is the connection relationship with transmission paths between samples, and the edge weight matrix W′=[w′ ij , and the edge weight calculation formula is:

[0139]

[0140] where w′ ij is the transmission priority weight, φ ij is the bandwidth utilization rate, λ ij is the observed real-time network delay, and μ j is the current load index.

[0141] In the present invention, a second convolutional restricted Boltzmann machine is introduced in the scheduling scoring stage, and semantic analysis and priority scoring are combined with the high-order feature tensor structure, further improving the discrimination ability of data scheduling scoring. By designing a weighted scoring function and combining convolutional activation, nonlinear transformation, and the mean of original features, the system can adaptively sort the scheduling priorities for data features in different scenarios and on different platforms, thereby optimizing the rationality of task scheduling and the accuracy of transmission paths.

[0142] In this embodiment, it is characterized in that during the training of the second convolutional restricted Boltzmann machine, the convolutional kernel weight G j needs to be updated:

[0143]

[0144] where ΔG j is the update value of the jth convolutional kernel weight matrix, η is the learning rate, is the expected value of the product of the input and output during the forward propagation process, is the reconstruction error, that is, the expected value of the product of the output and input of the reconstructed data through backpropagation.

[0145] In the present invention, by defining the update formula of the convolutional kernel weight of the convolutional restricted Boltzmann machine and introducing the expected difference between forward propagation and backward reconstruction as the gradient basis, it is ensured that the parameters converge stably during the training process, effectively improving the generalization ability and learning efficiency of the model in a multi-platform environment. This training mechanism combines the semantic features of actual toll data, which helps the convolutional kernel quickly capture high-value structural patterns and enhances the model's ability to learn deep semantic features.

[0146] In this embodiment, the S6 specifically includes:

[0147] S61. In the weighted directed graph G′ = (V′, E′, W′), initialize the set of ant individuals for the ant colony optimization algorithm. Each ant individual starts from the starting node in the graph to perform path search. Define the transition probability of the d-th ant from node i to node j as:

[0148]

[0149] where represents the probability that the d-th ant transfers from node i to node j at time t, τ ij (t) is the current pheromone concentration on the edge e′ ij , η ij (t) = 1 / w′ ij is the value of the path heuristic function, w′ ij is the transmission priority weight, is the set of adjacent nodes that the d-th ant can reach from node i at the current time, and α and β are respectively the pheromone intensity and the heuristic factor weight parameters;

[0150] S62. After all ants complete path traversal, define the total cost function of each path as:

[0151]

[0152] where C d is the total cost of the path corresponding to the d-th ant, is the ordered edge set in the path traversed by the d-th ant, s i is the scheduling priority score, w′ ij is the transmission priority weight, ∈ ij is the instantaneous packet loss rate on the edge e′ ij , ψ j is the network congestion degree of the target node j, ρ is the path cost smoothing factor, and δ is the regularization term balance factor;

[0153] S63. According to the total cost function C n , sort each path, select the path with the minimum total cost as the optimal path, and update the pheromone concentration on each path edge according to the following pheromone update rule:

[0154]

[0155] where τ ij (t + 1) is the pheromone concentration on the edge e′ ij at time t + 1, τ ij (t) is the pheromone concentration on the edge e′ ij at time t, ω is the pheromone evaporation coefficient, is the amount of pheromone deposited by the ant on the edge e′ ijThe contribution of pheromone, where Q is the pheromone release constant and N is the number of ant individuals.

[0156] The present invention introduces an ant colony optimization algorithm in the scheduling path selection, defines a transfer probability function and a total path cost function, integrates multiple factors such as scheduling score, bandwidth status, delay index and network congestion degree, and constructs an optimal path search mechanism in a real scenario. By introducing a pheromone update strategy and a cost feedback mechanism, the system can achieve global optimal path learning based on historical experience, improve the stability and transmission efficiency of the scheduling path, and is applicable to high-dynamic scenarios of multi-platform collaborative transmission tasks.

[0157] Example 1:

[0158] To verify the feasibility of the present invention in implementation, the present invention is applied to the cross-platform data fusion and transmission transformation project of the networked toll collection system of the expressway group company under the Provincial Transportation Management Bureau. The company currently has 6 expressways with a total of 48 toll stations. The networked toll collection platform is constructed by three different suppliers, respectively using different data structures, communication protocols and scheduling mechanisms. For a long time, there have been problems such as inability to unify the data fusion between platforms, low transmission efficiency, and unstable scheduling paths, which have seriously restricted the further improvement of road network collaborative scheduling and toll management intelligence.

[0159] In this scenario, the three systems are respectively responsible for data collection and settlement of different sections. Among them, Platform A uses the JSON format with a fixed field structure, Platform B uses the semi-structured XML format, and Platform C directly uploads compressed binary data messages. The field names of the three platforms are not unified, and some fields are missing. The original system adapts the data through manual scripts, with low processing efficiency and high error rate. Originally, the toll data transmission between the three platforms relied on FTP + manual scheduling, and only about 12,000 data could be synchronized per hour, and problems such as packet loss, delay, and data coverage were likely to occur during peak periods. To solve the above problems, the method of the present invention is introduced to unify and fuse the toll data of the three platforms and achieve efficient cross-platform transmission through intelligent scheduling.

[0160] In practical applications, the project team first accessed and preprocessed the charging data of the three platforms, and uniformly extracted 22 fields such as vehicle information (license plate number, vehicle type, vehicle speed), payment information (payment method, amount, time), channel information (in / out station code, station name), and recognition image data (lane image summary hash, target detection result). After performing missing value filling, time format standardization, and feature normalization operations, a sample feature matrix was constructed. Subsequently, the similarity between samples was calculated using the Gaussian kernel function to obtain a similarity matrix, and data modality partitioning was performed through spectral clustering. The results show that the system can automatically identify 7 different data modalities during the fusion stage, with an accuracy rate of 96.3%, which is higher than that of traditional manual rules (about 78.5%) and single feature matching methods (about 85.2%).

[0161] The fused data is subjected to deep feature extraction through a convolutional restricted Boltzmann machine to generate high-order fusion vectors, and combined with the feedback metrics of the target platform (such as historical success rate, latency, bandwidth), scheduling scores are calculated through a second convolutional restricted Boltzmann machine, and finally a multi-factor weighted transmission directed graph is constructed. An improved ant colony algorithm is used to search for paths on this graph, and the system dynamically adjusts the transmission path according to the network status in different time periods, thereby achieving efficient and stable cross-platform data transmission.

[0162] After the system was deployed, during the continuous 7-day operation (from October 10th to October 16th, 2024), the total covered data volume reached 2,376,450 records, and the system operation metrics were significantly improved. The original average hourly data processing volume was 12,000 records, which was increased to 49,300 records, with an increase rate of 310%; the average fusion time per single record decreased from the original 1.26 seconds to 0.38 seconds; the packet latency during the peak period decreased from the original maximum of 6.2 seconds to 1.1 seconds, and the packet loss rate decreased from 1.7% to 0.02%; the average switching response latency of the scheduling paths among the 3 platforms decreased from 3.9 seconds to 1.2 seconds, effectively alleviating the transmission bottleneck in the case of network congestion.

[0163] The system also synchronously recorded the comparison between the scheduling score and the path cost function. During the morning and evening rush hours every day, the average cost of the automatically scheduled paths decreased by 42.5%. Among them, in the case of a sudden large traffic volume (52,000 vehicle trips per hour) on October 12th, the traditional scheduling system had multi-node blockages and data accumulation for more than 7 minutes, while the system adopting the technical solution of the present invention automatically reconstructed the path and completed data transfer. The maximum hourly forwarding volume reached 72,800 records, and the synchronization rate between platforms was maintained above 99.98%, ensuring business continuity and settlement accuracy.

[0164] The following table shows the performance of the key technical indicators of the present invention in practical applications:

[0165] Table 1 Comparative Experiment Table of Multi-platform Networked Toll Collection Data Fusion Transmission Performance

[0166]

[0167] In summary, in the complex traffic networked toll collection scenario, the present invention realizes the deep fusion and efficient transmission of multi-platform toll collection data through an intelligent scheduling algorithm, which not only significantly improves the data processing efficiency and system response ability, but also greatly enhances the overall operation stability and intelligent level, and has significant application and promotion value.

[0168] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A cross-platform fusion and transmission method for network toll collection data based on an intelligent scheduling algorithm, characterized in that, It includes the following steps: S1. Collect the charging data of the networking platform and perform preprocessing; S2. Calculate the feature similarity of the preprocessed charging data, and construct an undirected graph structure based on the feature similarity. The nodes in the graph represent the charging data samples, and the edge weights represent the feature similarity between the samples; S3. Calculate the normalized Laplacian matrix according to the undirected graph structure, extract the low-order eigenvectors corresponding to the first k minimum eigenvalues, and perform clustering on the low-order eigenvectors using the k-means algorithm to obtain the feature cluster labels representing different data modalities; S4. Construct a modality mapping matrix based on the feature cluster labels, map the low-order eigenvectors to standard modality vectors, and input them into the first convolutional restricted Boltzmann machine to perform feature extraction and depth fusion operations, and output the fused high-order eigenvectors; S5. Input the high-order eigenvectors into the second convolutional restricted Boltzmann machine, and combine with the normalized Laplacian matrix to perform data scheduling priority sorting, and construct a weighted directed graph according to the sorting result; S6. Initialize multiple virtual scheduling factors in the weighted directed graph, and use the ant colony optimization algorithm to iteratively search for the path sequence with the minimum cost to determine the optimal cross-platform transmission path; S7. Perform real-time distribution on the fused high-order eigenvectors according to the optimal cross-platform transmission path to achieve efficient and stable transmission of the networking charging data among multiple platforms.

2. The cross-platform fusion and transmission method of network toll data based on the intelligent scheduling algorithm according to claim 1, wherein The charging data includes vehicle information, payment records, charging timestamps, channel numbers, and corresponding image recognition results. The preprocessing includes missing value filling, field standardization, time format unification, and noise data elimination operations.

3. The cross-platform fusion and transmission method of networked toll collection data based on the intelligent scheduling algorithm according to claim 1, characterized in that, The specific content of S2 includes: S21. Map each preprocessed toll data into a feature vector x i = [x i1 , x i2 , …, x in T , where x i is the feature vector of the i-th toll data, n is the total number of fields contained in each toll data, and x ij is the normalized value of the i-th toll data on the j-th field. All data samples form a feature matrix X = [x1, x2, …, x m T , and m is the total number of toll data samples;​​ S22. Based on the eigenvectors and eigenmatrices, calculate the comprehensive feature distance between two samples based on the weighted feature kernel function: where d ij is the comprehensive feature distance between the i-th and j-th charging data, and α k is the feature weight coefficient of the k-th field, satisfying x ik , x jk represent the normalized feature values of the i-th and j-th data in the k-th field respectively, is the mean value of the k-th field in all charging data samples, that is ε is a positive real number used to avoid a zero denominator; S23. Construct a feature similarity matrix \(W = [w ij \) based on the comprehensive feature distance. Using all toll data as nodes and the non-zero elements in the feature similarity matrix as edge weights, construct an undirected graph \(G=(V, E)\), where: V = {v1, v2, …, v m}, representing the set of nodes composed of toll data samples; Represents a set of edges between nodes; For any v i , v j ∈ V, if w ij > 0, then there exists an edge e ij ∈ E, and the weight corresponding to this edge is w ij .

4. The cross-platform fusion and transmission method of network toll data based on the intelligent scheduling algorithm according to claim 3, characterized in that, The feature similarity matrix W = [w ij , and the Gaussian kernel function is used to generate the similarity weights and perform normalization processing: where exp(·) is the natural exponential function with base e, d ij is the comprehensive feature distance between the i-th toll data and the j-th toll data, σ is the hyperparameter that adjusts the smoothness of the kernel function, d ij is the comprehensive feature distance between the i-th toll data and the j-th toll data, w ij is the similarity weight between the toll data x i and x j to ensure that 5. The cross-platform fusion and transmission method of networked toll collection data based on the intelligent scheduling algorithm according to claim 1, wherein The specific content of S3 includes: S31. Based on the feature similarity matrix W = [w ij , calculate the degree matrix D = diag(d1, d2, …, d m ), where m is the total number of toll data samples, w ij is the similarity weight between toll data x i and x j , diag(·) is a diagonal matrix, the i-th diagonal element is d i , and represents the sum of the connection weights of node v i ; S32. Based on the feature similarity matrix W and the degree matrix D, calculate the normalized graph Laplacian matrix: L sym = I - D -1 / 2 WD -1 / 2 ; Among them, L sym is the normalized Laplacian matrix, I is the identity matrix, and D -1 / 2 is the element-wise negative square root diagonal matrix of the degree matrix D, which is used to balance the weight influence between different nodes; S33. Perform eigenvalue decomposition on the Laplacian matrix, extract the eigenvectors corresponding to the smallest first k eigenvalues, and construct the eigenvector matrix U = [u1, u2, …, u k , where k is the preset number of clusters and also the dimension of the low-dimensional subspace, and u l is the first eigenvector, l ∈ {1, 2, ..., k}; Normalize the eigenvector matrix U row by row to obtain the normalized eigenmatrix Z = [z1, z2, …, z m T , where z i is the representation vector of the i-th sample in the low-dimensional feature subspace;​ S34. Based on the normalized eigenmatrix Z, perform clustering on all node representation vectors, and use the k-means algorithm to minimize the following objective function: Among them, J is the clustering loss function, and c l(i) is the cluster center vector to which the i-th sample belongs, is the sample set of the p-th cluster, m is the total number of toll data samples, and |C p | is the number of samples in the cluster C p , and λ is the balance hyperparameter that adjusts the clustering compactness and separation degree; S35. Use the clustering result of each sample z i as the feature cluster label for the subsequent modality mapping process.

6. The cross-platform fusion and transmission method of networked toll collection data based on the intelligent scheduling algorithm according to claim 1, wherein The specific content of S4 includes: S41. Based on the feature cluster labels of each toll data sample, construct a modal mapping matrix M = [m ij , where m ij = 1 indicates that the i-th toll data belongs to the j-th feature cluster, otherwise m ij = 0; Through the mapping relationship, the normalized feature matrix Z = [z1, z2, …, z m T is transformed into the standard modal vector matrix:​ where m is the total number of toll data samples, is the standard modal vector matrix, R is the feature cluster weight mapping matrix, Q is the feature compression matrix, γ is the modal enhancement coefficient, and d is the standard modal dimension; S42. Reconstruct the standard modal vector matrix into a tensor representation As the input of the first convolutional restricted Boltzmann machine, define the set of learnable convolutional kernels as and the hidden layer activation output of the first convolutional restricted Boltzmann machine is: where h is the number of convolutional kernels, and r is the spatial size of each convolutional kernel, is the hidden activation unit value at the (j, t) position of the i-th sample under the action of the convolutional kernel K t , σ(·) is the Sigmoid activation function, is the double summation within the two-dimensional convolutional window, and K t (p, q) are the weight coefficients of the t-th convolutional kernel at the position (p, q), is the standard modal vector matrix is the set of sample tensors obtained by reconstruction, is the i-th sample tensor is the element value at the position (i + p - 1, j + q - 1), and b t is the convolutional kernel bias term; S43. Aggregate the activation outputs of all samples under the convolution kernel to form a three-dimensional tensor Fully connect and compress the convolution output of each sample to generate a set of high-order feature vectors V = [v1, v2, …, v m T , and the set of high-order feature vectors is used as the input of the subsequent second convolutional restricted Boltzmann machine​ 7. The method for cross-platform fusion and transmission of network toll data based on the intelligent scheduling algorithm according to claim 6, wherein The first convolutional restricted Boltzmann machine adopts a shared weight structure and includes no less than 3 groups of convolutional kernels. The size of each group of convolutional kernels is r×r, where r is the spatial size of each convolutional kernel and r≥5. The convolutional operation performs a sliding convolution with a step size of 1 on the standard modal vector tensor representation, and the activation function is After the convolutional output is subjected to max pooling processing by a pooling layer, local invariant features are generated. The local invariant features are converted into a one-dimensional vector through a flattening operation and used as part of the high-order feature vector.

8. The cross-platform fusion and transmission method of networked toll collection data based on the intelligent scheduling algorithm according to claim 1, wherein The specific content of S5 includes: S51. Use the set of high-order feature vectors \(V = [v_1, v_2, \ldots, v\) m T as the input of the second convolutional restricted Boltzmann machine. First, reconstruct each high-order feature vector \(v\) i into a tensor and combine it with the set of convolution kernels of the second convolutional restricted Boltzmann machine as Define the convolution activation output as:​ where q is the number of convolution kernels, and r′ is the size of the convolution kernel, is the activation value at the (u, v) position under the action of the j-th convolution kernel for the i-th sample, is the i-th sample tensor The element value at the position (u + a - 1, v + b - 1), G j (a, b) is the weight coefficient of the j-th convolution kernel at the position (a, b), β j is the bias term of the j-th convolution kernel, and σ(·) is the Sigmoid activation function; S52. Unfold all convolutional outputs of each sample into a one-dimensional vector r i , and for the vector r i perform priority score calculation: where s i is the scheduling priority score of the i-th sample, g is the length of the one-dimensional vector after convolution expansion, f is the dimension of the high-order feature vector, δ1, δ2, δ3 are the weighting coefficients of different scoring factors, and r it is the t-th dimensional component of the i-th sample in the convolution expansion vector r i and v ik is the k-th component value of the high-order feature vector v i , log2(·) is the logarithmic function, and max is the maximum value function; S53. Construct a weighted directed graph \(G'=(V', E', W')\) based on the priority scores of each sample and the real-time network status, where the node set \(V'=\{v'_1, v'_2, \ldots, v' m \}\) is the toll data sample, the edge set \(E'\) is the connection relationship with a transmission path between samples, and the edge weight matrix \(W' = [w' ij \), and the edge weight calculation formula is: where, w′ ij is the transmission priority weight, φ ij is the bandwidth utilization rate, λ ij is the observed real-time network delay, μ j is the current load index.

9. The cross-platform fusion and transmission method of networked toll collection data based on the intelligent scheduling algorithm according to claim 8, wherein, During the training of the second convolutional restricted Boltzmann machine, the convolutional kernel weights G j need to be updated: Among them, ΔG j is the updated value of the weight matrix of the j-th convolution kernel, η is the learning rate, is the expected value of the product of the input and output during the forward propagation process, is the reconstruction error, that is, the expected value of the product of the output and input of the reconstructed data through backpropagation.

10. The cross-platform fusion and transmission method of networked toll collection data based on the intelligent scheduling algorithm according to claim 8, characterized in that, The specific content of S6 includes: S61. In the weighted directed graph G′=(V′,E′,W′), initialize the set of ant individuals of the ant colony optimization algorithm. Each ant individual starts from the starting node in the graph to search for a path. Define the transition probability of the d-th ant from node i to node j as: where represents the probability that the d-th ant transfers from node i to node j at time t, and τ ij (t) is the current pheromone concentration on edge e′ ij , η ij (t) = 1 / w′ ij is the value of the path heuristic function, and w′ ij is the transmission priority weight, is the set of adjacent nodes that the d-th ant can reach from node i at the current moment, and α and β are respectively the pheromone intensity and the heuristic factor weight parameters; S62. After all ants complete the path traversal, define the total cost function of each path as: Among them, C d is the total path cost corresponding to the d-th ant, is the ordered edge set in the path traversed by the d-th ant, s i is the scheduling priority score, w′ ij is the transmission priority weight, ∈ ij is the instantaneous packet loss rate on edge e′ ij is ψ j is the network congestion degree of the target node j, ρ is the path cost smoothing factor, and δ is the regularization term balance factor; S63. According to the total cost function C n Sort each path, select the path with the minimum total cost as the optimal path, and update the pheromone concentration on each path edge according to the following pheromone update rule: where τ ij (t + 1) is the pheromone concentration on edge e′ at time t + 1 ij τ′ ij (t) is the pheromone concentration on edge e at time t ij ω is the pheromone evaporation coefficient, is the contribution of the ant to the pheromone on edge e′ ij Q is the pheromone release constant, and N is the number of ant individuals.

Images