Virtual point location data generation method with sparse layout of ETC door frame

Abstract

The invention belongs to the technical field of intelligent traffic, and discloses a virtual point location data generation method for sparse layout of an ETC door frame, which comprises the following steps: S1, taking irregular distribution of ETC as graph structure data, and modeling ETC door frame space distribution by using a spectral domain-based graph convolutional neural network; s2, predicting ETC data, and capturing adjacent features between a target gantry and upstream and downstream gantries through linear combination; s3, constructing a GCN-GRU neural network model; and S4, optimizing the model by adopting a genetic algorithm, and automatically searching an optimal parameter combination through population iteration, fitness evaluation and genetic manipulation. The method provided by the invention effectively overcomes the limitation that the spatial-temporal heterogeneity characteristics of the traffic flow are difficult to capture in the typical scene of the expressway in the existing data generation method, and provides higher-precision data supplement.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent transportation technology, and in particular relates to a method for generating virtual point data for sparsely arranged ETC gantries. Background Art

[0002] The multidimensional spatiotemporal data generated by highway electronic toll collection (ETC) systems serves as a critical data foundation for intelligent transportation system research. It holds significant theoretical research value and engineering practical significance in key areas such as traffic flow situational awareness, short-term traffic flow forecasting, and congestion propagation mechanism analysis. From a data science perspective, the integrity of ETC data directly impacts the reliability and effectiveness of subsequent modeling and analysis. When data contain non-random missing data or uneven spatiotemporal distribution, it can lead to systematic biases in the prediction model, affecting the accuracy of traffic control strategies and potentially even leading to secondary traffic safety risks. Therefore, multidimensional ETC data completion technology for sparse road sections has become a critical technical bottleneck that urgently needs to be overcome in the field of intelligent transportation.

[0003] Although a variety of data completion methods have been proposed, they still have many shortcomings in practical applications. Current research mainly focuses on data interpolation methods based on spatiotemporal correlation, such as the spatiotemporal matrix completion algorithm based on tensor decomposition and the deep learning framework of generative adversarial networks. However, existing methods have significant limitations when dealing with the typical scenario of highways where there is a lack of sufficient data: first, traditional interpolation methods have difficulty capturing the spatiotemporal heterogeneity of traffic flow, especially under sudden changes in road network structure (such as construction diversions) or sudden events; second, existing models have insufficient ability to integrate multi-source heterogeneous data (such as meteorological data and social public opinion data), resulting in a lack of environmental context awareness in the data completion process.

[0004] Patent document CN115081689B discloses a method for generating ETC data. This method obtains the coordinates of the starting and ending points of highways through a map API, and uses QGIS tools to perform high-precision densification of road segment points to address sparsity issues. A binary optimization algorithm (GPG algorithm) is then used to iteratively optimize the gantry coordinates, achieving automated, high-precision generation of gantry positions. However, this method is highly data-dependent and relies entirely on ETC transaction data. This can affect accuracy in sections with low vehicle traffic or sparse data, resulting in insufficient adaptability and robustness in different data scenarios. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method for generating virtual point data for sparsely arranged ETC gantries, aiming to solve the problem of insufficient adaptability and robustness of existing ETC data completion methods in scenarios with low vehicle traffic volume or sparse data.

[0006] The present invention provides a method for generating virtual point data of sparsely arranged ETC gantries, comprising the following steps:

[0007] S1. Treat the irregular distribution of ETC as a graph-structured data and use a spectral-domain-based graph convolutional neural network to model the spatial distribution of ETC gantries.

[0008] S2. Predict the ETC data and capture the proximity features between the target gantry and the upstream and downstream gantries through linear combination;

[0009] S3. Build a GCN-GRU neural network model;

[0010] S4. A genetic algorithm is used to optimize the model, automatically searching for the optimal parameter combination through population iteration, fitness evaluation, and gene manipulation.

[0011] Furthermore, step S1 includes the following sub-steps:

[0012] S1.1 treats the traffic flow parameters extracted by ETC as signals on the spectrum and captures the implicit relationship in the spatial state by processing the signals;

[0013] S1.2 converts the convolution kernel g and the input signal x in the GCN time domain into the corresponding frequency domain inner product form through convolution operation, and performs approximate calculation. The expression is as follows:

[0014]

[0015] Where g = U T g=diag(θ),θ∈R N ;U T x represents mapping x into the frequency domain space based on U; represents the frequency domain filter function constructed based on the adjacency matrix A; A represents the adjacency matrix, which is used to represent the connection between highway nodes; I n represents the self-connection matrix, which is used to ensure that the node's own information can be effectively retained and transmitted in the process of capturing the feature relationship between adjacent gantries to prevent information attenuation; D represents the degree matrix, which provides normalization for the adjacency matrix and balances the node influence;

[0016] and exists The final output at layer l is:

[0017]

[0018] Where σ represents the activation function; W represents the learnable weight matrix;

[0019] S1.3 construct a highway map model based on ETC location information;

[0020] Among them, the set of ETC gantries is V = {v1,v2,...,v N};

[0021] The construction rules of the adjacency matrix A are:

[0022]

[0023] S1.4 reconstructs data based on GCN’s virtual nodes. After inserting the virtual ETC gantry, a virtual node is added to the graph structure by modifying the direct connection relationship between the gantries.

[0024] The estimation of the virtual point is calculated as follows:

[0025] X v =[x x-T ,...,x t-1 ,x t ]=f(A v ;(X i ,X j ))

[0026] Where, X v is the estimated historical traffic value of the virtual ETC point; T is the required historical time series length; f(A v ;(X i ,X j )) is the mapping function; A v is the adjacency matrix after embedding virtual nodes; X i is the historical traffic data of the upstream ETC gantry closest to the virtual point; X j It is the historical traffic data of the downstream ETC gantry closest to the virtual point.

[0027] Furthermore, step S2 includes the following sub-steps:

[0028] S2.1 processes temporal dependencies through gated recurrent units;

[0029] S2.2 introduces graph attention mechanism to enhance model performance;

[0030] The adjacency matrix A between the actual road directly connected node pairs is used as a mask to control the attention calculation range. The attention calculation formula is:

[0031]

[0032] Where, represents the learnable attention weight vector; W represents the learnable weight matrix; is the new feature of node i after integrating domain information; is the new feature of node j after integrating domain information;

[0033] S2.3 introduces a distance attenuation mechanism to allow distant and nearby nodes to receive differentiated treatment in attention allocation;

[0034] Assume that the physical distance between door frame i and door frame j is δ(i,j), then the attention calculation formula after distance adjustment is:

[0035]

[0036] Where β is the hyperparameter obtained through training; f(δ(i,j)) is a nonlinear function;

[0037] S2.4 normalizes the attention weights using the softmax function;

[0038]

[0039] Based on the normalized attention weights, update the feature representation of each node:

[0040]

[0041] Where, is the set of neighbor nodes of node i.

[0042] Furthermore, step S3 includes the following sub-steps:

[0043] S3.1 The traffic flow time-varying feature matrix X in the ETC gantry data node feature matrix t And the adjacency matrix A is input into the GCN layer to extract spatial features at each time node t. The extraction results are as follows:

[0044]

[0045] Where, is the normalized adjacency matrix; W (l) is the GCN weight matrix of the lth layer;

[0046] S3.2 The spatial feature H obtained in step S3.2 (t) As the input of the GRU model, it captures the dynamic changes of traffic flow over time.

[0047] Furthermore, step S4 includes the following sub-steps:

[0048] S4.1 Population initialization;

[0049] During the initialization process, it is necessary to preset the range of hyperparameter values ​​in the population, population size, chromosome length, chromosome encoding method, maximum number of iterations, and crossover probability;

[0050] S4.2 Fitness calculation;

[0051] The prediction accuracy in the GRU neural network model is used as the value of the individual's fitness, and the calculation formula is as follows:

[0052]

[0053] Where, f(x i ) is the fitness function; n is the number of corresponding data; is the prediction result under the jth set of hyperparameters; j is the corresponding actual value;

[0054] S4.3 Select operation;

[0055] The calculated fitness values ​​are used to select individuals for the next generation of crossover and mutation, and the optimal hyperparameters are achieved through multiple iterations and evolutionary processes;

[0056] S4.4 Crossover operation;

[0057] Combine and cross the hyperparameters of two individuals to generate new individuals;

[0058] S4.5 mutation operation;

[0059] Randomly change the hyperparameters in individuals;

[0060] S4.6 terminates the judgment and outputs the optimal solution;

[0061] After meeting the maximum number of iterations or the model judgment value, the hyperparameter combination represented by the highest fitness is output.

[0062] Beneficial effects:

[0063] The present invention proposes a method for generating virtual point data for sparsely distributed ETC gantries. In view of the situation where the large spacing between ETC gantries on highways leads to limited monitoring range and sparse data, which in turn affects the model effect, a GCN model and a GRU model are combined to construct a GCN-GRU neural network. The model captures spatial topological associations through graph convolution operations, processes temporal dependencies through gated recurrent units, and fuses distance-weighted heterogeneous influences through a graph attention mechanism, forming a model that can simultaneously model spatial dependencies, temporal evolution, and distance attenuation characteristics. This multi-dimensional fusion network structure is particularly suitable for processing the spatiotemporal heterogeneity of highway ETC gantry data, achieving high-precision data supplementation for missing detection points, and improving adaptability and accuracy in complex traffic environments. The above method effectively overcomes the limitation of existing data generation methods in capturing the spatiotemporal heterogeneity of traffic flows in typical highway scenarios, and provides higher-precision data supplementation.

[0064] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a flow chart of a method for generating virtual point data for sparse ETC gantry layout according to the present invention;

[0066] Figure 2 A framework diagram for generating virtual ETC point data on data-sparse road sections. DETAILED DESCRIPTION

[0067] To make the technical solutions, advantages, and purposes of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0068] like Figure 1 As shown, the present invention provides a method for generating virtual point data of sparsely arranged ETC gantry, comprising the following steps:

[0069] S1. Treat the irregular distribution of ETC as a graph-structured data and use a spectral-domain graph convolutional neural network to model the spatial distribution of ETC gantries.

[0070] S1.1 treats the traffic flow parameters extracted by ETC as signals on the spectrum and captures the implicit relationship in the spatial state by processing the signals. The Laplacian matrix of the graph is defined as:

[0071] L=DA

[0072] After regularization:

[0073]

[0074] Where, l n ∈R n×n is the identity matrix; the degree matrix D ii =∑ i A ii .

[0075] After performing eigenvalue decomposition on L, we can get L=U∧U T , where ∧=diag([λ1,...,λ n ]) is a diagonal matrix composed of the eigenvalues ​​of L, and U={u1,u2,....,u n} is an orthogonal matrix consisting of the standard orthonormal eigenvectors of L.

[0076] S1.2 According to the input signal X∈R n , then the Fourier transform in the figure is The corresponding inverse Fourier transform is The convolution kernel g and the input signal x in the GCN time domain can be converted into the corresponding frequency domain inner product form through the convolution operation. The calculation formula is as follows:

[0077]

[0078] Where g = U T g=diag(θ),θ∈R N ⊙ is the Hadamard product; U T g means mapping g into the frequency domain space based on U.

[0079] Due to g θ The calculation process of is quite complex. In conventional applications, it is approximated by using hierarchical linear model constraints and Chebyshev polynomials. After the approximate calculation, the calculation process is as follows:

[0080]

[0081] and exists The final output at layer l is:

[0082]

[0083] S1.3 constructs a highway graph model based on the ETC location information, where A represents the adjacency matrix, which is used to represent the connection between highway nodes.

[0084] Each row of A represents the corresponding situation of an ETC gantry, and each value in A represents the connection between ETC gantries. This is to prevent the ETC gantry on the highway from losing or being unable to transmit its own information during the process of capturing the feature relationship between adjacent gantries.

[0085] The set of ETC gantries is V = {v1,v2,...,v N}Yes, the construction rule of the corresponding adjacency matrix A is:

[0086]

[0087] S1.4 reconstructs data based on virtual nodes of GCN. After inserting virtual ETC gantry, the direct connection relationship between the gantry is modified to add a virtual node to the graph structure. The flow data corresponding to each gantry is used as the node attribute feature, which is represented as X n*p , where n is the number of nodes and p is the number of features of node attributes.

[0088] Correspondingly, the estimation of the embedded virtual point can be considered as the mapping function f for the road network topology A and the flow characteristic matrix X n*p The corresponding expression for the calculation after learning is:

[0089] X v =[x x-T ,...,x t-1 ,x t ]=f(A v ;(X i ,X j ))

[0090] Where, X v is the estimated historical traffic value of the virtual ETC point; T is the required historical time series length; f(A v ;(X i ,X j )) is the mapping function; A v is the adjacency matrix after embedding virtual nodes; X i is the historical traffic data of the upstream ETC gantry closest to the virtual point; X j It is the historical traffic data of the downstream ETC gantry closest to the virtual point.

[0091] After constructing the corresponding updated adjacency matrix and its distance feature matrix after inserting the virtual gantry, the data of the virtual ETC point is reconstructed through the GCN model.

[0092] S2. Predict the ETC data by capturing the proximity features between the target gantry and its upstream and downstream gantries through linear combination, while also considering the evolutionary characteristics of the node itself. This includes the following sub-steps:

[0093] S2.1 In the GRU model, there is a reset gate (r t ) and update gate (Z t ) Two gate control units, where Z t Indicates how much the state at the previous moment is transferred to the current state, and r t It indicates how much of the state information of the previous moment has been forgotten. The calculation process is as follows:

[0094] Z t =σ(W z ·[x t ,H t-1 ])

[0095] r t =σ(W z ·[x t ,H t-1 ])

[0096] H t =σ(W h ·[r t ×H t-1 ,x t ])

[0097] H t =(1-z t )×h t-1 +z t ×H t

[0098] Where W z Represents the reset gate r t The weight of ; σ is the activation function; x t Represents the ETC traffic and speed matrix at the current moment; H t-1 is the hidden state at time t-1.

[0099] S2.2 introduces a graph attention mechanism with spatial perception to enhance model performance. The attention function a is defined as ij It indicates the degree of attention of gantry i to gantry j. The calculation process is as follows:

[0100] First, the adjacency matrix A between the actual road directly connected node pairs is used as a mask to control the attention calculation range.

[0101]

[0102] The attention weight is controlled by the adjacency matrix value to achieve effective spatial information screening.

[0103] S2.3 takes into account the impact of the actual physical distance between nodes on information transmission, and introduces a distance attenuation mechanism so that distant and near nodes receive differentiated treatment in attention allocation.

[0104] Assume that the physical distance between door frame i and door frame j is δ(i,j), then the attention calculation formula after distance adjustment is:

[0105]

[0106] Where β is the hyperparameter obtained through training; f(δ(i,j)) is a nonlinear function;

[0107] S2.4 Taking all the above factors into consideration, the final attention score function expression is:

[0108]

[0109] The attention weights are then normalized using the softmax function:

[0110]

[0111] Based on the normalized attention weights, update the feature representation of each node:

[0112]

[0113] Where, is the new feature of node i after integrating domain information; is the new feature of node j after integrating domain information.

[0114] S3. Build a GCN-GRU neural network model to generate virtual ETC point data;

[0115] S3.1 data features are connected to the GCN layer, and the traffic flow time-varying feature matrix X in the input ETC gantry data node feature matrix is ​​converted into t And the adjacency matrix A is input into the GCN layer to realize the extraction of spatial features at each time node t. The extraction results are as follows:

[0116]

[0117] Where, σ is the nonlinear activation function ReLU; is the normalized adjacency matrix; W (l) is the GCN weight matrix of the lth layer;

[0118] S3..2 will get the spatial feature H output by the GCN layer (t) As the input of the GRU model, it captures the dynamic changes in time. The calculation process of its update gate when changing the input is as follows:

[0119] z (t) =σ(W z [h (t-1) ,H (t) ]+U z h (t-1) )

[0120] r (t) =σ(W r [h (t-1) ,H (t) ]+U r h (t-1) )

[0121]

[0122] After the above process analysis, GCN fully processes the spatial dependencies between road traffic topological structures and captures the spatial propagation characteristics. Then, the GRU model is used to process the spatiotemporal continuity characteristics of traffic flow and finally achieve the purpose of generating data for virtual ETC points.

[0123] S4. Use a genetic algorithm to optimize the model. This algorithm automatically searches for the optimal parameter combination through population iteration, fitness evaluation, and genetic manipulation, taking into account both globality and randomness. This includes the following sub-steps:

[0124] S4.1 Population Initialization. In a genetic algorithm, each individual represents a candidate solution for a hyperparameter, and a population is a collection of individuals. During the genetic algorithm optimization process, the hyperparameter value range and corresponding encoding method are converted into the corresponding chromosome encoding method. During the initialization process, it is necessary to preset the hyperparameter value range, population size, chromosome length, chromosome encoding method, maximum number of iterations, crossover probability, and other parameters in the population.

[0125] S4.2 Fitness calculation. The prediction accuracy in the GRU neural network model is used as the fitness value of the individual. The calculation formula is as follows:

[0126]

[0127] Where, f(xi ) is the fitness function; n is the number of corresponding data; is the prediction result under the jth set of hyperparameters; j is the corresponding actual value.

[0128] S4.3 Selection Operation. The calculated fitness values ​​are used to select which individuals will be used for crossover and mutation information in the next generation. Through this iterative and evolutionary process, optimal hyperparameter selection is achieved.

[0129] S4.4 Crossover operation. The crossover operation generates new individuals by combining the hyperparameters of two individuals. Through the crossover operation, the individuals in the evolving crowd gradually tend to a better solution through multiple iterations.

[0130] S4.5 Mutation operation. The mutation operation introduces randomness and diversity by randomly changing the hyperparameters in individuals, preventing them from falling into local optimal solutions.

[0131] S4.6 terminates the judgment and outputs the optimal solution, and outputs the hyperparameter combination represented by the highest fitness after meeting the maximum number of iterations or the model judgment value.

[0132] In this embodiment, the overall framework diagram of virtual ETC point data generation can be found in Figure 2 .

[0133] It is hereby stated that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for generating virtual point data for sparsely distributed ETC gantry, characterized in that: The following steps are involved: S1. Treat the irregular distribution of ETC as a graph-structured data and use a spectral-domain-based graph convolutional neural network to model the spatial distribution of ETC gantries. S2. Predict the ETC data and capture the proximity features between the target gantry and the upstream and downstream gantries through linear combination; S3. Build a GCN-GRU neural network model; S4. A genetic algorithm is used to optimize the model, automatically searching for the optimal parameter combination through population iteration, fitness evaluation, and gene manipulation.

2. The method for generating virtual point data for sparsely arranged ETC gantry according to claim 1, characterized in that: The step S1 includes the following sub-steps: S1.1 treats the traffic flow parameters extracted by ETC as signals on the spectrum and captures the implicit relationship in the spatial state by processing the signals; S1.2 converts the convolution kernel g and the input signal x in the GCN time domain into the corresponding frequency domain inner product form through convolution operation, and performs approximate calculation. The expression is as follows: Where g = U T g=diag(θ),θ∈R N ;U T x represents mapping x into the frequency domain space based on U; represents the frequency domain filter function constructed based on the adjacency matrix A; A represents the adjacency matrix, which is used to represent the connection between highway nodes; I n represents the self-connection matrix, which is used to ensure that the node's own information can be effectively retained and transmitted in the process of capturing the feature relationship between adjacent gantries to prevent information attenuation; D represents the degree matrix, which provides normalization for the adjacency matrix and balances the node influence; and exists The final output at layer l is: Where σ represents the activation function; W represents the learnable weight matrix; S1.3 construct a highway map model based on ETC location information; Among them, the set of ETC gantries is V = {v1,v2,...,v N }; The construction rules of the adjacency matrix A are: S1.4 reconstructs data based on GCN’s virtual nodes. After inserting the virtual ETC gantry, a virtual node is added to the graph structure by modifying the direct connection relationship between the gantries. The estimation of virtual nodes is calculated as follows: X v =[x x-T ,...,x t-1 ,x t ]=f(A v ;(X i ,X j )) Where, X v is the estimated historical traffic value of the virtual ETC point; T is the required historical time series length; f(A v ;(X i ,X j )) is the mapping function; A v is the adjacency matrix after embedding virtual nodes; X i is the historical traffic data of the upstream ETC gantry closest to the virtual point; X j It is the historical traffic data of the downstream ETC gantry closest to the virtual point.

3. The method for generating virtual point data for sparsely arranged ETC gantry according to claim 2, characterized in that: The step S2 includes the following sub-steps: S2.1 processes temporal dependencies through gated recurrent units; S2.2 introduces graph attention mechanism to enhance model performance; The adjacency matrix A between the actual road directly connected node pairs is used as a mask to control the attention calculation range. The attention calculation formula is: Where, represents the learnable attention weight vector; W represents the learnable weight matrix; is the new feature of node i after integrating domain information; is the new feature of node j after integrating domain information; S2.3 introduces a distance attenuation mechanism to allow distant and nearby nodes to receive differentiated treatment in attention allocation; Assume that the physical distance between door frame i and door frame j is δ(i,j), then the attention calculation formula after distance adjustment is: Where β is the hyperparameter obtained through training; f(δ(i,j)) is a nonlinear function; S2.4 normalizes the attention weights using the softmax function; Based on the normalized attention weights, update the feature representation of each node: Where, is the set of neighbor nodes of node i.

4. The method for generating virtual point data for sparsely arranged ETC gantry according to claim 3 is characterized in that: The step S3 includes the following sub-steps: S3.1 The traffic flow time-varying feature matrix X in the ETC gantry data node feature matrix t And the adjacency matrix A is input into the GCN layer to extract spatial features at each time node t. The extraction results are as follows: Where, is the normalized adjacency matrix; W (l) is the GCN weight matrix of the lth layer; S3.2 The spatial feature H obtained in step S3.2 (t) As the input of the GRU model, it captures the dynamic changes of traffic flow over time.

5. The method for generating virtual point data for sparsely arranged ETC gantry according to claim 4, characterized in that: The step S4 includes the following sub-steps: S4.1 Population initialization; During the initialization process, it is necessary to preset the range of hyperparameter values ​​in the population, population size, chromosome length, chromosome encoding method, maximum number of iterations, and crossover probability; S4.2 Fitness calculation; The prediction accuracy in the GRU neural network model is used as the value of the individual's fitness, and the calculation formula is as follows: Where, f(x i ) is the fitness function; n is the number of corresponding data; is the prediction result under the jth set of hyperparameters; j is the corresponding actual value; S4.3 Select operation; The calculated fitness values ​​are used to select individuals for the next generation of crossover and mutation, and the optimal hyperparameters are achieved through multiple iterations and evolutionary processes; S4.4 Crossover operation; Combine and cross the hyperparameters of two individuals to generate new individuals; S4.5 mutation operation; Randomly change the hyperparameters in individuals; S4.6 terminates the judgment and outputs the optimal solution; After meeting the maximum number of iterations or the model judgment value, the hyperparameter combination represented by the highest fitness is output.

Images