Sparse data automatic labeling method and system based on graph neural network

Abstract

The invention discloses an automatic sparse data labeling method and system based on a graph neural network, and the method comprises the steps: constructing a space-time diagram structure, carrying out the embedding and mapping of traffic flow and speed features through a Grassmann manifold, and generating node features and edge weights; constructing a traffic flow field based on the node features and the edge weights, decomposing the traffic flow field into diffused gradient features and direction-changing curl features, and updating the edge weights; designing a drawing convolution mechanism, and fusing scalar potential and vector potential convolution kernels to generate a propagation kernel; constructing a space-time propagation mechanism through a long and short-term memory network to update node features; a dynamic action range is introduced, and a propagation range is adjusted by combining node density and a traffic flow field; the traffic flow field is used as a confidence index, traffic flow information of marked nodes is spread to sparse nodes through a Poisson equation, and high-precision automatic marking is achieved; the method overcomes the limitation that a traditional method neglects nonlinear dynamic and static propagation, can improve the precision and robustness of sparse node labeling, and is suitable for peak and emergency scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of data annotation processing under deep learning, and in particular to a method and system for automatic annotation of sparse data based on graph neural networks. Background Art

[0002] In recent years, with the rapid development of intelligent transportation systems, traffic flow prediction technology based on graph neural networks (GNNs) has made significant progress in processing complex spatiotemporal data. By modeling traffic networks as graph structures and leveraging the topological relationships between nodes and edges to capture the spatiotemporal correlations of traffic flow characteristics such as volume and speed, GNNs have been widely used in urban traffic management, route planning, and congestion prediction. Existing technologies typically optimize data propagation within networks by enhancing the expressive power of traffic flow characteristics. Currently, for sparse data scenarios, some studies have attempted to achieve predictions for unlabeled nodes through interpolation methods or semi-supervised learning. These technological advances have provided powerful tools for traffic flow prediction and are driving the development of intelligent transportation systems towards higher accuracy and real-time performance.

[0003] However, existing technologies still have significant deficiencies in the automatic labeling of sparse data. First, when dealing with sparse nodes, traditional methods often rely on simple interpolation or neighborhood averaging strategies, ignoring the nonlinear dynamic characteristics of traffic flow in the time and space dimensions, resulting in high labeling errors, especially poor performance during peak hours or in emergency scenarios (such as traffic accidents). Secondly, existing technologies usually adopt a static propagation mechanism, which makes it difficult to adaptively adjust the propagation range to cope with topological changes in sparse areas and sudden changes in traffic flow patterns, thereby limiting the ability to capture long-distance dependencies. In addition, existing technologies lack accurate modeling of local changes in traffic flow during confidence assessment and label propagation, and are prone to reducing the labeling accuracy of sparse nodes due to information loss or error accumulation. Summary of the Invention

[0004] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of this application to avoid obscuring the purpose of this section, the abstract and the title of the invention, and such simplifications or omissions should not be used to limit the scope of the present invention.

[0005] In view of the above existing problems, the present invention is proposed. Therefore, the present invention provides a sparse data automatic annotation method based on graph neural network to solve the problems mentioned in the background technology.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for automatic labeling of sparse data based on a graph neural network, comprising:

[0008] The traffic network is modeled as a spatiotemporal graph structure. The traffic flow and speed features are embedded into the graph structure through feature mapping method to generate node features and edge weights.

[0009] Constructing a traffic flow field using the node features and edge weights, generating diffusion features and directional features based on local changes in vehicle volume and speed, and updating edge weights;

[0010] A graph convolution mechanism is designed based on the diffusion and direction features to handle the diffusion and direction interactions of traffic flow, and a propagation kernel is generated through feature fusion.

[0011] Based on the edge weights and propagation kernels, a spatiotemporal propagation mechanism is constructed, and node features are updated by combining a coupled propagation method with a time series model;

[0012] By utilizing the density of the traffic flow field and node features, a dynamic range of action is introduced to adjust the propagation range to adapt to traffic flow changes in sparse areas;

[0013] The traffic flow field is used as a confidence indicator and combined with the node features to propagate the traffic flow information of the labeled nodes to sparse nodes to achieve automatic labeling.

[0014] As a preferred solution of the sparse data automatic annotation method based on graph neural network described in the present invention, the traffic network is modeled as a spatiotemporal graph structure, including:

[0015] The nodes in the spatiotemporal graph structure represent location points in the transportation network, and the edges in the spatiotemporal graph structure represent connections between the location points;

[0016] Collect real-time traffic flow information from sensors, including vehicle volume and speed, and combine it with historical traffic patterns to generate node features;

[0017] By embedding the traffic flow information into a low-dimensional space, the geometric characteristics of the traffic flow are preserved.

[0018] The edge weights are calculated based on the connection distance between location points and the similarity of traffic flows.

[0019] As a preferred solution of the sparse data automatic labeling method based on graph neural network of the present invention, the traffic flow field is constructed by using the node features and edge weights, including:

[0020] According to the node characteristics, the difference of traffic flow information between nodes is calculated to determine the traffic flow field;

[0021] The traffic flow field is decomposed into a diffusion gradient feature and a direction change curl feature, and the edge weights are updated simultaneously.

[0022] As a preferred solution of the sparse data automatic labeling method based on graph neural network described in the present invention, a graph convolution mechanism is designed according to the diffusion characteristics and directional characteristics, including:

[0023] Based on the gradient characteristics of diffusion, a scalar potential convolution kernel is constructed to handle the diffusion of traffic flow from high-flow areas to low-flow areas;

[0024] Based on the curl characteristics of direction changes, a vector potential convolution kernel is constructed to process the interaction of traffic flow directions at the location point;

[0025] Through feature fusion, the convolution results of the scalar potential convolution kernel and the vector potential convolution kernel are combined to generate the propagation kernel.

[0026] As a preferred solution of the sparse data automatic annotation method based on graph neural network described in the present invention, wherein: constructing a spatiotemporal propagation mechanism based on the edge weight and propagation kernel includes:

[0027] Based on the propagation kernel, the node features are updated using a long short-term memory network.

[0028] As a preferred solution of the sparse data automatic annotation method based on graph neural network described in the present invention, the introduction of dynamic scope includes:

[0029] Calculating node density based on the traffic flow information distribution of neighboring nodes according to the node characteristics;

[0030] According to the traffic flow field, the dynamic action range is set to change with the node density.

[0031] As a preferred solution of the sparse data automatic labeling method based on graph neural network described in the present invention, wherein: the traffic flow field is used as a confidence indicator, combined with the node features, the traffic flow information of the labeled nodes is propagated to the sparse nodes, including:

[0032] Based on the traffic flow field, calculate the confidence of each node;

[0033] Using the confidence level, a confidence threshold is set, and when the confidence threshold is met, traffic flow information of the node with the highest confidence level is preferentially transmitted;

[0034] According to the node characteristics, the Poisson equation is constructed, and the traffic flow information of the labeled nodes is used as the boundary condition. The multi-grid method is used to iteratively solve the problem and propagate the labels to the sparse nodes.

[0035] In a second aspect, the present invention provides a sparse data automatic annotation system based on a graph neural network, which includes:

[0036] The spatiotemporal graph modeling and multi-source feature embedding module is configured to model the traffic network as a spatiotemporal graph structure, embed and map traffic flow and speed features into the graph structure through feature mapping methods, and generate node features and edge weights;

[0037] a traffic flow field decomposition and dynamic edge weight reconstruction module, configured to construct a traffic flow field using the node features and edge weights, generate diffusion features and directional features based on local changes in vehicle volume and speed, and update edge weights;

[0038] a dual-channel graph convolution kernel generation module, configured to design a graph convolution mechanism based on the diffusion feature and the directional feature, process the diffusion and directional interaction of the traffic flow, and generate a propagation kernel through feature fusion;

[0039] a spatiotemporal coupled propagation and time series integration module, configured to construct a spatiotemporal propagation mechanism based on the edge weights and propagation kernels, and update node features by combining a time series model with a coupled propagation method;

[0040] a dynamic field adaptive adjustment module configured to utilize the density of the traffic flow field and node features, introduce a dynamic range of action, and adjust the propagation range to adapt to traffic flow changes in sparse areas;

[0041] The confidence-guided annotation propagation module is configured to use the traffic flow field as a confidence indicator and, in combination with the node features, propagate the traffic flow information of the annotated nodes to sparse nodes to achieve automatic annotation.

[0042] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: the processor implements any step of the above method when executing the computer program.

[0043] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, any step of the above method is implemented.

[0044] Compared with the prior art, the invention has the following beneficial effects:

[0045] 1. By constructing a spatiotemporal graph structure and traffic flow field, and utilizing Grassmann manifold embedding, a dual-channel graph convolution mechanism, and a Poisson equation propagation method, this method not only achieves automatic annotation of sparse nodes but also reduces annotation errors caused by the nonlinear dynamic characteristics of traffic flow. This method achieves higher prediction accuracy than traditional interpolation or neighborhood averaging strategies, especially during peak hours or in emergency scenarios.

[0046] 2. In addition, by introducing a dynamic range and confidence-guided annotation propagation mechanism, the propagation range can be adaptively adjusted to adapt to the topological changes of sparse areas, effectively capturing the long-distance dependencies of traffic flow, thereby improving the robustness and adaptability of sparse node annotation. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort. Among them:

[0048] Figure 1 This is an overall flow chart of a method for automatic labeling of sparse data based on graph neural networks according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0050] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0051] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0052] The present invention is described in detail with reference to schematic diagrams. For ease of illustration, cross-sectional views of device structures may be partially enlarged and not to scale when describing embodiments of the present invention. Furthermore, the schematic diagrams are merely illustrative and should not limit the scope of the present invention. Furthermore, in actual production, the three-dimensional dimensions of length, width, and depth should be included.

[0053] In the description of the present invention, it should be noted that the terms "upper, lower, inner, and outer" and other references to orientations or positional relationships are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first, second, or third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0054] In this disclosure, unless otherwise specified or limited, the terms "mounted," "connected," and "connected" should be interpreted broadly. For example, they may refer to fixed, removable, or integral connections. They may also refer to mechanical, electrical, or direct connections, indirect connections through an intermediary, or internal communication between two components. Those skilled in the art will understand the specific meanings of these terms in this disclosure.

[0055] Example 1

[0056] Reference Figure 1 , which is the first embodiment of the present invention, provides a sparse data automatic labeling method based on a graph neural network, comprising:

[0057] S1. Model the traffic network as a spatiotemporal graph structure, embed and map traffic flow and speed features into the graph structure through feature mapping method, and generate node features and edge weights;

[0058] Furthermore, real-time traffic data, including traffic volume (number of vehicles passing per hour) and average speed (km / h), is obtained from sensors in the traffic network (e.g., cameras, radars, and vehicle-mounted GPS);

[0059] Specifically, historical traffic data (e.g., data from the past week or month) is collected to capture periodic trends (e.g., peak hours on weekdays and low traffic periods on weekends);

[0060] In addition, if you need to call historical traffic data, you can download the traffic dataset from the GRIP Global Roads website. The traffic dataset contains sensor nodes (coverage rate is about 30-40%) or obtain it by calling the sensor traffic flow interface.

[0061] Specifically, the acquired real-time / historical traffic data is cleaned, and statistical methods (such as Z-score) are used to identify and remove abnormal traffic data. Abnormal traffic data refers to situations where the traffic volume suddenly changes to 0 or the speed is abnormally high due to sensor failure. For example, if the traffic volume of a node exceeds 3 times the standard deviation of the historical mean, it is marked as an anomaly and replaced with the average value of the neighboring nodes; for missing real-time traffic data, interpolation is performed based on the periodic trend of historical traffic data or the weighted average of neighboring nodes; for example, if the traffic volume of a node is missing at time t, the traffic volume of the node's neighboring nodes is calculated, where the weighted weight is the inverse of the road distance; after processing the abnormal traffic data and missing traffic data, the traffic volume and speed in the traffic data are normalized to the interval [0,1] to avoid the impact of dimensional differences on subsequent processing, where the normalization uniformly adopts Min-Max;

[0062] Specifically, a spatiotemporal graph structure is modeled based on the normalized data and a graph neural network architecture, wherein nodes in the spatiotemporal graph structure represent locations in the traffic network, and edges in the spatiotemporal graph structure represent connections between the locations;

[0063] It should be noted that the solution of the present invention is mainly designed around node embedding and graph convolution in the graph neural network architecture;

[0064] Specifically, a feature vector is generated for each node (representing a road intersection or sensor location) as a node feature. The feature vector includes real-time features: the current traffic volume and speed; historical features: the traffic volume and speed sequence over the past T time steps (e.g., T = 12, with one time step every 5 minutes); and contextual features: the node's geographic coordinates and road type (e.g., highway, branch road).

[0065] Specifically, the dimension of the node feature vector is based on the traffic flow, speed, and historical window. For example, if each time step contains traffic flow and speed, and the historical window is 12, then the feature vector is 2×12+2=26 dimensions (1 dimension each for real-time traffic flow and speed, and 24 dimensions for historical data);

[0066] It should be noted that for sensorless nodes (i.e., unlabeled nodes), also known as sparse nodes, their node features are estimated by weighted average of neighboring nodes, and the weighted weights are obtained based on road distance or historical traffic similarity;

[0067] Furthermore, through Grassmann manifold embedding, we aim to map high-dimensional node features into low-dimensional space, preserve the geometric characteristics of traffic flow information, optimize the expressiveness of spatiotemporal graph structures, and provide efficient feature representation for automatic labeling.

[0068] Specifically, the generated node feature vectors are combined into a feature matrix Where N is the number of nodes and d is the feature dimension. For example, for 1100 nodes, the feature dimension d = 26, and X is a 1100 × 26 feature matrix.

[0069] Specifically, singular value decomposition is performed on X, and the first k singular vectors are taken as an orthogonal basis to generate low-dimensional features. On the Grassmann manifold Gr(k, d), the similarity between nodes is calculated by projecting the F-norm, retaining the geometric characteristics of the traffic flow (such as the local smoothness of the traffic flow distribution), forming an embedding operation for the low-dimensional features;

[0070] Specifically, the embedded features are normalized to ensure numerical stability, and regularization terms (such as L2 regularization) are added to prevent overfitting, so that each node can obtain reduced dimensions while retaining spatiotemporal correlation. The optimization formula is:

[0071]

[0072] Among them, X embed Represented as low-dimensional features, V is represented as the projection matrix, and the dimension is Used to transform low-dimensional features X embed Map back to the high-dimensional space to reconstruct the original feature X; is the square of the Frobenius norm (F-norm), which represents the difference between the original feature matrix X and the reconstructed matrix X embed V T The reconstruction error between is used to measure the degree of information loss in the dimensionality reduction process. The goal is to minimize this error to ensure that X embed Keep the traffic flow information of X as much as possible (such as the spatiotemporal correlation of traffic volume and speed) to reduce the impact of traffic flow information loss on annotation accuracy; is the square of the L2 norm of the projection matrix V, which prevents overfitting. In traffic flow prediction, overfitting may make the model sensitive to noise (such as abnormal sensor data) and reduce the robustness of sparse node automatic labeling. The L2 norm can maintain the model's generalization ability by constraining the element size of V. λ is the regularization coefficient, a non-negative scalar with an initial value of 0.01 (which can be optimized in the range of [0.001, 0.1] through cross-validation or grid search). It is used to balance the weights of the reconstruction error term and the regularization term. A smaller λ emphasizes minimizing the reconstruction error and retaining more original information. A larger λ can enhance regularization, simplify the model, and improve generalization ability.

[0073] It should be noted that in the Grassmann manifold embedding, the row vectors of V are orthogonal bases (generated by the right singular vectors of SVD), which represent the principal component directions of the high-dimensional feature space. The optimization process minimizes the reconstruction error by adjusting V;

[0074] Furthermore, based on the connection distance between location points and the similarity of traffic flows, edge weights are calculated to reflect the correlation of traffic flows between nodes;

[0075] Specifically, for two connected nodes, calculate the geographical distance of the road:

[0076]

[0077] Among them, d ij represents the distance between node i and node j, x i Represented as the horizontal coordinate of node i, y i Expressed as the vertical coordinate of node i, x j and y j Similarly;

[0078] Specifically, the normalized distance is the edge weight

[0079]

[0080] Where θ is the distance scale parameter (the initial value is 1 / 2 of the average road length);

[0081] In addition, you can also consider the road type (for example, the weight of the highway is higher than the branch road) and adjust the edge weight:

[0082]

[0083] Among them, α type is the road type coefficient (e.g. 1.2 for expressway, 0.8 for branch road);

[0084] Specifically, for two connected nodes, calculate the traffic flow similarity sim ij , using cosine similarity:

[0085]

[0086] Among them, f i 、f j are the traffic flow sequences of node i and node j respectively, and sim ij Convert to weight

[0087]

[0088] It should be noted that if the node is an unlabeled node, the traffic flow data is missing, and the node features obtained by the aforementioned sparse node processing are used;

[0089] Furthermore, the final edge weight is calculated by combining the road length and traffic flow similarity:

[0090]

[0091] Among them, β∈[0,1] is the fusion coefficient (the initial value is 0.5);

[0092] Specifically, by normalizing the edge weights, we ensure that the sum of the node neighborhood weights is 1;

[0093] S2. Constructing a traffic flow field using the node features and edge weights, generating diffusion features and directional features based on local changes in traffic volume and speed, and updating edge weights;

[0094] Specifically, according to the node characteristics, the difference of traffic flow information between nodes is calculated to determine the traffic flow field;

[0095] Specifically, the difference in traffic flow information includes traffic volume difference and speed difference:

[0096] Δf ij =f i -f j

[0097] Δv ij =v i -v j

[0098] Where Δf ij is the traffic volume difference, Δv ij is the velocity difference, v i and v j The velocity vectors of nodes i and j respectively (two-dimensional, including direction and magnitude);

[0099] Specifically, combined with the edge weights, the weighted traffic flow difference and speed difference are calculated:

[0100]

[0101] in, The weighted traffic flow difference, Weighted speed difference;

[0102] Furthermore, the traffic flow field is defined as the weighted difference field between nodes, and we get:

[0103] Scalar field (traffic flow):

[0104] Vector field (velocity):

[0105] Among them, F f (f i ) represents the local change intensity of the traffic flow at node i, F v (v i) represents the change in the velocity direction of node i, and N(i) represents the set of neighboring nodes of node i;

[0106] Further, the traffic flow field is decomposed into a diffusion gradient feature and a direction change curl feature, and the edge weights are updated at the same time;

[0107] Specifically, the traffic flow field is decomposed into gradient features (scalar potential, reflecting the diffusion of traffic flow) and curl features (vector potential, reflecting the change in speed direction);

[0108] Specifically, we define the gradient feature to represent the spatial rate of change of traffic flow, analogous to the scalar potential:

[0109]

[0110] Among them, n ij Represented as an edge direction unit vector;

[0111] Specifically, by calculating the weighted traffic flow gradient of each node, it is stored as the gradient feature g i ;

[0112] Specifically, the curl characteristic is defined to represent the rotational component of the velocity, analogous to the vector potential:

[0113]

[0114] Where × represents the cross product, which is used to capture the rotation of the velocity direction (such as turning at an intersection);

[0115] Specifically, the curl feature c is generated by calculating the cross product of the velocity vector i ;

[0116] Furthermore, the edge weights are dynamically adjusted based on the gradient and curl characteristics of the diffusion;

[0117] Specifically, calculate the traffic flow intensity s ij :

[0118] s ij =||g i -g j ||2+||c i -c j ||2

[0119] It should be noted that adjusting the edge weights through the intensity of the traffic flow field can reflect the latest dynamics of the traffic flow field;

[0120] S3. Designing a graph convolution mechanism based on the diffusion and directional features to process the diffusion and directional interactions of traffic flow, and generating a propagation kernel through feature fusion;

[0121] It should be explained that before designing the graph convolution mechanism, it is necessary to model the dynamic changes of traffic flow and construct a non-Euclidean space-time vortex propagation equation to simulate the diffusion, convection and external constraints of traffic flow, providing a dynamic propagation mechanism for node feature updates;

[0122] Specifically, the dynamic changes of traffic flow are modeled as partial differential equations:

[0123]

[0124] Where F represents the traffic flow field, which includes traffic volume (scalar) and velocity (vector) components, and v is the field viscosity coefficient, which is used to control the diffusion speed (initial value is 0.1 and is adjusted based on the smoothness of the traffic flow). is a diffusion term, which is used to simulate the spread of traffic flow from high-density areas (such as city centers) to low-density areas (such as suburbs). is the convection term, which is used to capture the directional propagation of traffic (such as flow transfer from main roads to branch roads); is a pressure term, which is used to reflect external constraints (such as signal light cycle and road capacity limit);

[0125] It should be noted that in a traffic network, traffic flow can be analogized to scalar potential (diffusion characteristics), speed can be analogized to vector potential (directional characteristics), and pressure terms can be defined by traffic rules (such as traffic lights) or road capacity;

[0126] In addition, the partial differential equation can be decomposed into three subproblems, dealing with diffusion, convection, and pressure terms respectively, to reduce the computational complexity;

[0127] Specifically, the diffusion term can be discretized by the graph Laplacian operator, and the neighborhood weighted difference of each node can be calculated to simulate the smooth propagation of traffic flow;

[0128] Specifically, the convection term can be discretized by the upwind difference method, and the flow is preferentially propagated to downstream nodes (e.g., from the main road to the branch road) according to the velocity direction;

[0129] Specifically, the pressure term can be solved by the Poisson equation, with the traffic light cycle (the flow rate is 0 when the light is red) or the road capacity (the upper limit flow rate) as the boundary condition, and the pressure field is solved iteratively;

[0130] It should be noted that if the above decomposition term processing method is adopted, the traffic flow field needs to be updated according to the processed diffusion term, convection term, and pressure term;

[0131] In addition, traffic rules and environmental factors can also be introduced:

[0132] Traffic light cycle: During the red light period, the node flow is set to 0; during the green light period, the flow is updated according to the actual value;

[0133] Road capacity: Set a maximum traffic flow rate for each node (e.g., 1,000 vehicles / hour), and cut off any traffic exceeding that limit;

[0134] Emergency events: such as traffic accidents, simulate traffic blockage by reducing the edge weight of the affected nodes (e.g., reducing it by 50%);

[0135] It should be noted that the above-mentioned traffic rules and environmental factors are embedded in the pressure items and updated synchronously;

[0136] Furthermore, based on the gradient characteristics of diffusion, a scalar potential convolution kernel is constructed to handle the diffusion of traffic flow from high-flow areas to low-flow areas;

[0137] Specifically, based on the node features and diffusion terms, a scalar potential convolution kernel G is constructed φ :

[0138]

[0139] Where h is the aforementioned node feature;

[0140] It should be noted that the scalar potential convolution kernel is constructed to simulate the diffusion of traffic flow from high-flow nodes (such as downtown intersections) to low-flow nodes (such as suburban road sections);

[0141] Furthermore, based on the curl characteristics of direction changes, a vector potential convolution kernel is constructed to handle the interaction of traffic flow directions at the location point;

[0142] Specifically, based on the node characteristics and the curl characteristics of the direction change, a vector potential convolution kernel G is constructed. A :

[0143]

[0144] It should be noted that the vector potential convolution kernel is constructed to simulate the change of vehicle speed direction at the intersection (such as turning or diverging);

[0145] In addition, for the constructed scalar potential convolution kernel and vector potential convolution kernel, the regularization L2 can be set to prevent excessive smoothing of features and distortion of directional features;

[0146] Furthermore, through feature fusion, the convolution results of the scalar potential convolution kernel and the vector potential convolution kernel are combined to generate the propagation kernel;

[0147] Specifically, the results of the scalar potential convolution kernel and the vector potential convolution kernel are combined to generate the propagation kernel G conv :

[0148] G conv =αG φ +γG A

[0149] Among them, α,γ∈[0,1] represent adjustment coefficients, and their initial values ​​are both 0.5;

[0150] S4. Constructing a spatiotemporal propagation mechanism based on the edge weights and propagation kernel, and updating node features by combining a coupled propagation method with a time series model;

[0151] Furthermore, based on the propagation kernel, the node features are updated using a long short-term memory network;

[0152] Specifically, the node characteristics of the input propagation kernel output Based on the aforementioned edge weights, we get the neighborhood aggregation feature z i :

[0153]

[0154] Furthermore, we use the LSTM model to update node features:

[0155]

[0156] in, is the node feature of the current layer, is a historical feature sequence obtained from historical traffic data;

[0157] Furthermore, the gate mechanism of the LSTM model includes:

[0158] Input gate: Based on edge weights, controls neighborhood aggregation feature z i The inflow ratio;

[0159] Forget Gate: Dynamically adjusts the retention of historical information based on node characteristics and traffic flow changes;

[0160] Output gate: fuses current and historical features to generate updated node features;

[0161] Specifically, the number of units in the LSTM model is consistent with the dimension of the input features, and the hidden layer dimension is 16. The model uses backpropagation to optimize weights during training, with a learning rate of 0.001 (using the Adam optimizer);

[0162] S5. Using the density of the traffic flow field and node features, a dynamic range of action is introduced to adjust the propagation range to adapt to traffic flow changes in sparse areas;

[0163] Furthermore, according to the node characteristics, the node density is calculated based on the traffic flow information distribution of the neighboring nodes;

[0164] Specifically, we input node features, edge weights, neighborhood structure, and traffic flow information, and calculate node density based on the traffic flow information distribution of neighborhood nodes to reflect the local information density:

[0165] p i =ε·|N(i)|+δ·info i

[0166] Among them, p i It is expressed as node density, ε and δ are fusion coefficients, with initial values ​​of 0.4 and 0.6 respectively, which are used to emphasize the contribution of traffic flow information. i It represents the distribution intensity of traffic flow information, which is obtained by calculating the weighted variance of vehicle volume and speed within the neighborhood node;

[0167] Specifically, for high-density nodes, such as intersections in the city center, there are many neighboring nodes, and the traffic volume and speed vary significantly (high variance), so the density value is high; for low-density nodes, such as sparse nodes in the suburbs, there are few neighboring nodes, and the traffic flow information is evenly distributed (low variance), so the density value is low;

[0168] Furthermore, based on the traffic flow field, a mapping relationship between node density and dynamic action range is established;

[0169] It should be explained that the purpose of setting a dynamic range is to adaptively adjust the scope of information dissemination according to the node density and traffic flow field, so as to adapt to the different topological and traffic characteristics of sparse areas (such as suburbs) and dense areas (such as city centers);

[0170] Specifically, the mapping relationship between node density and dynamic range is established as follows:

[0171] r i =r max ·(1-η·p′ i )+κ·||F f (f i )||2

[0172] Among them, r max is the maximum range of action, and its initial value is twice the road length; η represents the density influence coefficient, which ranges from [0, 1] and has an initial value of 0.7, and is used to control the scaling of density on range; κ is the traffic flow field influence coefficient, which ranges from [0, 1] and has an initial value of 0.3, and is used to enhance the regulation of flow intensity on range; p′ i represents the normalized node density;

[0173] Furthermore, the scope of action is dynamically adjusted according to the node density and traffic field to ensure that the propagation range adapts to the topology and traffic changes in sparse areas;

[0174] Specifically, the adjustment mechanism is: sparse area: for low-density nodes (p′ i <0.5) increases the scope of action and expands the node neighborhood to obtain more information about the labeled nodes; dense area: for high-density nodes (p′ i ≥0.5) Narrow the scope of action and give priority to using local traffic flow information;

[0175] Specifically, if the proportion of labeled nodes in the neighborhood of a node is low (|N(i)∩L| / |N(i)|<0.3), the range of action r′ is enlarged. i =r i ·(1+θ·(1-|N(i)∩L| / |N(i)|)), θ=0.5 is the amplification factor, L is the set of labeled nodes; if the traffic field intensity suddenly changes (such as an emergency, ||F f (f i )||2 increases by 50%), temporarily enlarging the range r′ i =r i ·(1+θ·Δ||F f (f i )||2 / ||F f (f i )||2),θ=0.2,Δ||F f (f i )||2 is the intensity change of traffic flow field;

[0176] In addition, the scope can be limited to [r min ,r max ] to ensure the rationality of the scope of action;

[0177] S6. Using the traffic flow field as a confidence indicator and combining it with the node features, propagating the traffic flow information of the labeled nodes to sparse nodes to achieve automatic labeling;

[0178] Specifically, based on the traffic flow field, the confidence of each node is calculated;

[0179] Furthermore, the statistical distribution of the confidence of all nodes (mean μ, standard deviation σ) is calculated. Since confidence belongs to conventional technical means, it will not be described in detail here;

[0180] Furthermore, the confidence level is used to set a confidence threshold, and when the confidence threshold is met, the traffic flow information of the node with the highest confidence level is preferentially transmitted;

[0181] Specifically, the confidence threshold τ = μ + k·σ is set according to the mean and standard deviation of the statistical distribution, where k is the selection coefficient in the range of [0.5, 1.5] to ensure that high-confidence nodes are selected;

[0182] Specifically, the labeled nodes that meet the confidence level greater than or equal to the confidence threshold are screened, and the screened nodes are sorted from high to low according to the confidence level, and the node information with the highest confidence level is propagated first;

[0183] In addition, the sorting method can adopt heap sorting to reduce the complexity of sorting;

[0184] Furthermore, based on the traffic flow information of high-confidence nodes, a Poisson equation is constructed. The traffic flow information of the labeled nodes is used as the boundary condition, and the multi-grid method is used to iteratively solve the problem and propagate the annotations to the sparse nodes.

[0185] Specifically, Poisson's equation is expressed as:

[0186]

[0187] Where u represents the target value of the node (traffic flow, speed); ρ represents the source term, which is obtained based on the change of traffic flow intensity;

[0188] Specifically, the boundary conditions are: for the labeled nodes, the target labeling value of the node is set to the traffic flow or speed; for the unlabeled nodes, the target labeling value of the node is set to 0;

[0189] Furthermore, the convergence of the Poisson equation can be accelerated by using the multigrid method to iteratively solve the Poisson equation;

[0190] Specifically, a multi-layer graph grid (3 to 5 layers) is constructed through a coarsening operation, from a fine grid (the original node graph) to a coarse grid (aggregated nodes). Spectral clustering is then used to cluster the 1,000-node graph into a coarse grid of 100 nodes. Smoothing is then performed, using Gauss-Seidel iterations on each grid layer. Finally, the correction is solved on the coarse grid by calculating the residual, and then interpolated back to the fine grid. This process is repeated 5 to 10 times until the residual converges, resulting in a solution to the Poisson equation.

[0191] It should be noted that since the multi-grid method is a tool for solving the Poisson equation in the solution of the present invention, it will not be discussed in more depth here;

[0192] Specifically, the solution of the Poisson equation is the label value of the unlabeled node, while for the labeled node, the solution of the Poisson equation is the predicted label value of the labeled node. The predicted label value of the labeled node is compared with the label value of the original labeled node, and the judgment is made by calculating the mean square error (MSE);

[0193] It should be noted that traffic field-guided sparse node labeling realizes information transfer from labeled nodes to unlabeled nodes through confidence evaluation and Poisson equation propagation, overcoming the limitation of traditional interpolation methods that ignore nonlinear dynamics. Compared with the static propagation mechanism, the use of traffic flow fields to dynamically adjust the propagation path can improve the accuracy and robustness of sparse node labeling.

[0194] Furthermore, this embodiment also provides a sparse data automatic annotation system based on graph neural network, including:

[0195] The spatiotemporal graph modeling and multi-source feature embedding module is configured to model the traffic network as a spatiotemporal graph structure, embed and map traffic flow and speed features into the graph structure through feature mapping methods, and generate node features and edge weights;

[0196] a traffic flow field decomposition and dynamic edge weight reconstruction module, configured to construct a traffic flow field using the node features and edge weights, generate diffusion features and directional features based on local changes in vehicle volume and speed, and update edge weights;

[0197] a dual-channel graph convolution kernel generation module, configured to design a graph convolution mechanism based on the diffusion feature and the directional feature, process the diffusion and directional interaction of the traffic flow, and generate a propagation kernel through feature fusion;

[0198] a spatiotemporal coupled propagation and time series integration module, configured to construct a spatiotemporal propagation mechanism based on the edge weights and propagation kernels, and update node features by combining a time series model with a coupled propagation method;

[0199] a dynamic field adaptive adjustment module configured to utilize the density of the traffic flow field and node features, introduce a dynamic range of action, and adjust the propagation range to adapt to traffic flow changes in sparse areas;

[0200] The confidence-guided annotation propagation module is configured to use the traffic flow field as a confidence indicator and, in combination with the node features, propagate the traffic flow information of the annotated nodes to sparse nodes to achieve automatic annotation.

[0201] This embodiment further provides a computer device applicable to the sparse data automatic labeling method based on a graph neural network, including:

[0202] Memory and processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the sparse data automatic labeling method based on graph neural network as proposed in the above embodiment.

[0203] The computer device may be a terminal, comprising a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0204] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the automatic labeling method for sparse data based on graph neural network proposed in the above embodiment is implemented.

[0205] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiment belong to the same inventive concept. Technical details not fully described in this embodiment can be found in the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0206] Those skilled in the art will appreciate that the embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Furthermore, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present application may be implemented in various computer languages, for example, object-oriented programming language Java and interpreted scripting language JavaScript, etc.

[0207] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0208] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0209] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0210] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0211] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A sparse data automatic labeling method based on graph neural network, characterized by: include: The traffic network is modeled as a spatiotemporal graph structure. The traffic flow and speed features are embedded into the graph structure through feature mapping method to generate node features and edge weights. Constructing a traffic flow field using the node features and edge weights, generating diffusion features and directional features based on local changes in vehicle volume and speed, and updating edge weights; A graph convolution mechanism is designed based on the diffusion and direction features to handle the diffusion and direction interactions of traffic flow, and a propagation kernel is generated through feature fusion. Based on the edge weights and propagation kernels, a spatiotemporal propagation mechanism is constructed, and node features are updated by combining a coupled propagation method with a time series model; By utilizing the density of the traffic flow field and node features, a dynamic range of action is introduced to adjust the propagation range to adapt to traffic flow changes in sparse areas; The traffic flow field is used as a confidence indicator and combined with the node features to propagate the traffic flow information of the labeled nodes to sparse nodes to achieve automatic labeling.

2. The sparse data automatic labeling method based on graph neural network according to claim 1, characterized in that: Modeling the transportation network as a spatiotemporal graph structure, including: The nodes in the spatiotemporal graph structure represent location points in the transportation network, and the edges in the spatiotemporal graph structure represent connections between the location points; Collect real-time traffic flow information from sensors, including vehicle volume and speed, and combine it with historical traffic patterns to generate node features; By embedding the traffic flow information into a low-dimensional space, the geometric characteristics of the traffic flow are preserved. The edge weights are calculated based on the connection distance between location points and the similarity of traffic flows.

3. The sparse data automatic labeling method based on graph neural network according to claim 2, characterized in that: The traffic flow field is constructed by using the node features and edge weights, including: According to the node characteristics, the difference of traffic flow information between nodes is calculated to determine the traffic flow field; The traffic flow field is decomposed into a diffusion gradient feature and a direction change curl feature, and the edge weights are updated simultaneously.

4. The sparse data automatic labeling method based on graph neural network according to claim 3 is characterized in that: The graph convolution mechanism is designed based on the diffusion and directional features, including: Based on the gradient characteristics of diffusion, a scalar potential convolution kernel is constructed to handle the diffusion of traffic flow from high-flow areas to low-flow areas; Based on the curl characteristics of direction changes, a vector potential convolution kernel is constructed to process the interaction of traffic flow directions at the location point; Through feature fusion, the convolution results of the scalar potential convolution kernel and the vector potential convolution kernel are combined to generate the propagation kernel.

5. The sparse data automatic labeling method based on graph neural network according to claim 2 or 4, characterized in that: A spatiotemporal propagation mechanism is constructed based on the edge weights and propagation kernels, including: Based on the propagation kernel, the node features are updated using a long short-term memory network.

6. The sparse data automatic labeling method based on graph neural network according to claim 5, characterized in that: The introduction of a dynamic scope includes: Calculating node density based on the traffic flow information distribution of neighboring nodes according to the node characteristics; According to the traffic flow field, the dynamic action range is set to change with the node density.

7. The sparse data automatic labeling method based on graph neural network according to claim 1, characterized in that: Using the traffic flow field as a confidence indicator and combining it with the node features, the traffic flow information of the labeled nodes is propagated to the sparse nodes, including: Based on the traffic flow field, calculate the confidence of each node; Using the confidence level, a confidence threshold is set, and when the confidence threshold is met, traffic flow information of the node with the highest confidence level is preferentially transmitted; According to the node characteristics, the Poisson equation is constructed, and the traffic flow information of the labeled nodes is used as the boundary condition. The multi-grid method is used to iteratively solve the problem and propagate the labels to the sparse nodes.

8. A sparse data automatic annotation system based on a graph neural network, based on the sparse data automatic annotation method based on a graph neural network according to any one of claims 1 to 7, characterized in that: include: The spatiotemporal graph modeling and multi-source feature embedding module is configured to model the traffic network as a spatiotemporal graph structure, embed and map traffic flow and speed features into the graph structure through feature mapping methods, and generate node features and edge weights; a traffic flow field decomposition and dynamic edge weight reconstruction module, configured to construct a traffic flow field using the node features and edge weights, generate diffusion features and directional features based on local changes in vehicle volume and speed, and update edge weights; a dual-channel graph convolution kernel generation module, configured to design a graph convolution mechanism based on the diffusion feature and the directional feature, process the diffusion and directional interaction of the traffic flow, and generate a propagation kernel through feature fusion; a spatiotemporal coupled propagation and time series integration module, configured to construct a spatiotemporal propagation mechanism based on the edge weights and propagation kernels, and update node features by combining a time series model with a coupled propagation method; a dynamic field adaptive adjustment module configured to utilize the density of the traffic flow field and node features, introduce a dynamic range of action, and adjust the propagation range to adapt to traffic flow changes in sparse areas; The confidence-guided annotation propagation module is configured to use the traffic flow field as a confidence indicator and, in combination with the node features, propagate the traffic flow information of the annotated nodes to sparse nodes to achieve automatic annotation.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images