Discontinuous perception area traffic flow estimation method, device and product

Abstract

The embodiment of the invention relates to the technical field of information, and discloses a discontinuous perception area traffic flow estimation method, device and product, and the method comprises the steps: obtaining traffic flow observation data of a target area, and constructing a mask matrix of a missing value position in the traffic flow observation data; calculating a time delay matrix between observation time points; inputting the traffic flow observation data and the mask matrix into a time feature extraction network, and processing by using a long short-term memory network to obtain hidden variables; generating a time decay estimation value containing the time dimension feature; filling missing positions in the traffic flow observation data by using the time attenuation estimation value, and constructing pre-filled graph structure data; inputting the pre-filled graph structure data into a graph attention network, aggregating traffic flow information of adjacent vertexes based on a graph attention mechanism, and generating a spatial dependency estimation value; and fusing the time attenuation estimation value and the spatial dependence estimation value by using a tradeoff coefficient to obtain traffic flow complementation data, thereby making up for the deficiency of traffic situation index data.

Description

Technical Field

[0001] This application relates to the field of traffic information, and in particular to a method, device and product for estimating traffic flow in discontinuous sensing areas. Background Technology

[0002] Traffic situation indicators (such as flow rate, speed, and density) are key data for quantifying the operational status of a transportation system, providing scientific support for planning and decision-making by traffic management departments and for travel services for the public. In Intelligent Transportation Systems (ITS), accurate and continuous traffic flow data is fundamental to achieving efficient traffic control.

[0003] However, in actual transportation networks, data acquisition often faces numerous difficulties. On the one hand, existing sensing data sources are relatively singular, and the deployment of roadside sensing devices (such as loop detectors, radar, and video surveillance) is usually sparse, failing to cover all road segments and resulting in numerous natural sensing blind spots. On the other hand, even on road segments where sensing devices have been deployed, due to uncontrollable factors such as detector malfunctions, weak or interrupted network signals, and interference from severe weather, the collected data often suffers from severe packet loss or long-term data gaps.

[0004] The aforementioned issues result in significant discontinuities and uneven distributions in the temporal and spatial distribution of existing traffic situation indicator data. Existing conventional interpolation methods often struggle to simultaneously address both of these complex missing data scenarios.

[0005] For road sections with sensing devices but missing data, traditional methods (such as mean imputation, simple interpolation, or conventional neural networks) often ignore the temporal correlation decay caused by long-term data loss, or fail to effectively handle missing values ​​of adjacent nodes when performing spatial aggregation, resulting in insufficient imputation accuracy.

[0006] For road sections without any sensing devices, conventional data-driven incomplete methods are not applicable due to the lack of historical observation data, making it difficult to predict the traffic conditions in the area.

[0007] Therefore, how to construct a traffic flow estimation scheme that can simultaneously solve the problems of partial data loss and complete lack of data coverage for discontinuous sensing areas is an urgent problem to be solved in the field of intelligent transportation. Summary of the Invention

[0008] One objective of this application is to provide a method, device, and product for estimating traffic flow in discontinuous sensing areas, at least to solve the technical problem in the prior art where traffic situation index data is discontinuous and unevenly distributed due to reasons such as sparse deployment of roadside sensing devices, detector failures, and unstable network transmission, thus failing to accurately reflect the overall traffic operation status.

[0009] To achieve the above objectives, some embodiments of this application provide the following aspects:

[0010] This application provides a traffic flow estimation method for discontinuous sensing areas, the method comprising:

[0011] Obtain traffic flow observation data collected by sensors in the traffic network of the target area, define the traffic network as a graph structure, where road segments are vertices and the spatial correlation between road segments are edges, and construct a mask matrix indicating the location of missing values ​​in the traffic flow observation data;

[0012] Based on the timestamps of the traffic flow observation data, a time lag matrix is ​​calculated between the observation time points. The time lag matrix represents the time interval between the current missing data and the most recent valid observation data.

[0013] The traffic flow observation data and the mask matrix are input into a time feature extraction network, and latent variables are obtained by processing them using a long short-term memory network.

[0014] The latent variables are filtered using a time decay gate generated based on the time delay matrix, thereby generating a time decay estimate that includes time dimension features;

[0015] The time decay estimate is used to fill in the missing locations in the traffic flow observation data to construct pre-filled graph structure data;

[0016] The pre-filled graph structure data is input into a graph attention network, and traffic flow information of adjacent vertices is aggregated based on the graph attention mechanism to generate spatial dependency estimates.

[0017] The time decay estimate and the spatial dependence estimate are fused using a tradeoff coefficient to obtain traffic flow completion data, and complete traffic situation data including the completed flow, speed or density indicators are output.

[0018] Secondly, some embodiments of this application also provide an electronic device, the electronic device comprising: one or more processors; and a memory storing computer program instructions, which, when executed, cause the processor to perform the steps of the method described above.

[0019] Thirdly, some embodiments of this application also provide a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the method described above.

[0020] Compared with related technologies, the solution provided in this application firstly introduces a time delay matrix and a time decay gate based on it to address the data loss problem under non-uniform time intervals. By explicitly controlling the filtering mechanism of the latent variables in the Long Short-Term Memory (LSTM) network through physical time intervals, the model can perceive the duration of data loss, thereby physically simulating the decay of the reliability of historical information over time. This mechanism effectively solves the problem that the influence weight of outdated historical information on the current moment cannot be reasonably adjusted in scenarios with long-term continuous data loss, making the feature extraction in the time dimension more consistent with the physical evolution of traffic flow.

[0021] Secondly, the missing positions are pre-filled using time decay estimates generated by a time-prior-based serial imputation strategy, constructing graph structure data containing prior information before inputting it into the graph attention network. This strategy ensures that when the graph attention network aggregates information from neighboring nodes, spatial information propagation is not interrupted or noise is introduced due to missing or zero values ​​in neighboring nodes, thus achieving deep coupling and complementarity between temporal continuity and spatial correlation.

[0022] By using the mask matrix and time decay gate as input features, the model can learn to generate a dynamic tradeoff coefficient, intelligently deciding whether to rely more on temporal or spatial inference based on the current context. Simultaneously, by combining a bidirectional feature extraction structure with an imputation consistency loss, the system can constrain the consistency of forward and backward imputation results, avoiding deviations in bidirectional inference logic and further improving the reliability of the results.

[0023] Finally, by organically coupling a data-driven repair method for partially missing road segments with a model-driven inference method for completely unperceived road segments, the physical inversion of blind spot conditions was achieved using measured data from boundary road segments and macroscopic traffic flow models. These two mechanisms complement each other, effectively overcoming the technical bottleneck of existing technologies that struggle to simultaneously address data gaps caused by detector malfunctions and perception blind spots caused by equipment sparsity. This enables the continuity and completeness of regional traffic flow indicators in the spatiotemporal dimensions. Attached Figure Description

[0024] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0025] Figure 1 A flowchart of a traffic flow estimation method for a discontinuous sensing area provided as an exemplary embodiment of this disclosure;

[0026] Figure 2A flowchart of another traffic flow estimation method for discontinuous sensing areas provided as an exemplary embodiment of this disclosure;

[0027] Figure 3 An exemplary structural diagram of the electronic device provided for some embodiments of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] Terminology Explanation

[0030] LSTM (Long Short-Term Memory): Long Short-Term Memory network.

[0031] GAT (Graph Attention Network): Graph attention network.

[0032] Figure 1 A flowchart of a traffic flow estimation method for a discontinuous sensing area provided as an exemplary embodiment of this disclosure, the method comprising:

[0033] S101. Obtain traffic flow observation data collected by sensors in the target area traffic network, define the traffic network as a graph structure, where road segments are vertices and the spatial correlation between road segments is an edge, and construct a mask matrix indicating the location of missing values ​​in the traffic flow observation data.

[0034] Specifically, to mathematically model the topology of the traffic network, it is defined as an undirected graph structure. Road segments are defined as the set of vertices, and the spatial adjacency between road segments is defined as the set of edges. The adjacency matrix of the graph is constructed based on the connectivity between sensors. Simultaneously, a mask matrix with the same dimension as the collected traffic flow observation data is constructed. This mask matrix is ​​used to accurately indicate the data status; for example, a location is marked as 1 when it is an observed value and 0 when it is a missing value, thus clearly distinguishing between valid information and information to be filled.

[0035] S102. Based on the timestamps of the traffic flow observation data, calculate the time lag matrix between observation time points. The time lag matrix represents the time interval between the current missing data and the most recent valid observation data.

[0036] Specifically, considering that missing traffic data is often not uniformly distributed, in order to capture the data characteristics under non-uniform time intervals, a time lag matrix can be calculated based on the timestamps of traffic flow observation data to determine the time lag between observation points. This time lag matrix does not record simple index differences, but rather represents the physical time interval between the currently missing data and the most recent valid observation data.

[0037] S103. Input the traffic flow observation data and the mask matrix into the time feature extraction network, and use the long short-term memory network to process and obtain the latent variables.

[0038] S104. The latent variables are filtered using a time decay gate generated based on the time delay matrix, thereby generating a time decay estimate containing time dimension features.

[0039] Specifically, to process the data more precisely, the temporal feature extraction network employs a dual-channel Long Short-Term Memory (LSTM) network structure. The actual value channel specifically processes traffic flow observation data to extract numerical change features, while the mask channel specifically processes the mask matrix to extract missing pattern features. The latent variables of these two channels interact through a fusion layer, filtering the actual value latent variables to remove interference from invalid information. Building upon this, step S104 introduces a time decay mechanism. Using the time delay matrix generated in step S102, a time decay gate with a value between 0 and 1 is calculated. This time decay gate is multiplied element-wise with the latent variables output by the LSTM, simulating the physical process of information reliability decreasing over time, thereby generating a time decay estimate containing accurate temporal dimension features.

[0040] S105. Use the time decay estimate to fill in the missing positions in the traffic flow observation data to construct pre-filled graph structure data.

[0041] S106. Input the pre-filled graph structure data into the graph attention network, aggregate traffic flow information of adjacent vertices based on the graph attention mechanism, and generate spatial dependency estimates.

[0042] Specifically, to address the issue of missing neighbor node data during spatial feature extraction, step S105 employs a serial imputation strategy. The generated time decay estimates are used to fill in the missing locations in the traffic flow observation data, constructing a "pre-filled" graph structure. This operation ensures that subsequent graph networks receive valid values ​​calibrated for the time dimension as input during convolution or aggregation, rather than zero values ​​or noise. Next, in step S106, this pre-filled graph structure data is input into a graph attention network (GAT). Based on the graph attention mechanism, the model can dynamically calculate the attention coefficients between the current vertex and its neighboring vertices, and aggregate the traffic flow information of neighboring vertices accordingly. To preserve the original temporal estimation information and prevent network degradation, the aggregated spatial information is fused with the time decay estimates through residual connections to generate spatial dependency estimates.

[0043] S107. The time decay estimate and the spatial dependence estimate are fused using a trade-off coefficient to obtain traffic flow completion data, and complete traffic situation data including the completed flow, speed or density indicators is output.

[0044] Specifically, in step S107, a tradeoff coefficient is used to fuse the time decay estimate and the spatial dependence estimate. This tradeoff coefficient is not manually set, but is generated by a neural network through learning, using a mask matrix and a time decay gate as input features. This coefficient adaptively determines whether the model should rely more on temporal or spatial inferences under specific missing patterns and time intervals. The fused result constitutes unidirectional traffic flow completion data. This outputs complete traffic situation data containing the completed flow, speed, or density indicators, achieving high-precision estimation of traffic flow in discontinuous sensing areas.

[0045] In the above embodiments, firstly, a physical time-based forgetting mechanism is constructed within the neural network by introducing a time delay matrix and a time decay gate. This mechanism enables the model to explicitly perceive the time span of the last effective observation and dynamically decays historical latent variables accordingly. Thus, in scenarios where data is continuously missing for a long period, it effectively reduces the misleading influence of outdated historical information on the current estimate, significantly improving the physical rationality and prediction accuracy of the imputation results.

[0046] Secondly, a sequential filling strategy of "time inference first, spatial aggregation last" is adopted, which uses high-confidence time decay estimates to pre-fill the graph structure. This strategy provides the graph attention network with an input base containing effective prior information, ensuring that the features of neighboring nodes can be effectively utilized when performing neighborhood aggregation, and realizing deep coupling and complementarity of spatiotemporal features.

[0047] Finally, a dynamic tradeoff coefficient determined by both the data missing pattern and the duration of the missing data is introduced to achieve adaptive switching of the inference mechanism. The model can automatically focus on time-based inference in the early stages of data missing data, while automatically switching to spatial-based inference when the missing time is long but the spatial neighbors are complete. This allows it to flexibly and robustly cope with various complex random or block-like missing data patterns.

[0048] Furthermore, in one embodiment, the mask matrix in step S101 is represented by the following formula:

[0049] If the road network is Traffic data is collected by data sensors at various locations, and the road network is defined as an undirected graph. ; This represents the set of vertices for road segments in a road network, with the number of vertices being... ; A graph represents the set of edges, and a graph represents the spatial dependence between vertices. The adjacency matrix is ​​composed of matrix Build, Represents the set of real numbers. If Then the sensor With sensors They are adjacent; if Then the sensor With sensors They are not adjacent. At each timestamp The real traffic flow data collected by all sensors in the transportation network is denoted as All timestamps The data collected internally is recorded as Introducing a mask matrix Record The location of missing values ​​in; when When it is an observation value, ;when When the value is missing, The mask matrix is ​​defined as follows:

[0050]

[0051] Furthermore, based on the mask matrix, non-missing observations are considered as valid information, and a historical observation variable is defined. , Used to measure historical observation data Can be used to complete How much effective information is provided? The formula for calculating historical observation variables is as follows:

[0052]

[0053] In the above embodiments, by constructing the traffic network as a graph structure containing vertices and edges, and using an adjacency matrix to describe the adjacency relationships between sensors, the complex real-world road topology can be effectively transformed into a computer-processable mathematical model. This allows for the accurate capture of the spatial connectivity and correlation of different road segments, laying a solid data foundation for subsequent mining of spatial dependency features between road segments. Simultaneously, this embodiment introduces a mask matrix to explicitly mark true observations and missing values, clarifying the validity boundaries of the data and effectively preventing the subsequent model from misjudging missing locations as valid zero-value signals during feature extraction, thus ensuring the purity of the input data. More importantly, this embodiment defines historical observation variables to measure the contribution of historical data to the current completion, enabling the model to no longer blindly rely on all historical data, but rather dynamically evaluate its reference value based on the time span since the last valid observation. This provides crucial physical prior information for solving the data missing problem under non-uniform time intervals, significantly improving the logical rationality and prediction accuracy of traffic flow estimation in long-term continuous missing scenarios.

[0054] In one embodiment, the step of calculating the time lag matrix between observation time points based on the timestamps of the traffic flow observation data specifically includes:

[0055] Calculate the forward time delay matrix and the reverse time delay matrix respectively;

[0056] The elements in the positive time delay matrix represent the time difference from the most recent historical observation point to the current missing time point;

[0057] The elements in the inverse time delay matrix represent the time difference from the most recent future observation point to the current missing time point;

[0058] The time feature extraction network adopts a bidirectional structure, performing forward imputation and backward imputation based on the forward time delay matrix and the reverse time delay matrix respectively, and generating forward imputation results and backward imputation results.

[0059] Specifically, to accurately capture the non-uniform distribution characteristics of missing traffic flow data over time, a time lag matrix is ​​calculated based on the timestamps of the traffic flow observation data. This time lag matrix includes a forward time lag matrix and a reverse time lag matrix. The forward time lag matrix measures the freshness of historical information, with its elements representing the time difference from the most recent historical observation time to the current missing time. The reverse time lag matrix assesses the reference value of future information, with its elements representing the time difference from the most recent future observation time to the current missing time. Based on this, the temporal feature extraction network is designed with a bidirectional structure, capable of processing bidirectional information in parallel: performing forward imputation based on the forward time lag matrix and backward imputation based on the reverse time lag matrix. Through this bidirectional processing mechanism, the network ultimately generates independent forward and backward imputation results, providing fundamental data support for the subsequent comprehensive utilization of past and future spatiotemporal information.

[0060] Furthermore, in one embodiment, forward filling and backward filling are performed based on the forward time delay matrix and the reverse time delay matrix, respectively, specifically including:

[0061] By leveraging the temporal correlation between observed and missing data, missing data can be filled in from a temporal perspective. The contribution of observations to the estimation of missing values ​​depends on the time interval between them, especially when data is continuously missing along the time axis. Therefore, a matrix is ​​introduced. Record time interval information. Among them, This represents the time difference from the most recent observation to the current unobserved value. Missing data is filled in both the forward and reverse directions, with each... Includes positive time difference and reverse time difference Positive time difference This can be expressed by the following formula:

[0062]

[0063] In the formula, This indicates the sampling time point of the data. Similarly, when calculating the reverse time difference... At that time, the timestamps are arranged in reverse order. .

[0064] In the above embodiments, by introducing forward and reverse time delay calculation mechanisms based on sampling time points, the model's perception accuracy and physical plausibility for missing data in non-uniform time intervals are significantly improved. Traditional processing methods based on sequence index step size often ignore the passage of actual physical time and cannot accurately reflect the changes in data correlation after long-term interruptions. This embodiment, however, uses the forward time difference calculated from the actual sampling time points to objectively characterize the degree of correlation decay of the influence of historical valid observation data on the current missing location over time. Simultaneously, by calculating the reverse time difference using reversed timestamps, the strength of the feedback constraint of future valid observation data on the current moment can be quantified. This bidirectional physical time quantification mechanism provides accurate prior parameters regarding the spatiotemporal correlation of data for the subsequent time feature extraction network, enabling the model to dynamically adjust its focus on past or future information based on the length of actual physical time, thereby significantly improving the accuracy and robustness of traffic flow estimation in discontinuous sensing areas.

[0065] In one embodiment, the step of inputting the traffic flow observation data and the mask matrix into a time feature extraction network and processing them using a long short-term memory network to obtain latent variables specifically includes:

[0066] Construct a dual-channel long short-term memory network, including an actual value channel and a mask channel;

[0067] The traffic flow observation data is processed using the actual value channel to obtain the actual value latent variables;

[0068] The mask matrix is ​​processed using the mask channel to obtain the mask hidden variables;

[0069] The actual value latent variable and the mask latent variable are fused through a fusion layer. The mask latent variable is processed by an activation function and then used as a filter to perform feature filtering on the actual value latent variable to obtain the latent variable.

[0070] Specifically, to handle the relationship between traffic flow observation data and missing patterns more precisely, the step of inputting traffic flow observation data and the mask matrix into the time feature extraction network for processing employs a unique dual-channel long short-term memory network structure. First, a dual-channel network containing an actual value channel and a mask channel is constructed. These two channels work in parallel but undertake different feature extraction tasks. Using the actual value channel, traffic flow observation data is received and processed to capture the dynamic characteristics of traffic flow values ​​changing over time, thereby outputting actual value latent variables. Simultaneously, the mask channel processes the mask matrix, specifically for learning the distribution patterns and rules of missing data, outputting mask latent variables. To effectively integrate the actual value latent variables and the mask latent variables, a fusion layer is set up in the model. In this layer, the mask latent variables are first processed by an activation function (such as Sigmoid), converted into values ​​with "gated" properties (i.e., filters), and then interact with the actual value latent variables. This process uses the learned missing patterns to filter numerical features, retaining the features of valid observation points and suppressing noise from missing points, thus obtaining pre-cleaned fused latent variables.

[0071] In the above embodiments, the model can retain the effective information in the actual value latent variables according to the missing data state, and suppress or discard invalid features or noise introduced by missing values, and finally obtain the cleaned and filtered latent variables.

[0072] In one embodiment, the step of filtering the latent variables using a time decay gate generated based on the time delay matrix to generate a time decay estimate containing time dimension features specifically includes:

[0073] Based on the time interval values ​​in the time delay matrix, a time decay gate with a value between 0 and 1 is calculated.

[0074] The time decay gate is used to multiply the output of the fusion layer element by element to obtain the filtered time dimension features.

[0075] Specifically, to further address the impact of non-uniform time intervals, the time delay matrix calculated in the previous steps is invoked. Based on the physical time interval values ​​recorded in the time delay matrix, a time decay gate with a value between 0 and 1 is generated. Then, this time decay gate is used to perform element-wise multiplication on the output of the aforementioned fusion layer (i.e., the initially cleaned latent variables). This operation is equivalent to applying a physical time-based "forgetting factor" at the feature level, assigning lower weights to historical information further back in time, thereby obtaining a time decay estimate that is calibrated for the time dimension and contains accurate time-dimensional features.

[0076] Furthermore, in one embodiment, the determination of the time decay estimate is primarily based on the temporal correlation learned from historical observation data, combined with the missing data pattern for decay estimation. Considering the heterogeneity of the input data—that is, the actual observations are numerical and the missing data are Boolean—these two are treated as two independent features and processed separately.

[0077] Specifically, an actual value matrix is ​​constructed to record the collected real data, while a mask matrix is ​​constructed to record the missing data distribution pattern. In the feature extraction stage, two standard Long Short-Term Memory (LSTM) networks are used, i.e., using... and Two standard LSTMs process the actual value matrix and the mask matrix, respectively. and The abbreviation for hidden variables is and Taking the forward propagation direction as an example:

[0078]

[0079]

[0080] In the formula, yes Forward filling, This is the mask vector. Forward imputation is the process of filling missing values ​​in the data by using the nearest known valid value before the missing value to fill the missing position. At the initial moment of computation, i.e., when... At that time, and Randomly initialized to and Used to calculate .

[0081] After extracting features separately, a fusion layer is used to interactively fuse the latent variables with actual values ​​and those with missing values. During this process, to effectively distinguish the information validity mask variables, LSTM variables are used. The sigmoid activation function is used as a filter to preserve... The useful information is extracted and useless information is discarded. The calculation formula for the fusion layer is as follows:

[0082]

[0083]

[0084] In the formula, and Represents the weight parameters. and Indicates the bias parameter. This represents the sigmoid activation function. This represents the hyperbolic tangent activation function.

[0085] Furthermore, considering that the reference value of historical observations for reconstructing current missing data is not constant, and that their influence gradually weakens as the time interval between the last observation and the current missing location increases, this embodiment introduces a time decay gate. The decay gate is determined by the time interval. Derived from this, it is used to hide variables in the output. Perform filtering. The value ranges from 0 to 1. (Hidden variable) The formula for generating it is as follows:

[0086]

[0087]

[0088] In the formula, These are weight parameters. It is a bias parameter.

[0089] Finally, although variables The data already includes filtered and decayed historical patterns and missing data. However, to further improve the accuracy of the estimation, it is not directly used as output. Instead, residual connections are used to retain the original information, resulting in a more accurate estimation result: the time decay estimate. The calculation formula is as follows:

[0090]

[0091] In the formula, For weight parameters, For bias parameters, and , Represents the set of real numbers. This indicates a splicing operation.

[0092] In the above embodiments, firstly, the numerical evolution of traffic flow and the missing data distribution pattern are characterized by independent channels, effectively decoupling the two types of heterogeneous features. Based on this, an intrinsic confidence gating mechanism is constructed by adaptively filtering the actual value latent variables using a gating signal generated by masked latent variables. This mechanism can automatically reduce the weight of low-confidence data features according to the missing data pattern, effectively suppressing noise interference introduced by imputation and ensuring high signal-to-noise ratio and purity.

[0093] Secondly, this embodiment introduces a time decay gate derived from the time delay matrix, thereby constructing a physical time-based forgetting mechanism within the neural network. This mechanism enables the model to explicitly perceive the time span since the last valid observation and dynamically decay historical latent variables accordingly. In scenarios with long-term continuous data gaps, this design effectively reduces the misleading influence of outdated historical information on current estimates, making the imputation results more consistent with the physical laws governing traffic flow evolution over time.

[0094] Finally, this embodiment uses residual connections to generate the final time decay estimate, that is, at the output end, the high-order features filtered by decay are concatenated and fused with the original forward imputation data. This design not only preserves the deterministic information in the original observation data, but also incorporates the nonlinear spatiotemporal variation laws learned by the model. This not only effectively prevents the information degradation problem in deep networks, but also significantly improves the accuracy and robustness of the final imputation result.

[0095] In one embodiment, the step of inputting the pre-filled graph structure data into a graph attention network, aggregating traffic flow information of adjacent vertices based on the graph attention mechanism, and generating a spatial dependency estimate specifically includes:

[0096] After replacing the missing data with the time decay estimate, the graph attention coefficients between vertices in the graph structure are calculated;

[0097] The graph attention coefficient is calculated based on the feature representations of the current vertex and its neighboring vertices;

[0098] The traffic flow information of adjacent vertices is weighted and aggregated according to the graph attention coefficient, and the aggregated spatial information is fused with the time decay estimate through residual connection to obtain the spatial dependency estimate.

[0099] Specifically, before processing the spatial dependencies between road segments in the traffic network, a crucial data preprocessing step is performed: the time decay estimates generated in the previous steps are used to replace missing data in the original traffic flow observation data. This operation effectively uses the predicted information in the time dimension to fill the gaps in the spatial dimension, constructing a pre-filled graph with a complete topological structure, ensuring that the input for subsequent graph computations no longer contains invalid zero or null values. Based on this, the graph attention coefficients between vertices in the graph structure are calculated. For any road segment vertex in the traffic network, the model calculates the correlation strength between them using a shared attention mechanism (usually including learnable weight vectors and the LeakyReLU activation function) based on the feature representation of that vertex and the feature representations of its one-hop neighbors, and normalizes it using the Softmax function to obtain dynamically changing attention coefficients. Subsequently, the model performs weighted aggregation of traffic flow information for all adjacent vertices based on these attention coefficients. This means that neighboring nodes with stronger relevance to the current road segment's traffic state will be assigned greater weights, thereby accurately capturing local spatial dependency features. To further integrate spatiotemporal information and ensure the stability of model training, this embodiment introduces a residual connection structure at the output. The spatial features obtained by aggregating the graph attention layer are fused with the temporal decay estimate from the input (e.g., through an addition operation after linear transformation) to obtain the final spatial dependency estimate. This design ensures that the spatial module is "incrementally corrected" based on the temporal estimate, rather than learning from scratch.

[0100] In this embodiment, high-confidence time decay estimates are used for pre-filling, providing the graph attention network with an input rich in prior information, ensuring the effectiveness and accuracy of the neighborhood aggregation operation. A graph attention mechanism replaces traditional static graph convolution. Considering that the spatial correlation of traffic flow is not static (e.g., the mutual influence between upstream and downstream road segments is drastically different during congestion and free flow), the graph attention mechanism can dynamically adjust the aggregation weights of neighboring nodes based on real-time traffic characteristics, thereby accurately capturing this time-varying, non-linear spatial dependency in the traffic network. Finally, by introducing a residual connection structure, the model can directly reuse the feature extraction results in the time dimension, transforming the localization of spatial modules into spatial calibration and correction of the time estimates. This effectively alleviates the gradient vanishing and degradation problems in deep neural networks and ensures that the final estimation result simultaneously considers the continuity of the time dimension and the correlation of the spatial dimension, improving the overall filling accuracy.

[0101] Furthermore, in one embodiment, the process of determining the spatial dependency estimate aims to fully utilize the spatial topological characteristics of traffic flow in the road network. In a traffic network, the traffic flow of any road segment is not only strongly correlated with its own historical flow evolution, but also has a significant dynamic correlation with the flow status of its neighboring road segments. Based on this characteristic, this embodiment employs a Graph Attention Network (GAT) to mine the dependencies between different sensors from a spatial dimension to assist in inferring missing data. To enable the Graph Attention Network to capture more complete spatial relationships, time decay estimates are used. Alternative The missing data was extracted, while the original observations were retained.

[0102]

[0103] In the formula, .

[0104] Graph Attention Networks (GAT) integrate self-attention mechanisms to learn dynamic spatial relevance. The graph self-attention operation formula is as follows:

[0105]

[0106] In the formula, This represents the sigmoid activation function. The timestamp is vertex The output value, It is the vertex One-hop neighbor (with vertex) (The set of all nodes directly connected by an edge) These are training parameters. Attention coefficients. The calculation formula is as follows:

[0107]

[0108] In the formula, These are training parameters. This represents an activation function.

[0109] The graph attention layer captures spatial information from neighboring vertices, and then fuses this spatial information using a standard LSTM. The LSTM used for processing spatial information is abbreviated as LSTM. Based on the learned spatial dependencies, a fully connected layer is applied, and the spatial dependency estimate is obtained through residual connections as a time estimate. :

[0110]

[0111]

[0112] In the formula, For weight parameters, This is the bias parameter.

[0113] In the above embodiments, by employing a "time-filling-space" preprocessing strategy, a complete graph input is constructed using high-confidence time decay estimates. This effectively solves the problem of spatial feature propagation interruption or distortion caused by zero-value input when facing missing data in traditional graph models, ensuring the effectiveness of neighborhood aggregation. Furthermore, a graph attention network replaces the traditional static graph convolution, dynamically adjusting the aggregation weights of neighboring nodes based on real-time traffic flow characteristics. This means the model can keenly perceive the changes in the intensity of mutual influence between upstream and downstream road segments under different traffic conditions (such as congestion or smooth flow), thereby accurately fitting the nonlinear dynamic spatial dependencies in the traffic network. By introducing spatial LSTM for feature fusion and employing a residual connection structure, not only is the model's ability to integrate spatiotemporal sequence features enhanced, but the residual path also ensures the direct transmission of original temporal prior information, effectively mitigating the gradient degradation problem in deep networks. This results in the final estimation results possessing both temporal continuity and spatial correlation.

[0114] In one embodiment, the step of fusing the time decay estimate and the spatial dependence estimate using a tradeoff coefficient specifically includes:

[0115] Using the mask matrix and the time decay gate as input features, a weighting coefficient with values ​​between 0 and 1 is generated.

[0116] The traffic flow filling result is obtained by weighting and summing the time decay estimate and the spatial dependence estimate using the aforementioned trade-off coefficient.

[0117] Specifically, the mask matrix and the time decay gate generated in the preceding steps are used as key input features. The mask matrix represents the spatial missing data pattern, while the time decay gate quantifies the time span between the current missing location and the last valid observation. The model linearly transforms these two features through a neural network layer containing learnable parameters (such as weight matrices and bias terms), and uses a sigmoid activation function to strictly map the transformed output to the interval between 0 and 1, thus generating a dynamically changing tradeoff coefficient. Subsequently, this tradeoff coefficient is used to perform a weighted summation operation on the time decay estimate and the spatial dependency estimate. This process automatically calculates the confidence ratio of temporal and spatial features in the current completion task based on the missing data state and the length of the time interval, ultimately generating a unidirectional traffic flow completion result.

[0118] In this embodiment, dynamic optimization and fusion of spatiotemporal features are achieved by introducing an adaptive tradeoff coefficient. Unlike traditional fusion methods that use fixed weights or simple averaging, this mechanism gives the model the ability to "self-judge": when data is recently missing (with a large time decay threshold), the model automatically assigns a higher weight to the time decay estimate, prioritizing recent historical information; while when data has been missing for a long time, leading to a decrease in the reliability of temporal features, but neighboring nodes have complete data, the model reduces the proportion of time weights and tends to rely on spatial dependency estimates. This dynamic weighting strategy based on missing patterns and time lag effectively overcomes the limitations of a single feature extraction perspective, ensuring that the model can achieve optimal imputation accuracy in various random or long-term continuous missing scenarios, significantly enhancing the system's robustness and environmental adaptability.

[0119] Furthermore, in one embodiment, the time decay estimate and the spatial dependence estimate are weighted and summed using the tradeoff coefficient to obtain the traffic flow filling result, which can be specifically determined by the following formula.

[0120] Specifically, for the mask vector and time decay coefficient Perform the conversion and learn the trade-off coefficients. .

[0121]

[0122] The formula for calculating the overall filling result is as follows:

[0123]

[0124]

[0125] In the formula, Represents the weight parameters. Indicates the bias parameter. This represents the activation function, used to... Scale to the range (0, 1).

[0126] In one embodiment, the method further includes:

[0127] A loss function is constructed to optimize the temporal feature extraction network and the graph attention network;

[0128] The loss function includes reconstruction loss and consistency-filling loss:

[0129] The reconstruction loss is used to calculate the time decay estimate, the spatial dependency estimate, and the mean absolute error between the traffic flow filling result and the actual observation data, in order to adjust the network parameters of the time feature extraction network and the graph attention network.

[0130] The filling consistency loss is used to calculate the average absolute error between the forward filling result and the backward filling result, in order to constrain the bidirectional feature consistency of the temporal feature extraction network and the graph attention network.

[0131] Specifically, the loss function consists of two parts: the reconstruction loss and the infill consistency loss. The former is the mean absolute error between the estimated values ​​(time decay estimate, spatial dependency estimate, and fusion estimate) and the original observations, and its purpose is to make the infilled values ​​as close as possible to the true values. The latter is the mean absolute error applied to the forward and backward infill results, used to ensure consistency between the infill results in both directions. The formula for calculating the loss function is as follows:

[0132]

[0133] In the formula, It represents the mean absolute error.

[0134] In the above embodiments, by applying reconstruction constraints to the three levels of temporal estimation, spatial estimation, and final fusion result, not only is the control over the accuracy of the final output strengthened, but the intermediate layer features are also forced to have explicit physical semantics, effectively alleviating the gradient vanishing and feature degradation problems in deep neural networks. The consistency loss for filling in the gaps utilizes the mutual corroboration of forward and backward inference to provide the model with additional regularization constraints. This is particularly crucial in data-sparse scenarios, effectively preventing logical discrepancies in bidirectional prediction results and significantly enhancing the model's robustness in long-term time-series filling tasks.

[0135] Furthermore, in one embodiment, the final filling result can be determined by the following formula:

[0136] The above embodiments are obtained Forward filling results and backfill results Using historical observations A fusion method replaces the mean method, integrating the two results. The fusion ratio of the forward and backward imputation results is related to the amount of effective information at each location. The percentage of the forward (fill-in) result is represented by the following formula:

[0137]

[0138] The final filling result is represented by the following formula:

[0139]

[0140] Based on the final completion results, complete traffic situation data including the completed flow, speed or density indicators is output, thus achieving high-precision estimation of traffic flow in discontinuous sensing areas.

[0141] In one embodiment, the method further includes inferring traffic situation indicators for road sections without sensing devices, specifically including:

[0142] The upstream road segment with sensing data adjacent to the road segment without sensing equipment is defined as the starting road segment, and the downstream road segment with sensing data adjacent to the road segment without sensing equipment is defined as the ending road segment.

[0143] The measured traffic situation indicators of the starting road segment and the ending road segment are obtained by sensors, and the measured traffic situation indicators include flow rate, speed and density;

[0144] Based on the measured traffic situation indicators of the starting and ending road segments, the traffic situation indicators of the road segments without sensing devices are calculated using a traffic flow model.

[0145] Specifically, such as Figure 2 For road segments with sensing devices but partially missing data, the data-driven traffic situation indicator continuity method described in the above embodiments is used to address the data discontinuity issue. For road segments without sensing devices and completely missing data, a model-driven traffic situation indicator continuity method is used to infer the traffic situation indicators for the data-free road segments, thus addressing the data discontinuity issue. An upstream road segment with sensing data adjacent to a data-free road segment is defined as the starting point of the data-free road segment. The first downstream road segment with sensing data adjacent to the road segment without sensing devices is defined as the endpoint of the road segment without sensing devices. Traffic flow, speed, density, and other traffic situation indicators at the starting and ending points are measured data obtained through sensing devices.

[0146] Indicates the starting point of the road segment In the Traffic flow during a given time period (unit: veh / h) Indicates the end point of the road segment In the Traffic flow during a given time period (unit: veh / h).

[0147] Indicates the starting point of the road segment In the Density at any given time (unit: veh / km / lane). Indicates the end point of the road segment In the Density at any given time (unit: veh / km / lane).

[0148] Indicates the starting point of the road segment In the Average speed at any given time (unit: km / h) Indicates the end point of the road segment In the Average speed at any given time (unit: km / h).

[0149] The formula for calculating the density of road sections without sensing devices is as follows:

[0150]

[0151] In the formula, Indicates road sections without sensing devices Each lane in the first Density at any given time (unit: veh / km / lane). Indicates a time scale. Indicates the data update cycle. Indicates road segment Length (unit: kilometers). Indicates road segment The number of lanes (unit: lanes). Indicates road sections without sensing devices In the Traffic flow during a given time period (unit: veh / h).

[0152] when hour, , Indicates road segment The traffic flow of the upstream adjacent starting segment. That is, when hour, , It is measured data obtained through sensing devices.

[0153] The formula for calculating the average speed on road sections without sensing devices is as follows:

[0154]

[0155] In the formula, the first term This indicates the velocity at the previous moment. The second term... Expressing vehicle speed in time The delay approximates the traffic flow speed under steady-state density , This represents the relaxation time parameter. Indicates road segment In the The velocity calculated based on density at all times. (Third item) This is the convective term, indicating that the average speed is related to the speed of the upstream adjacent road segment; that is, the upstream traffic flow tends to enter the next segment at the same speed. The fourth term... This is the expected value, indicating that the driver will adjust the vehicle speed based on the density ahead. Indicates the expectation index. This represents a constant parameter.

[0156] The calculation formula is as follows:

[0157]

[0158] In the formula, Indicates road segment Free-flow velocity (unit: km / h). This represents the natural exponential function. It is a constant parameter. Indicates road segment Critical density (unit: veh / km / lane).

[0159] when hour, , Indicates road segment The average speed of the upstream adjacent starting segment. That is, when hour, , It is the starting section of the road. In the The average speed at any given time is measured data obtained through sensing devices.

[0160] use This represents the total number of road segments without sensing devices in the analysis area. hour, , This indicates the first downstream road segment with sensing data adjacent to the road segment without sensing devices, i.e., the endpoint of the road segment without sensing devices. Therefore, , It is measured data obtained through sensing devices.

[0161] The formula for calculating traffic flow in road sections without sensing devices is as follows:

[0162]

[0163] In the formula, Indicates road sections without sensing devices In the Traffic flow during a given time period (unit: veh / h). Indicates road sections without sensing devices In the Average speed at any given time (unit: km / h). Indicates road segment The number of lanes (unit: lanes).

[0164] In the above embodiments, a macroscopic traffic flow extrapolation mechanism based on boundary constraints was constructed for road sections with complete blind spots caused by sparse deployment of roadside sensing devices. This effectively overcomes the failure problem of traditional data-driven methods in scenarios without historical observation data. This mechanism establishes sensing road sections upstream and downstream of the blind spot as physical boundary conditions, and combines the traffic flow conservation principle and the dynamic evolution mechanism of speed to perform reverse physical extrapolation of the traffic situation in the intermediate road section. This method deeply quantifies the relaxation effect of traffic flow regressing to steady-state density, the convection effect influenced by upstream speed, and the expectation effect constrained by downstream density. This ensures that the calculated flow, speed, and density indicators not only fit numerically but also strictly follow the essence of traffic flow operation in terms of physical mechanisms. It achieves logical self-consistency and continuity of traffic situation indicators across the entire spatial distribution, filling the spatial gap in regional traffic sensing.

[0165] Furthermore, some embodiments of this application also provide an electronic device. The electronic device can be various forms of digital computer, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, etc. The electronic device can also be various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices.

[0166] The electronic device includes: one or more processors; and a memory storing computer program instructions that, when executed, cause the processor to perform the steps of the methods provided in any one or more of the above embodiments. Figure 3 An exemplary structural diagram of the electronic device is disclosed. The electronic device includes one or more processors 1101, a memory 1102, and interfaces for connecting the various components, including high-speed interfaces and low-speed interfaces. The various components are interconnected via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the electronic device, including instructions stored in or on memory to display graphical information of a GUI on an external input / output device (such as a display device coupled to the interface). In some other embodiments, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple electronic devices can be connected, each providing some of the necessary operations. The components, their connections and relationships, and their functions shown herein are merely examples and are not intended to limit the implementation of the present application described and / or claimed herein.

[0167] The electronic device may further include an input device 1103 and an output device 1104. The processor 1101, memory 1102, input device 1103 and output device 1104 may be connected by a bus or other means, as shown in the figure, which is connected by a bus.

[0168] Input device 1103 can receive input numerical or character information, and generate key signal inputs related to user settings and function control of the electronic device, such as a touch screen, keypad, mouse, trackpad, touchpad, joystick, one or more mouse buttons, trackball, joystick, etc. Output device 1104 may include a display device, auxiliary lighting device (e.g., LED), and haptic feedback device (e.g., vibration motor). The display device may include, but is not limited to, a liquid crystal display, a light-emitting diode display, and a plasma display. In some embodiments, the display device may be a touch screen.

[0169] To provide interaction with the user, the electronic device can be a computer. The computer has: a display device (e.g., a cathode ray tube or LCD monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback); and input from the user can be received in any form (e.g., voice input or tactile input).

[0170] In this embodiment, a computer-readable medium stores a computer program / instructions that, when executed by a processor, implement the steps of the methods provided in any one or more of the above embodiments. This computer-readable medium may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into that device. The aforementioned computer-readable medium carries one or more computer-readable instructions.

[0171] The memory 1102 can serve as a non-transitory computer-readable storage medium, used to store non-transitory software programs, non-transitory computer-executable programs, and modules. The processor 1101 executes various functional applications and data processing of the server by running the non-transitory software programs, instructions, and modules stored in the memory 1102, thereby implementing the program instructions / modules corresponding to the methods provided in any one or more of the embodiments described above in this application.

[0172] The memory 1102 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the electronic device. Furthermore, the memory 1102 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory 1102 may optionally include memory remotely located relative to the processor 1101, and these remote memories can be connected to the electronic device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0173] It should be noted that the computer-readable medium described in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. Computer-readable media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, electrical connections having one or more wires, portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, optical fibers, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0174] Computer-readable media include permanent and non-permanent, removable and non-removable media, which can store information by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory, static random access memory, dynamic random access memory, other types of random access memory, read-only memory, electrically erasable programmable read-only memory, flash memory or other memory technologies, read-only optical discs, digital versatile optical discs or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0175] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including local area networks (LANs) or wide area networks (WANs), or it can be connected to an external computer (e.g., via the Internet through an Internet service provider).

[0176] In the above embodiments, all or part of the implementation can be achieved through software, hardware, firmware, or any combination thereof. For example, it can be implemented using an application-specific integrated circuit (ASIC), a general-purpose computer, or any other similar hardware device. In some embodiments, the software program of this application can be executed by a processor to implement the above steps or functions. Similarly, the software program of this application (including related data structures) can be stored in a computer-readable recording medium, such as RAM memory, magnetic or optical drives, floppy disks, and similar devices. In addition, some steps or functions of this application can be implemented in hardware, for example, as circuitry that cooperates with a processor to perform the various steps or functions.

[0177] The computer program product provided in this application includes one or more computer programs / instructions. When executed by a processor, these computer programs / instructions generate, in whole or in part, the processes or functions described in this application. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0178] The flowcharts or block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of devices, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-specific system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0179] The scope of this application is defined by the appended claims rather than the foregoing description, and is therefore intended to encompass all variations falling within the meaning and scope of equivalents of the claims. No reference numerals in the claims should be construed as limiting the scope of the claims. Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a device claim may also be implemented by a single unit or device in software or hardware. Terms such as "first," "second," etc., are used only for distinguishing descriptions and do not indicate any particular order, nor should they be construed as indicating or implying relative importance.

[0180] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily made by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims, and the above embodiments should be regarded as exemplary and non-limiting.

Claims

1. A method for estimating traffic flow in discontinuous sensing areas, characterized in that, Includes the following steps: Obtain traffic flow observation data collected by sensors in the traffic network of the target area, define the traffic network as a graph structure, where road segments are vertices and the spatial correlation between road segments are edges, and construct a mask matrix indicating the location of missing values ​​in the traffic flow observation data; Based on the timestamps of the traffic flow observation data, a time lag matrix is ​​calculated between the observation time points. The time lag matrix represents the time interval between the current missing data and the most recent valid observation data. The traffic flow observation data and the mask matrix are input into a time feature extraction network, and latent variables are obtained by processing them using a long short-term memory network. The latent variables are filtered using a time decay gate generated based on the time delay matrix, thereby generating a time decay estimate that includes time dimension features; The time decay estimate is used to fill in the missing locations in the traffic flow observation data to construct pre-filled graph structure data; The pre-filled graph structure data is input into a graph attention network, and traffic flow information of adjacent vertices is aggregated based on the graph attention mechanism to generate spatial dependency estimates. The time decay estimate and the spatial dependence estimate are fused using a tradeoff coefficient to obtain traffic flow completion data, and complete traffic situation data including the completed flow, speed or density indicators are output.

2. The method according to claim 1, characterized in that, The step of calculating the time lag matrix between observation time points based on the timestamps of the traffic flow observation data specifically includes: Calculate the forward time delay matrix and the reverse time delay matrix respectively; The elements in the positive time delay matrix represent the time difference from the most recent historical observation point to the current missing time point; The elements in the inverse time delay matrix represent the time difference from the most recent future observation point to the current missing time point; The time feature extraction network adopts a bidirectional structure, performing forward imputation and backward imputation based on the forward time delay matrix and the reverse time delay matrix respectively, and generating forward imputation results and backward imputation results.

3. The method according to claim 1, characterized in that, The step of inputting the traffic flow observation data and the mask matrix into a time feature extraction network, and processing them using a long short-term memory network to obtain latent variables, specifically includes: Construct a dual-channel long short-term memory network, including an actual value channel and a mask channel; The traffic flow observation data is processed using the actual value channel to obtain the actual value latent variables; The mask matrix is ​​processed using the mask channel to obtain the mask hidden variables; The actual value latent variable and the mask latent variable are fused through a fusion layer. The mask latent variable is processed by an activation function and then used as a filter to perform feature filtering on the actual value latent variable to obtain the latent variable.

4. The method according to claim 3, characterized in that, The step of filtering the latent variables using a time decay gate generated based on the time delay matrix to generate a time decay estimate containing time dimension features specifically includes: Based on the time interval values ​​in the time delay matrix, a time decay gate with a value between 0 and 1 is calculated. The time decay gate is used to multiply the output of the fusion layer element by element to obtain the filtered time dimension features.

5. The method according to claim 4, characterized in that, The step of inputting the pre-filled graph structure data into a graph attention network, aggregating traffic flow information of adjacent vertices based on the graph attention mechanism, and generating spatial dependency estimates specifically includes: After replacing the missing data with the time decay estimate, the graph attention coefficients between vertices in the graph structure are calculated; The graph attention coefficient is calculated based on the feature representations of the current vertex and its neighboring vertices; The traffic flow information of adjacent vertices is weighted and aggregated according to the graph attention coefficient, and the aggregated spatial information is fused with the time decay estimate through residual connection to obtain the spatial dependency estimate.

6. The method according to claim 4, characterized in that, The step of fusing the time decay estimate and the spatial dependence estimate using a tradeoff coefficient specifically includes: Using the mask matrix and the time decay gate as input features, a weighting coefficient with values ​​between 0 and 1 is generated. The traffic flow filling result is obtained by weighting and summing the time decay estimate and the spatial dependence estimate using the aforementioned trade-off coefficient.

7. The method according to claim 2, characterized in that, The method further includes: A loss function is constructed to optimize the temporal feature extraction network and the graph attention network; The loss function includes reconstruction loss and consistency-filling loss: The reconstruction loss is used to calculate the time decay estimate, the spatial dependency estimate, and the mean absolute error between the traffic flow filling result and the actual observation data, in order to adjust the network parameters of the time feature extraction network and the graph attention network. The filling consistency loss is used to calculate the average absolute error between the forward filling result and the backward filling result, in order to constrain the bidirectional feature consistency of the temporal feature extraction network and the graph attention network.

8. The method according to claim 1, characterized in that, The method also includes inferring traffic situation indicators for road sections without sensing devices, specifically including: The upstream road segment with sensing data adjacent to the road segment without sensing equipment is defined as the starting road segment, and the downstream road segment with sensing data adjacent to the road segment without sensing equipment is defined as the ending road segment. The measured traffic situation indicators of the starting road segment and the ending road segment are obtained by sensors, and the measured traffic situation indicators include flow rate, speed and density; Based on the measured traffic situation indicators of the starting and ending road segments, the traffic situation indicators of the road segments without sensing devices are calculated using a traffic flow model.

9. An electronic device, characterized in that, The electronic device includes: One or more processors; and A memory storing computer program instructions, which, when executed, cause the processor to perform the steps of the method as described in any one of claims 1 to 8.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 8.

Images