Highway running state recognition and prediction method and system based on ETC gantry data

Abstract

The application discloses a highway operation state recognition and prediction method and system based on ETC gantry data, and belongs to the technical field of highway operation state recognition and prediction, and the method specifically comprises the following steps: collecting highway ETC gantry data in real time, performing multi-scale fractal compression, aligning information of the highway ETC gantry data after the multi-scale fractal compression, constructing a four-element traffic state tensor, performing graph-time sequence coding on the highway and the four-element traffic state tensor, recognizing the highway operation state based on the graph-time sequence coding result, and predicting the highway operation state based on the highway operation state recognition result; and the application can more comprehensively depict traffic flow dynamics and evolution rules, and improves the predication ability of congestion nodes and the response efficiency of road network level operation scheduling.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of expressway operation state recognition and prediction, in particular to an expressway operation state recognition and prediction method and system based on ETC gantry data. BACKGROUND

[0002] Real-time recognition and accurate prediction of expressway operation state are important foundations for supporting dynamic traffic scheduling and intelligent management and control. ETC gantry systems, as the core facilities of networked non-stop toll collection, have gradually evolved into wide coverage, high frequency, and highly structured vehicle-road data collection infrastructure. ETC gantry systems can collect core data such as vehicle identification, travel time, travel path segment, and vehicle type information every time a vehicle passes, and through coordination with toll station entrance and exit systems, license plate recognition systems, and video monitoring systems, form a multi-source heterogeneous traffic big data system.

[0003] Existing methods for traffic operation analysis using ETC data mostly focus on static index evaluation, traffic volume statistics, or post-event situation analysis, making it difficult to support dynamic modeling of traffic state and real-time prediction of future trends. In addition, existing researches mostly use traditional time series methods or shallow neural networks in model construction, but it is difficult to effectively model complex road network structures and nonlinear state transition processes. There is a lack of multi-source data fusion processing capability. SUMMARY

[0004] In view of the deficiencies of the prior art, the application provides an expressway operation state recognition and prediction method and system based on ETC gantry data.

[0005] To achieve the above purpose, the application provides the following technical solutions:

[0006] The expressway operation state recognition and prediction method based on ETC gantry data comprises:

[0007] Real-time collection of expressway ETC gantry data, and multi-scale fractal compression of the collected expressway ETC gantry data;

[0008] Information alignment of the multi-scale fractal compressed expressway ETC gantry data, construction of a four-element traffic state tensor, the four-element traffic state tensor comprising a vehicle speed vector, a lane occupancy rate, a vehicle type, and ETC gantry information;

[0009] Graph-time sequence coding of the expressway and the four-element traffic state tensor, and recognition of the expressway operation state based on the graph-time sequence coding result;

[0010] Prediction of the expressway operation state based on the expressway operation state recognition result.

[0011] Specifically, the highway ETC gantry data compressed by the multi-scale fractal is aligned in information, and a four-element traffic state tensor is constructed, including:

[0012] The highway ETC gantry data compressed by the multi-scale fractal is divided into a first fragment, a second fragment, a third fragment and a fourth fragment according to the data source label;

[0013] A four-dimensional coordinate mapping model is used to uniformly encode the gantry latitude and longitude, lane number, longitudinal mileage and timestamp of any fragment;

[0014] Fill the curve to generate an adaptive space-time grid in the uniform coding domain, and assign a grid index to each grid cell;

[0015] Based on the grid index, the first fragment, the second fragment, the third fragment and the fourth fragment, the ETC gantry information, the vehicle speed vector, the vehicle type and the lane occupancy rate are identified or calculated;

[0016] The ETC gantry information, the vehicle speed vector, the vehicle type and the lane occupancy rate are constructed into a four-element traffic state tensor.

[0017] Specifically, the highway and the four-element traffic state tensor are graph-time sequence encoded, and the highway running state is identified based on the graph-time sequence encoding result, including:

[0018] Map the four-element traffic state tensor to a multi-scale graph containing ETC gantry nodes, construct directed edges and initialize weights according to the historical traffic and trajectory similarity between nodes;

[0019] Based on the multi-scale graph transmission attention message, the node state tensor is aggregated to generate a graph context embedding sequence;

[0020] The graph context embedding sequence is input into a bidirectional time sequence encoder to extract time state features;

[0021] Under the condition of no external label, the time state features are fuzzy clustered and self-supervised contrast learned to generate an initial highway running state of the ETC gantry node;

[0022] The initial highway running state is propagated for consistency using the adjacency relationship in the multi-scale graph structure to obtain the highway running state.

[0023] Specifically, the multi-scale graph transmission attention message is used to aggregate the node state tensor to generate a graph context embedding sequence, including:

[0024] The four-element traffic state tensor corresponding to the target node and its adjacent nodes is extracted from the multi-scale graph node relationship, and an attention query index table between node pairs is constructed;

[0025] For each target node, the four-dimensional traffic state tensor of its adjacent nodes is weighted and scored using multi-channel attention, and arranged in the order of spatial position in the attention query index table;

[0026] The weighted score result is matrix aggregated with the four-dimensional traffic state tensor of the corresponding node to generate the first round of context representation of the target node;

[0027] The matrix aggregation process is performed for all target nodes for n rounds, and the upstream and downstream weight distribution is adjusted using the ETC gantry timestamp offset factor in each round;

[0028] After all rounds are completed, the intermediate representation of each node is normalized and mapped to output the graph context embedding sequence.

[0029] Specifically, the graph context embedding sequence is input into a bidirectional time series encoder to extract time state features, including:

[0030] The graph context embedding sequence is rearranged in time sequence according to the ETC gantry node representation, and a historical state sequence set corresponding to each ETC gantry node is constructed;

[0031] For each historical state sequence, set the forward and backward time window range, and perform time alignment processing of state embedding in the forward and backward directions respectively;

[0032] Recursively input the embedding vector of the current time step and the previous state context in the forward sequence, and perform the same operation in the reverse direction in the backward sequence, and record the evolution trajectory sequence in the two directions respectively;

[0033] The outputs of the forward sequence and the backward sequence are spliced and projected into a unified state vector space to obtain the time dependence feature of each ETC gantry node at the current time;

[0034] The time dependence features of all ETC gantry nodes are collected in topological order to obtain the time state feature.

[0035] Specifically, the time state feature is fuzzy clustered and self-supervised contrast learned under the condition of no external label to generate the initial highway running state of the ETC gantry node, including:

[0036] The time state feature is divided into non-overlapping training subsets, each training subset containing the time dependence features of the ETC gantry nodes in different time periods;

[0037] For each training subset, the initialized fuzzy clustering center is used to estimate the membership degree to obtain the membership probability distribution of each ETC gantry node time state feature to different clusters;

[0038] Under the condition of no external label, a positive and negative sample pair is constructed, ETC gantry nodes belonging to the same time window are taken as positive samples, and ETC gantry nodes belonging to different time periods or different road areas are taken as negative samples;

[0039] A plurality of rounds of self-supervised contrast coding are performed using the positive and negative sample pair, a cluster center vector of fuzzy clustering is updated, and the ETC gantry node is adjusted in the embedding space towards a direction of consistency enhancement;

[0040] According to the category index to which the maximum membership degree of the ETC gantry node belongs, an initial expressway running state is determined.

[0041] Specifically, based on the expressway running state recognition result, the expressway running state is predicted, including:

[0042] Based on the time-dependent feature of the ETC gantry node and the corresponding expressway running state, a state-feature pair is constructed, and a prediction sample set is generated;

[0043] The prediction sample set is divided according to a preset time window, and a multi-period input sample sequence and an output sample sequence are constructed, wherein the input sample sequence is used to capture the historical state evolution trajectory, and the output sample sequence represents the target prediction period;

[0044] The input sample sequence is input into a time sequence strategy decision network, the state transition path is iterated, and the control variable of state evolution is calculated, to obtain a historical state vector and a control variable;

[0045] Based on the historical state vector and the control variable, a traffic state prediction result of a future plurality of time steps is generated;

[0046] The traffic state prediction results of all ETC gantry nodes are spatiotemporally reorganized to obtain a running prediction state of the expressway.

[0047] Specifically, the expressway ETC gantry data is collected in real time, and the collected expressway ETC gantry data is compressed by multi-scale fractals, including:

[0048] A pseudo-random hash function is used to generate a unique spatiotemporal index code for each piece of expressway ETC gantry data;

[0049] Based on the spatiotemporal index code, the same-period expressway ETC gantry data segments are subjected to secondary fractal segmentation, and the corresponding fractal dimension spectrum is calculated;

[0050] The fractal dimension spectrum is mapped to an orthogonal basis to construct a quantum fractal measure matrix;

[0051] Based on the preset topological entropy threshold, the quantum fractal measure matrix is pruned layer by layer to obtain a compressed data sequence.

[0052] The preset Markov correlation length is used for sparse recombination of the compressed data sequence, and multi-scale fractal compressed highway ETC gantry data is obtained.

[0053] The highway operation state recognition and prediction system based on ETC gantry data is used to realize the highway operation state recognition and prediction method based on ETC gantry data, and includes a data processing module, a state tensor construction module, an operation state recognition module and an operation state prediction module.

[0054] The data processing module is used for real-time acquisition of highway ETC gantry data, and multi-scale fractal compression of the acquired highway ETC gantry data.

[0055] The state tensor construction module is used for information alignment of the multi-scale fractal compressed highway ETC gantry data, and constructs a four-element traffic state tensor.

[0056] The operation state recognition module is used for graph-time sequence coding of the highway and the four-element traffic state tensor, and recognizes the highway operation state based on the graph-time sequence coding result.

[0057] The operation state prediction module is used for prediction of the highway operation state based on the highway operation state recognition result.

[0058] Specifically, the operation state recognition module includes a graph-time sequence coding unit and an operation state recognition unit.

[0059] The graph-time sequence coding unit is used for mapping the four-element traffic state tensor to a multi-scale graph containing ETC gantry nodes, generating a graph context embedding sequence, and extracting time state features by using a bidirectional time sequence encoder.

[0060] The operation state recognition unit is used for fuzzy clustering and self-supervised contrast learning of the time state features, and uses the adjacent relationship in the multi-scale graph structure to perform consistency propagation on the initial highway operation state, and obtains the highway operation state.

[0061] Compared with the prior art, the beneficial effects of the present application are:

[0062] The application provides a highway operation state recognition and prediction method and system based on ETC gantry data, wherein the highway ETC gantry data is compressed, a four-element traffic state tensor is constructed, then graph attention transmission and bidirectional time sequence coding are performed in a multi-scale graph structure, fuzzy clustering and self-supervised contrast learning are combined to recognize the operation state under a label-free condition, and finally multi-period state prediction is performed; compared with a traditional prediction method based on a single index or artificial rules, the method can more comprehensively depict traffic flow dynamics and evolution rules, improves the congestion node prediction ability and response efficiency of road network operation scheduling, and is suitable for various application requirements such as operation monitoring, congestion treatment and traffic diversion in a highway scene. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 A highway operation state recognition and prediction method based on ETC gantry data is provided in the application;

[0064] Figure 2 A highway operation state recognition method is provided in the application;

[0065] Figure 3 A highway operation state prediction method is provided in the application;

[0066] Figure 4 A highway operation state recognition and prediction system architecture based on ETC gantry data is provided in the application. DETAILED DESCRIPTION

[0067] The application will be described in detail below with specific embodiments. The following embodiments will help those skilled in the art to further understand the application, but do not limit the application in any form. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the application. These all belong to the protection scope of the application.

[0068] In order to make the purpose, technical scheme and advantages of the application clearer and more apparent, the application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application.

[0069] It should be noted that the various features of the embodiments of the present application can be combined with each other, and are within the protection scope of the present application, if there is no conflict. In addition, although the functional modules are divided in the device schematic diagram, and the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order from the module division in the device or the order in the flowchart. In addition, the "first", "second", "third" and the like used in the present application do not limit the data and execution order, but only distinguish the same items or similar items with basically the same function and effect.

[0070] Unless otherwise defined, all technical and scientific terms used in the present application have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used in the present application are only for the purpose of describing the specific embodiments of the present application, and are not used to limit the present application. The term "and / or" used in the present application includes any and all combinations of one or more related listed items.

[0071] Embodiment 1:

[0072] Please refer to Figure 1 The present application provides an embodiment: a highway running state recognition and prediction method based on ETC gantry data, comprising the following specific steps:

[0073] Step S1: real-time collection of highway ETC gantry data, and multi-scale fractal compression of the collected highway ETC gantry data.

[0074] Specifically, there are three types of data generated by the ETC gantry: first, the gantry transaction data, the data generated by the interaction between the CPC card or OBU tag and the ETC gantry when the vehicle passes through the ETC gantry, including vehicle entry information, passing time, vehicle type, gantry billing mileage, billing amount and other information; second, the picture data captured by the gantry camera, the real-time vehicle picture captured by the camera installed on the ETC gantry when the vehicle passes through the gantry; third, the license plate recognition stream data, which is an information description of the gantry captured picture, including gantry number, capture time, recognized license plate number, recognized vehicle type, shooting position and the like.

[0075] The specific steps of step S1 are:

[0076] Step S101: using a pseudo-random hash function to generate a unique space-time index code for each piece of highway ETC gantry data.

[0077] In the embodiment, a pseudo-random hash function with input disturbance sensitivity and output uniform distribution characteristics is selected, and the input is jointly modeled: after standardizing the vehicle identification, the time stamp field is normalized in precision, and then spliced with the gantry number and geographic coordinates into a high-dimensional input vector; after performing hash calculation, the output is a string of fixed length, discrete uniform coding space, and extremely low reversibility binary hash value, which is encoded as the unique spatiotemporal index code of the data.

[0078] It should be noted that the spatiotemporal index code not only bears the identification function, but also participates in the tensor key construction as one of the variables in the structured tensor construction, ensuring consistency and traceability in the spatiotemporal dimension. Through the above hash generation mechanism, the problems of data redundancy, field heterogeneity, and uneven distribution are effectively overcome.

[0079] Step S102: Based on the spatiotemporal index code, the same period highway ETC gantry data segments are subjected to secondary fractal segmentation, and the corresponding fractal dimension spectrum is calculated.

[0080] In the embodiment, all ETC gantry data segments located in the same time period are divided into a unified data group based on the unified spatiotemporal index code as the key index key. Then, the data group is subjected to fractal segmentation process. First, a primary subset is constructed according to the vehicle passing path trajectory, and then a secondary division is performed on each subset based on the flow density fluctuation and path density to form a multi-level spatial nested structure.

[0081] Further, within each divided subset, the fractal dimension index of the subset at a specific segmentation scale is derived by calculating the logarithmic ratio between its spatial coverage structure and frequency change. The estimation process of fractal dimension introduces multiple overlapping windows and moving sample points. By accumulating the measurement results of multiple scale layers, the fractal dimension spectrum of the data segment is finally formed, i.e., the complexity change spectrum at different time scales.

[0082] Step S103: Map the fractal dimension spectrum to the orthogonal basis to construct a quantum fractal measure matrix.

[0083] In the embodiment, the fractal dimension spectrum is regarded as a sequence of function values changing in the continuous scale domain. First, normalization operation is performed to make it evenly distributed in the standard value domain. Then, the inner product expansion is performed between the scale level and the corresponding sequence in the selected orthogonal basis in turn, so that each dimension component of the fractal dimension spectrum is projected into the orthogonal basis space to obtain the mapping coefficient set of the fractal dimension spectrum in the orthogonal basis space. The coefficient set is filled into the matrix structure as weights, and the row index represents the mapping dimension, and the column index corresponds to the original fractal spectrum dimension.

[0084] Further, to strengthen the compressibility and computational compatibility of the matrix, a threshold selection mechanism based on spectral density reconstruction error is introduced, and structural pruning is implemented for low contribution terms, and window smoothing is performed on high-frequency perturbation coefficients to reduce the influence of characteristic dispersion on the stability of the subsequent model. After the above processing, the quantum fractal measure matrix is obtained.

[0085] Step S104: based on a preset topological entropy threshold, pruning operation is performed on the quantum fractal measure matrix layer by layer to obtain a compressed data sequence.

[0086] In this embodiment, when performing the step, first, the quantum fractal measure matrix is divided into layers along the scale dimension direction, each layer corresponds to a fixed scale window, and in each layer, the topological entropy value of the layer is obtained by calculating the information gain distribution per unit area or per node; then the topological entropy value is compared with the topological entropy threshold set in the system, if it is lower than the topological entropy threshold, it is determined that the layer is a weak contribution layer, and enters the pruning set.

[0087] Further, the matrix sub-blocks entering the pruning set are not directly deleted, but are redirected to zero vectors using a structural sparse mapping strategy, the position relationship in the tensor position is retained, but the actual numerical content is erased; finally, after performing topological entropy evaluation and pruning marking of all layers, the residual part of the matrix structure constitutes the compressed data sequence.

[0088] Step S105: the compressed data sequence is sparsely reorganized according to a preset Markov correlation length to obtain a multi-scale fractal compressed highway ETC gantry data.

[0089] Step S2: aligning information of the multi-scale fractal compressed highway ETC gantry data, constructing a four-element traffic state tensor, the four-element traffic state tensor including a vehicle speed vector, a lane occupancy rate, a vehicle type, and ETC gantry information.

[0090] The specific steps of step S2 are as follows:

[0091] Step S201: dividing the multi-scale fractal compressed highway ETC gantry data into a first slice, a second slice, a third slice, and a fourth slice according to data source labels.

[0092] Specifically, the first slice, the second slice, the third slice, and the fourth slice are ETC gantry system slices, toll station entrance and exit slices, license plate recognition slices, and video monitoring slices, respectively.

[0093] Step S202: using a four-dimensional coordinate mapping model to uniformly encode the gantry longitude and latitude, lane number, longitudinal mileage, and timestamp of any slice.

[0094] In the embodiment, four types of indicators are spliced into a four-tuple coordinate (X, Y, L, T) in a fixed order, where X and Y are the conversion results of longitude and latitude, L is the longitudinal mileage, and T is the standardized timestamp.

[0095] Step S203: Fill the adaptive spatio-temporal grid with the curve and assign a grid index to each grid cell in the unified coding domain.

[0096] In the embodiment, a four-dimensional coordinate space is divided into basic grid cells, each dimension uses different cutting scales: longitude and latitude use fixed spatial resolution, lane number is mapped with integer number, longitudinal mileage uses milepost interval constraint, and time dimension dynamically adjusts sampling frequency according to traffic density; then all grid cells are numbered and sorted to form a continuous spatial traversal sequence; in the sequence, each four-dimensional grid cell is assigned a unique grid index according to the curve traversal order.

[0097] Step S204: Based on the grid index, the first, second, third and fourth fragments, the ETC gantry information, the vehicle speed vector, the vehicle type and the lane occupancy rate are identified or calculated.

[0098] Step S205: The ETC gantry information, the vehicle speed vector, the vehicle type and the lane occupancy rate are constructed into a four-dimensional traffic state tensor.

[0099] In the embodiment, the construction of the four-dimensional traffic state tensor here also includes information such as space cell number and timestamp.

[0100] Step S3: Graph-temporal coding is performed on the expressway and the four-dimensional traffic state tensor, and the running state of the expressway is identified based on the graph-temporal coding result.

[0101] As shown in Figure 2 , the specific steps of step S3 are:

[0102] Step S301: Map the four-dimensional traffic state tensor to a multi-scale graph containing ETC gantry nodes, and construct directed edges and initialize weights according to the historical traffic and trajectory similarity between nodes.

[0103] In this embodiment, each gantry entity in the four-dimensional traffic state tensor is extracted as a node in the graph, which retains all state vectors in the corresponding four-dimensional traffic state tensor as node attributes, and the time dimension in the four-dimensional traffic state tensor is expanded by a sliding window to form multiple time segments, and the state of the gantry node in each time segment is regarded as an independent state snapshot to constitute a static graph of the time layer; the vehicle passing trajectory sequence within a certain time range is extracted, and the co-occurrence frequency and order offset of vehicles between any two gantry nodes are identified; the traffic state vectors of the two gantry nodes in adjacent time segments are normalized to calculate the state similarity score, which is combined with the passing trajectory weight as the final edge initialization weight input to the graph structure.

[0104] Step S302: based on the multi-scale graph passing attention message, the node state tensor is aggregated to generate a graph context embedding sequence.

[0105] The specific steps of step S302 are:

[0106] Step S3021: extract the four-dimensional traffic state tensor corresponding to the target node and its adjacent nodes from the multi-scale graph node relationship, and construct an attention query index table between the node pairs.

[0107] Step S3022: for each target node, use a multi-channel attention to weight and score the four-dimensional traffic state tensor of its adjacent nodes, and arrange them in the order of spatial position in the attention query index table.

[0108] In this embodiment, first, the attention query index table is called to determine the set of all adjacent nodes connected to the target node, and the corresponding tensor encoding is obtained; then the state vector of the target node at the current time is set as the query vector, and the matching degree between the query vector and each channel sub-vector of each adjacent node is calculated; a channel fusion strategy is introduced to normalize and fuse the channel attention scores, and a learnable fusion coefficient is used to weight and sum to form a comprehensive attention score; after completing the attention weight scoring, the adjacent nodes are reordered according to the preset spatial position order in the attention query index table.

[0109] Step S3023: matrix aggregation processing is performed on the weighted scoring result and the four-dimensional traffic state tensor of the corresponding node to generate the first round of context representation of the target node.

[0110] In this embodiment, first, the adjacent state tensor matrix is established, which takes all adjacent nodes of the target node as a row vector set, and each row corresponds to the four-dimensional traffic state tensor of the adjacent node in the current time window. The channel alignment operation is performed on the adjacent node tensor matrix; then the element-wise sum operation is performed on the matrix after weighting all channels according to the channel dimension, and the aggregation result is standardized to a uniform dimension vector to form the first round of context representation of the target node.

[0111] Step S3024: Perform the matrix aggregation process for all target nodes for n rounds, and adjust the upstream and downstream weight distribution with the ETC gantry timestamp offset factor in each round.

[0112] In this embodiment, first set the number of processing rounds n, which is usually determined by the complexity of the road topology and the required state coverage; in each round of propagation processing, attention scoring and matrix aggregation operations are performed, but in this step, a timestamp offset factor is specially introduced to adjust the weighted weight distribution of each adjacent node.

[0113] Further, after each round of propagation processing, the current target node's four-element traffic state tensor is weighted and fused with the last round state representation, and a residual connection structure is introduced to retain the information difference of each iteration.

[0114] Step S3025: After all rounds are completed, normalize the intermediate representation of each node and output the graph context embedding sequence.

[0115] Step S303: Input the graph context embedding sequence into a bidirectional time series encoder to extract time state features.

[0116] The specific steps of step S303 are:

[0117] Step S3031: Rearrange the graph context embedding sequence in time sequence according to the ETC gantry node representation, and construct a set of historical state sequences corresponding to each ETC gantry node.

[0118] Step S3032: For each historical state sequence, set the forward and backward time window ranges, and perform time alignment processing of state embedding in the forward and backward directions respectively.

[0119] In this embodiment, based on the current time point of the target ETC gantry node, a pre-defined length of time window is set forward and backward, forming a forward historical window and a backward future window, the length of the time window is set according to the periodicity of traffic, data sampling frequency and prediction requirements; Then extract the state records of the current node in the forward and backward time windows to form two independent time series subsets.

[0120] Further, the state vectors in the forward window sequence are arranged in time increasing order, and the state vectors in the backward window sequence are arranged in time decreasing order, to construct forward and backward recursive sequences respectively.

[0121] Step S3033: recursively input the embedding vector of the current time step and the previous state context in the forward sequence, and perform the same operation in the reverse direction, record the evolution trajectory sequence of the two directions respectively.

[0122] In this embodiment, in the forward recursive process, first, the embedding vector sequence arranged in time order in the forward time window is selected as input; at the first time step, the state context is empty or initialized as a zero vector, and the current embedding vector is taken as the first recursive input; at each subsequent time step, the state context output at the previous time step and the embedding vector at the current time step are input into the recursive unit together, the state evolution result at the current time step is calculated, and the state evolution result is stored as the current context representation; the above recursive process continues to develop in the forward direction until the end of the time window, thereby forming a complete forward evolution trajectory sequence.

[0123] It should be noted that the structure and logic of the backward recursive process are similar, but the time sequence is reversed.

[0124] Step S3034: concatenate and project the outputs of the forward sequence and the backward sequence to a unified state vector space to obtain the time-dependent feature of each ETC gantry node at the current time.

[0125] In this embodiment, the forward evolution trajectory sequence and the backward evolution trajectory sequence constructed in step S3033 are obtained, and for the current target gantry node, the output state vectors corresponding to the current time in the two direction sequences are extracted, respectively denoted as the forward state vector and the reverse state vector; then the two state vectors are subjected to a dimension concatenation operation to form a double-channel state vector structure; further, the concatenated vector is input into a preset linear transformation module or tensor mapping structure to perform state space projection processing.

[0126] Step S3035: collect the time-dependent features of all ETC gantry nodes in topological order to obtain the time state feature.

[0127] In this embodiment, a highway network topology graph is constructed, the latitude and longitude positions, road numbers and passing directions of the gantry devices actually installed are taken as the basis for node topological relationship, and a logical path graph is established; a uniform time slice is set, for example, at a certain minute, the time-dependent features of all ETC gantry nodes at that time are extracted and labeled with a time label; further, the node topological index table of the road network topology graph is called to arrange all the gantry state vectors in topological order.

[0128] Step S304: under the condition of no external label, the time state feature is subjected to fuzzy clustering and self-supervised contrast learning to generate the initial highway running state of the ETC gantry node.

[0129] The specific steps of step S304 are:

[0130] Step S3041: dividing the time state features into non-overlapping training subsets, each training subset containing time-dependent features of ETC gantry nodes in different time periods.

[0131] Step S3042: for each training subset, membership estimation is performed using the initialized fuzzy clustering center to obtain the membership probability distribution of each ETC gantry node time state feature to different clusters.

[0132] In this embodiment, based on the training data set preset in step S3041, all gantry node state vectors in the same time window are input into the fuzzy clustering calculation module as input samples. When clustering is initialized, the number of clusters C and the corresponding initial clustering center vector are set. The vector can be obtained by performing pre-training on a small number of samples or using a graph aggregation statistical method. Then, for each gantry node state vector, the spatial distance or similarity index between it and each clustering center vector is calculated. Further, according to the above spatial distance result, combined with the preset fuzziness control parameter, the membership function of each node under different clustering centers is constructed.

[0133] It should be noted that the membership estimation process is not to output the final label, but as a fuzzy label distribution, it participates in the subsequent self-supervised representation optimization process and state consistency propagation.

[0134] Step S3043: under the condition of no external label, a positive and negative sample pair is constructed, and ETC gantry nodes belonging to the same time window are taken as positive samples, and ETC gantry nodes belonging to different time periods or different road areas are taken as negative samples.

[0135] In this embodiment, all ETC gantry node time state feature vectors in a fixed time window are extracted as a reference; then, the embedding space distance between any two nodes is calculated, and the physical distance is combined for similarity evaluation; if the distance between two nodes in the embedding space is close, and their physical positions are in the same trunk road or the same area, they are marked as a positive sample pair.

[0136] At the same time of constructing the positive sample pair, two categories of negative sample generation mechanisms are introduced, one is the cross-time negative sample, that is, the gantry node state vectors in different time windows are randomly paired with the current time window node; the other is the cross-area negative sample, that is, the nodes in different road sections and different traffic behavior mode areas are selected from the current time window.

[0137] Step S3044: Perform multi-round self-supervised contrastive coding using the positive and negative sample pairs, update the cluster center vectors of the fuzzy clustering, and adjust the ETC gantry nodes in the embedding space towards the direction of consistency enhancement.

[0138] In this embodiment, first, several batches are randomly sampled from the constructed positive and negative sample pair set as the input of the contrastive coding task; then the minimization operation of the self-supervised contrastive loss function is performed, and for each node, its embedding vector should be as close as possible to the positive sample and as far as possible from the negative sample in the representation space. In the embedding representation update process, the gradient backpropagation mechanism is adopted.

[0139] Further, after each round of contrastive training is completed, based on the updated node embedding representation, the center vector of each cluster in the fuzzy clustering is recalculated, and the node vectors belonging to a cluster are weighted and averaged according to their current membership probability to form a new center representation.

[0140] Step S3045: Determine the initial highway operating state according to the class index to which the maximum membership degree of the ETC gantry node belongs.

[0141] In this embodiment, for any ETC gantry node, the current clustering membership degree vector is retrieved, the vector dimension is consistent with the number of cluster centers, the maximum dimension index is marked by traversing the element values of the vector, and the index is mapped to the corresponding traffic state category, such as the predefined categories of free flow, light congestion, moderate congestion, and heavy congestion, as the initial state label of the node.

[0142] It should be noted that the determination process of the maximum membership degree class index does not depend on any external manual annotation information, but completely depends on the distribution characteristics of the node in the self-organizing embedding space, which is a semantic inference method conforming to the unsupervised paradigm.

[0143] Step S305: Propagate the initial highway operating state using the adjacency relationship in the multi-scale graph structure to obtain the highway operating state.

[0144] In this embodiment, the nodes with initial state labels are regarded as labeled nodes, and unlabeled nodes are regarded as nodes to be predicted. The propagation iteration is carried out on the graph structure based on the principle of state consistency. In each round of propagation process, a state smoothing coefficient and an edge guide adjustment factor are introduced to control the absorption degree of node state and the proportion of upstream / downstream label influence. After the propagation iteration is executed to the stable state, i.e., when the label distribution of all nodes converges or the change amplitude is less than a threshold, the propagation process is terminated.

[0145] Step S4: Based on the highway operating state recognition result, the highway operating state is predicted.

[0146] AsFigure 3 As shown, the specific steps of step S4 are:

[0147] Step S401: Based on the time-dependent features of the ETC gantry node and the corresponding highway operating state, a state-feature pair is constructed, and a prediction sample set is generated.

[0148] In this embodiment, a time series template is established for each ETC gantry node, recording the time-dependent feature vectors and highway state labels of consecutive time steps; a sliding window mechanism is used to construct sample subsequences, each window is divided into an input section and a target section: the input section collects state-feature pairs of a number of historical time steps to learn the state evolution trend; on this basis, all gantry nodes and all time periods are comprehensively sampled to form a cross-period and cross-section prediction sample set, each sample contains a group of time series state-feature pairs as input.

[0149] Step S402: The prediction sample set is divided by a sliding window according to a preset time window, to construct a multi-period input sample sequence and an output sample sequence, wherein the input sample sequence is used to capture the historical state evolution trajectory, and the output sample sequence represents the target prediction period.

[0150] In this embodiment, first, the length parameters of the sliding window are set, including the historical input window length L1 and the future output window length L2; for each ETC gantry node, in its state-feature pair sequence, the window is sequentially slid from the starting position, and the state-feature pairs of consecutive L1 steps are extracted as the input sequence, and the state labels of the subsequent L2 steps are taken as the output sequence.

[0151] Further, a sampling interval parameter is introduced in the sliding window division process, and the window sliding step is set; in actual process, different types of road sections, traffic flow density or gantry layout density will lead to different change rates of time series; according to the historical flow fluctuation amplitude, road topological complexity or prediction accuracy requirement, the values of L1 and L2 can be adaptively adjusted.

[0152] Step S403: The input sample sequence is input into the time series strategy decision network, the state transition path is iterated and the control variable of state evolution is calculated, to obtain the historical state vector and the control variable.

[0153] In this embodiment, each input sample sequence is input into the constructed time series strategy decision network; the network is based on a recursive structure, an attention mechanism or a graph sequence fusion module, receives the node historical state-feature pair sequence as input, and gradually simulates the state transition path in the time dimension.

[0154] Further, the control variables generally include but are not limited to: weight evolution factor of potential correlation between nodes, intervention intensity of historical traffic flow intensity difference on future flow trend, influence amount of local state stability on prediction window scale, etc.; after each iteration deduction is completed, the model simultaneously outputs two intermediate results: compressed historical state vector, representing the overall state evolution trend aggregated in the current input window; and strategy control variable matrix in the corresponding time period, used to describe the structural representation of state transition mechanism at each time.

[0155] Step S404: generating traffic state prediction results of multiple future time steps based on the historical state vector and the control variable.

[0156] In the embodiment, the historical state vector is taken as the starting input vector of the prediction stage, used to initialize the traffic state; then at each prediction time step, the state is updated and evolved in combination with the state input of the current step and the corresponding control variable; the output state vector of each time step will represent the predicted traffic state of the ETC gantry node at the time.

[0157] Step S405: spatiotemporal reorganization of traffic state prediction results of all ETC gantry nodes to obtain the running prediction state of the expressway.

[0158] In the embodiment, the running prediction state representation of all ETC gantry nodes at each target prediction time step is collected, and the prediction results are grouped and arranged according to the spatial positioning information, to establish a time-node index table; the prediction results of adjacent gantry nodes are fused into the corresponding road section level traffic state representation according to the road section division scheme of the gantry.

[0159] Embodiment 2:

[0160] Please refer to Figure 4 Another embodiment provided by the application: an expressway running state recognition and prediction system based on ETC gantry data, comprising: a data processing module, a state tensor construction module, a running state recognition module and a running state prediction module;

[0161] The data processing module is used for collecting expressway ETC gantry data in real time, and performing multi-scale fractal compression on the collected expressway ETC gantry data;

[0162] The state tensor construction module is used for information alignment on the multi-scale fractal compressed expressway ETC gantry data, to construct a four-element traffic state tensor;

[0163] The running state recognition module is used for graph-temporal coding on the expressway and the four-element traffic state tensor, to recognize the running state of the expressway based on the graph-temporal coding result;

[0164] The operation state prediction module predicts the operation state of the expressway based on the expressway operation state recognition result.

[0165] The operation state recognition module comprises a graph-temporal encoding unit and an operation state recognition unit.

[0166] The graph-temporal encoding unit is configured to map the four-element traffic state tensor into a multi-scale graph containing an ETC gantry node, generate a graph context embedding sequence, and extract time state features by using a bidirectional time sequence encoder.

[0167] The operation state recognition unit is configured to perform fuzzy clustering and self-supervised contrast learning on the time state features, and perform consistent propagation on the initial expressway operation state by using the adjacency relationship in the multi-scale graph structure to obtain the expressway operation state.

[0168] In addition, the part of the above technical solution in the embodiments of the present application that is consistent with the implementation principle of the corresponding technical solution in the prior art is not described in detail to avoid excessive repetition.

[0169] The specific embodiments described above further explain the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for identifying and predicting the operating state of a highway based on ETC gantry data, characterized by, The method comprises the following steps: Real-time collection of highway ETC gantry data, and multi-scale fractal compression of the collected highway ETC gantry data; Information alignment of the multi-scale fractal compressed highway ETC gantry data, and construction of a four-element traffic state tensor, wherein the four-element traffic state tensor comprises a vehicle speed vector, a lane occupancy rate, a vehicle type and ETC gantry information; Graph-temporal coding of the highway and the four-element traffic state tensor, and identification of the highway running state based on the graph-temporal coding result; Prediction of the highway running state based on the highway running state identification result; The graph-temporal coding of the highway and the four-element traffic state tensor, and the identification of the highway running state based on the graph-temporal coding result, comprise: Mapping the four-element traffic state tensor to a multi-scale graph containing ETC gantry nodes, constructing directed edges and initializing weights according to the historical traffic flow and trajectory similarity between nodes; Based on the multi-scale graph, the attention message is transmitted to aggregate the node state tensor and generate a graph context embedding sequence; The graph context embedding sequence is input into a bidirectional time sequence encoder to extract time state features; Under the condition of no external label, fuzzy clustering and self-supervised contrast learning are performed on the time state features to generate an initial highway running state of the ETC gantry node; The initial highway running state is propagated based on the adjacency relationship in the multi-scale graph structure to obtain the highway running state. 2.The ETC gantry data-based expressway running state recognition and prediction method of claim 1, wherein, The information alignment of the multi-scale fractal compressed highway ETC gantry data, and the construction of the four-element traffic state tensor, comprise: The multi-scale fractal compressed highway ETC gantry data is divided into a first fragment, a second fragment, a third fragment and a fourth fragment according to data source labels, and the first fragment, the second fragment, the third fragment and the fourth fragment are ETC gantry system fragments, toll station entrance and exit fragments, license plate recognition fragments and video monitoring fragments, respectively; A four-dimensional coordinate mapping model is used to uniformly encode the latitude and longitude, lane number, longitudinal mileage and timestamp of any fragment; An adaptive spatio-temporal grid is generated by filling curves in the uniform encoding domain, and a grid index is assigned to each grid cell; Based on the grid index, the first fragment, the second fragment, the third fragment and the fourth fragment, ETC gantry information, vehicle speed vectors, vehicle types and lane occupancy rates are identified or calculated; The ETC gantry information, vehicle speed vectors, vehicle types and lane occupancy rates are constructed into a four-element traffic state tensor. 3.The ETC gantry data-based expressway running state recognition and prediction method of claim 2, wherein, The aggregation of the node state tensor based on the multi-scale graph attention message transmission, and the generation of the graph context embedding sequence, comprise: From the multi-scale graph node relationship, the four-element traffic state tensors corresponding to the target node and its adjacent nodes are extracted, and an attention query index table between the node pairs is constructed; For each target node, the four-element traffic state tensors of its adjacent nodes are weighted and scored using multi-channel attention, and are arranged in the order of spatial position in the attention query index table; The weighted score result and the four-element traffic state tensor of the corresponding node are subjected to matrix aggregation processing to generate the first round of context representation of the target node; Performing the matrix aggregation process for all target nodes for n rounds, and adjusting the upstream and downstream weight distribution with the ETC gantry timestamp offset factor in each round; After all rounds are completed, perform normalization mapping processing on the intermediate representation of each node, and output the graph context embedding sequence.

4. The ETC gantry data-based expressway operation state recognition prediction method of claim 3, characterized by, Input the graph context embedding sequence into the bidirectional time series encoder to extract time state features, including: Rearrange the graph context embedding sequence in time sequence according to the ETC gantry node representation, and construct a set of historical state sequences corresponding to each ETC gantry node; For each historical state sequence, set the forward and backward time window ranges, and perform time alignment processing of state embedding in the forward and backward directions respectively; Recursively input the embedding vector of the current time step and the previous state context in the forward sequence, and perform the same operation in the reverse direction in the backward sequence, and record the evolution trajectory sequence in the two directions respectively; Concatenate and project the outputs of the forward sequence and the backward sequence into a unified state vector space to obtain the time-dependent features of each ETC gantry node at the current time; Collect the time-dependent features of all ETC gantry nodes in topological order to obtain the time state features. 5.The ETC gantry data-based expressway running state recognition and prediction method of claim 4, wherein, The fuzzy clustering and self-supervised contrast learning of the time state features under the condition of no external label to generate the initial highway running state of the ETC gantry node, including: Divide the time state features into non-overlapping training subsets, each training subset containing the time-dependent features of the ETC gantry node in different time periods; For each training subset, use the initialized fuzzy clustering center to estimate the membership, and obtain the membership probability distribution of each ETC gantry node time state feature to different clusters; Under the condition of no external label, construct positive and negative sample pairs, and take the ETC gantry nodes belonging to the same time window as positive samples, and take the ETC gantry nodes belonging to different time periods or different road regions as negative samples; Use the positive and negative sample pairs for multiple rounds of self-supervised contrast coding to update the cluster center vector of fuzzy clustering and adjust the ETC gantry node in the embedding space towards the direction of consistency enhancement; Determine the initial highway running state according to the class index to which the maximum membership of the ETC gantry node belongs. 6.The ETC gantry data-based expressway running state recognition and prediction method of claim 5, wherein, The prediction of the highway running state based on the highway running state recognition result, including: Based on the time-dependent features of the ETC gantry node and the corresponding highway running state, construct a state-feature pair and generate a prediction sample set; Slide and divide the prediction sample set according to the preset time window to construct multiple time period input sample sequences and output sample sequences, where the input sample sequence is used to capture the historical state evolution trajectory, and the output sample sequence represents the target prediction period; Input the input sample sequence into the time series strategy decision network to iterate the state transition path and calculate the control variable of state evolution to obtain the historical state vector and control variable; Based on the historical state vector and control variable, generate traffic state prediction results for multiple future time steps; The traffic state prediction results of all ETC gantry nodes are spatiotemporally reorganized to obtain the running prediction state of the expressway. 7.The ETC gantry data-based expressway running state recognition and prediction method of claim 6, wherein, The ETC gantry data of the expressway is collected in real time, and the collected ETC gantry data of the expressway is subjected to multi-scale fractal compression, including: A pseudo-random hash function is used to generate a unique spatiotemporal index code for each piece of ETC gantry data of the expressway; Based on the spatiotemporal index code, the same-period ETC gantry data segments of the expressway are subjected to secondary fractal segmentation, and the corresponding fractal dimension spectrum is calculated; The fractal dimension spectrum is mapped to an orthogonal basis to construct a quantum fractal measure matrix; Based on a preset topological entropy threshold, the quantum fractal measure matrix is subjected to pruning operation layer by layer to obtain a compressed data sequence; A preset Markov correlation length is used to sparsely reorganize the compressed data sequence to obtain the multi-scale fractal compressed ETC gantry data of the expressway.

8. A highway running state recognition and prediction system based on ETC gantry data, for implementing the highway running state recognition and prediction method based on ETC gantry data according to any one of claims 1-7, characterized in that, It includes: a data processing module, a state tensor construction module, a running state recognition module, and a running state prediction module; The data processing module is configured to collect ETC gantry data of the expressway in real time and perform multi-scale fractal compression on the collected ETC gantry data of the expressway; The state tensor construction module is configured to align information of the multi-scale fractal compressed ETC gantry data of the expressway and construct a four-element traffic state tensor; The running state recognition module is configured to perform graph-temporal coding on the expressway and the four-element traffic state tensor, and recognize the running state of the expressway based on the graph-temporal coding result; The running state prediction module is configured to predict the running state of the expressway based on the running state recognition result of the expressway. 9.The ETC gantry data-based highway operation state recognition and prediction system of claim 8, wherein, The running state recognition module includes a graph-temporal coding unit and a running state recognition unit; The graph-temporal coding unit is configured to map the four-element traffic state tensor to a multi-scale graph containing ETC gantry nodes, generate a graph context embedding sequence, and extract time state features using a bidirectional time series encoder; The running state recognition unit is configured to perform fuzzy clustering and self-supervised contrast learning on the time state features, and perform consistency propagation on the initial running state of the expressway using the adjacency relationship in the graph structure to obtain the running state of the expressway.

Images