An artificial intelligence-based urban spatiotemporal feature mining method and system

Abstract

The application discloses an artificial intelligence-based city space-time feature mining method and system, and relates to the field of deep learning.The method comprises the following steps: acquiring traffic network data and connectivity of a target area to construct a space-time data matrix; mapping the network into a dynamic road network graph, and dynamically determining a graph convolution receptive field range according to node connectivity and a local clustering coefficient; combining the receptive field and the data matrix to extract spatial and temporal feature vectors and perform preliminary weighted fusion; introducing macro statistical features to identify a current traffic scene category, and dynamically and adaptively adjusting feature fusion weights to reconstruct target space-time features according to the category; and inputting the target features into a pre-trained decoding network to perform dimension reduction and quantization, and outputting an adaptive traffic signal timing strategy.The technical scheme provided by the application improves the dynamic adaptability of an algorithm model to heterogeneous network structures and sudden traffic scenes.

Description

Technical Field

[0001] This application relates to the field of deep learning, and in particular to a method and system for mining urban spatiotemporal features based on artificial intelligence. Background Technology

[0002] With the continuous acceleration of urbanization and the surge in motor vehicle ownership, artificial intelligence-based intelligent transportation systems play a crucial role in alleviating traffic pressure and improving the overall traffic efficiency of the road network.

[0003] In related technologies, a static road network map with a fixed topology is pre-constructed based on the physical distance or adjacency relationship between road nodes. Then, a graph convolutional network is used to extract the spatial correlation features of the road network within a pre-defined fixed receptive field. Finally, a temporal model such as a long short-term memory network is used to fuse the spatiotemporal features along time and apply them to the control and scheduling of traffic signals.

[0004] However, when a local road network encounters severe weather or a sudden traffic accident that causes a nonlinear change in traffic flow patterns, feature fusion based on historical normal patterns can lead to insufficient dynamic adaptive capability of the algorithm model for heterogeneous road network structures and sudden traffic scenarios, making it prone to outputting ineffective control strategies under abnormal conditions. Summary of the Invention

[0005] This application provides an artificial intelligence-based method and system for mining urban spatiotemporal features, which can improve the dynamic adaptability of the algorithm model to heterogeneous road network structures and sudden traffic scenarios.

[0006] The first aspect of this application provides an artificial intelligence-based method for mining urban spatiotemporal features, the method comprising: The system acquires spatiotemporal correlation data and road connectivity relationships of the traffic network within the target area to construct a spatiotemporal data matrix. Intersections and road connectivity relationships in the traffic network are mapped to graph nodes and edges to construct a dynamic road network graph. The corresponding graph convolution receptive field is dynamically determined based on the connectivity degree and local clustering coefficient of each graph node. Spatial feature vectors of the dynamic road network graph are extracted based on the spatiotemporal data matrix and the graph convolution receptive field, and temporal feature vectors are obtained by performing time-series calculations on the spatiotemporal data matrix. The spatial and temporal feature vectors are weighted and fused to obtain an initial spatiotemporal feature vector. The current traffic scenario category is determined based on the macroscopic statistical features of the spatiotemporal data matrix and the initial spatiotemporal feature vector. The feature weight elements in the weighted fusion calculation are dynamically adjusted accordingly to recalculate the target spatiotemporal feature vector. The target spatiotemporal feature vector is input into a pre-trained timing parameter decoding network for mapping, dimensionality reduction, and quantization calculations, outputting the corresponding traffic signal timing strategy.

[0007] In the above embodiments, by dynamically establishing the receptive field of graph convolution and fusing two-dimensional spatiotemporal features, and by introducing scene category as a condition in the feature fusion stage to perform dynamic adaptive weight adjustment, the underlying network has the ability to dynamically perceive heterogeneous road network structures and complex traffic scenarios, thereby improving the accuracy and robustness of urban traffic hub signal control strategies in non-steady-state environments.

[0008] In conjunction with some embodiments of the first aspect, in some embodiments, before inputting the target spatiotemporal feature vector into a pre-trained timing parameter decoding network for mapping, dimensionality reduction, and quantization calculations, the method further includes: This paper extracts real-time microscopic anomaly distribution features of traffic flow at intersections experiencing abnormal changes from the spatiotemporal data matrix to locate topological failure nodes causing severe traffic congestion and deadlock. It then performs topological isolation processing on the graph edge weights between the topological failure node and its adjacent connected nodes in the dynamic road network diagram to perform local graph topological disconnection. Based on the abnormal detour directions of traffic flow characterized by the real-time microscopic anomaly distribution features, it determines the surrounding secondary diversion nodes receiving abnormal detour traffic flow and establishes temporary graph edges between the adjacent nodes of the topological failure node and the surrounding secondary diversion nodes. It obtains the real-time traffic flow transfer load parameters corresponding to the temporary graph edges and dynamically assigns edge weights, updating the dynamic road network diagram to a transient reconstructed road network diagram. The transient reconstructed road network diagram is then re-substituted into a graph convolutional network to perform secondary spatial feature extraction, obtaining a mutation spatial feature vector reflecting the extreme detour diffusion pattern of traffic flow. The mutation spatial feature vector is used as a compensation term and fused with the initial spatiotemporal feature vector to obtain a deeply corrected spatiotemporal feature vector for sudden nonlinear mutation scenarios. This deeply corrected spatiotemporal feature vector replaces the target spatiotemporal feature vector and is input into the timing parameter decoding network.

[0009] In the above embodiments, for extreme scenarios such as sudden traffic accidents or partial paralysis of the road network, by forcibly cutting off the failed nodes in the underlying mathematical graph and establishing temporary detour edges, the abnormal shock wave transmission of traffic flow in the physical road network is captured, ensuring that the spatiotemporal features extracted by the model are not skewed by the erroneous data of local nodes.

[0010] In conjunction with some embodiments of the first aspect, in some embodiments, edge weights are dynamically assigned based on real-time traffic flow transfer load parameters, and the dynamic road network map is updated to a transiently reconstructed road network map, specifically including: When a sudden nonlinear variation scenario is detected and the real-time meteorological monitoring value exceeds the preset disaster warning threshold, the spatial physical environment data and real-time meteorological evolution sequence of the surrounding secondary diversion nodes and the corresponding areas of the temporary map edge are acquired. Based on the spatial physical environment data and the real-time meteorological evolution sequence, environmental disaster-related traffic loss calculation is performed to calculate the dynamic traffic loss rate of the physical road segment corresponding to the temporary map edge. A traffic constraint penalty mechanism is constructed according to the dynamic traffic loss rate, and the environmental impedance coefficient for the temporary map edge is calculated. The environmental impedance coefficient is used to attenuate and correct the real-time traffic flow transfer load parameter to obtain the restricted transfer load parameter that integrates environmental constraints. The restricted transfer load parameter is compared with the set topological connectivity validity threshold. If it is lower than the threshold, secondary topological pruning is performed on the temporary map edge. If it is not lower than the threshold, the restricted transfer load parameter is assigned as the final weight to the temporary map edge to complete the update of the transient reconstructed road network map.

[0011] In the above embodiments, by integrating severe weather thresholds and physical environment evolution sequences during the graph network topology generation process, physical-level dimensionality reduction constraints on road network capacity are achieved. The secondary topology pruning operation based on environmental disaster impedance prevents AI algorithms from blindly guiding traffic flow to high-risk areas during disasters such as rainstorms, thus endowing the reconstructed road network with a high degree of safety compliance and physical realism.

[0012] In conjunction with some embodiments of the first aspect, in some embodiments, the current traffic scenario category is determined based on the macroscopic statistical characteristics of the spatiotemporal data matrix and the initial spatiotemporal feature vector, specifically including: Low-dimensional macroscopic traffic and environmental statistical elements are extracted from the spatiotemporal data matrix to obtain a macroscopic statistical feature matrix representing the explicit physical state as macroscopic statistical features. The macroscopic statistical feature matrix and the initial spatiotemporal feature vector are concatenated across dimensions to generate a joint representation vector that integrates explicit and implicit features. The joint representation vector is then input into a preset feature classification algorithm for classification and identification to determine the current traffic scenario category.

[0013] In the above embodiments, a cross-dimensional feature splicing mechanism is adopted to directly inject low-dimensional explicit physical statistical data into high-dimensional AI implicit mathematical inference vectors, avoiding feature extraction bias when a single graph neural network faces sudden abnormal events, and improving the recognition accuracy of traffic disaster scenarios under complex and intertwined conditions.

[0014] In conjunction with some embodiments of the first aspect, in some embodiments, the macroscopic statistical feature matrix and the initial spatiotemporal feature vector are concatenated across dimensional feature channels to generate a joint representation vector that integrates explicit and implicit features, specifically including: Extract multidimensional spatiotemporal fluctuation assessment parameters from the macroscopic statistical feature matrix; if the multidimensional spatiotemporal fluctuation assessment parameters exceed the set modal confidence safety threshold, it is determined that the current explicit physical data is subject to modal distortion due to environmental interference; calculate the feature space mapping difference between the macroscopic statistical feature matrix with modal distortion and the initial spatiotemporal feature vector to quantify the degree of cross-modal feature conflict; construct a dynamic adaptive feature isolation mask based on the degree of cross-modal feature conflict, and use the feature isolation mask to adaptively suppress the high-noise conflict channels in the macroscopic statistical feature matrix; concatenate the macroscopic statistical feature matrix after weight suppression with the initial spatiotemporal feature vector across dimensions to generate a joint representation vector.

[0015] In the above embodiments, an abnormal noise data detection and adaptive suppression mechanism is set before performing explicit and implicit feature stitching, which solves the risk caused by the failure of front-end sensors due to severe weather. By constructing a soft-gated mask to accurately isolate contaminated high-noise channels, the algorithm model's anti-interference immunity and cross-modal generalization ability are enhanced when the front-end physical data is severely degraded.

[0016] In conjunction with some embodiments of the first aspect, in some embodiments, intersections in the traffic network are mapped as graph nodes, and road connectivity relationships are mapped as graph edges to construct a dynamic road network graph, specifically including: Obtain the road grade coefficients and historical traffic flow transmission rates corresponding to road connectivity relationships; calculate the weights of the road grade coefficients and historical traffic flow transmission rates according to a preset ratio to obtain the weights of the graph edges and complete the construction of the dynamic road network map.

[0017] In the above embodiments, the graph edge weights between nodes are initialized by comprehensively considering the static physical road level of infrastructure and the dynamic historical traffic flow transmission rate, thus constructing an initial bottom-level graph that takes into account both the hard constraints of traffic and the soft evolution law of traffic flow, thereby improving the simulation fidelity of spatial feature extraction.

[0018] In conjunction with some embodiments of the first aspect, in some embodiments, time-series calculations are performed on the spatiotemporal data matrix to obtain a time feature vector, specifically including: The spatiotemporal data matrix is ​​input into a pre-defined bidirectional temporal neural network; historical dependency context features of the spatiotemporal data matrix are extracted along the forward calculation dimension of the time series, and reverse temporal evolution features are extracted along the reverse calculation dimension of the time series; the historical dependency context features and reverse temporal evolution features are concatenated and calculated to obtain a time feature vector representing the bidirectional fluctuation pattern of traffic flow.

[0019] In the above embodiments, deep bidirectional temporal modeling technology is used to capture the patterns of historical states in the forward direction while taking into account the inference and correlation of the reverse temporal context. This avoids the information forgetting blind spots that are easy to occur in traditional unidirectional temporal networks, and provides a more three-dimensional and comprehensive time series feature base for downstream graph processing.

[0020] Secondly, embodiments of this application provide an artificial intelligence-based urban spatiotemporal feature mining system, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, which includes computer instructions, and the one or more processors call the computer instructions to cause the artificial intelligence-based urban spatiotemporal feature mining system to perform the method described in the first aspect and any possible implementation thereof.

[0021] Thirdly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on an AI-based urban spatiotemporal feature mining system, cause the AI-based urban spatiotemporal feature mining system to execute the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on an AI-based urban spatiotemporal feature mining system, cause the AI-based urban spatiotemporal feature mining system to perform the method described in the first aspect and any possible implementation thereof.

[0023] Understandably, the AI-based urban spatiotemporal feature mining system provided in the second aspect, the computer program product provided in the third aspect, and the computer storage medium provided in the fourth aspect are all used to execute the AI-based urban spatiotemporal feature mining method provided in the embodiments of this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. This application dynamically establishes the receptive field of graph convolution and fuses two-dimensional spatiotemporal features. At the same time, scene category is introduced as a condition in the feature fusion stage to perform dynamic adaptive weight adjustment, which enables the underlying network to have dynamic perception capabilities for heterogeneous road network structures and complex traffic scenarios, thereby improving the accuracy and robustness of urban traffic hub signal control strategies in non-steady-state environments.

[0025] 2. In extreme scenarios such as sudden traffic accidents or partial paralysis of the road network, this application captures the abnormal shock wave transmission of traffic flow in the physical road network by forcibly cutting off the failed nodes in the underlying mathematical graph and establishing temporary detour edges, ensuring that the spatiotemporal features extracted by the model will not be skewed by the erroneous data of local nodes.

[0026] 3. This application achieves physical-level dimensionality reduction constraints on road network capacity by integrating severe weather thresholds and physical environment evolution sequences during the graph network topology generation process. The secondary topology pruning operation based on environmental disaster impedance prevents AI algorithms from blindly guiding traffic flow to high-risk areas during disasters such as rainstorms, thus endowing the reconstructed road network with a high degree of safety compliance and physical realism. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating an artificial intelligence-based urban spatiotemporal feature mining method in an embodiment of this application. Figure 2 This is a schematic diagram of dynamic road network graph construction and edge weight calculation in an embodiment of this application; Figure 3 This is a schematic diagram of the hierarchical structure of the spatiotemporal feature extraction and fusion model in the embodiments of this application; Figure 4 This is another flowchart illustrating the artificial intelligence-based urban spatiotemporal feature mining method in this application embodiment; Figure 5 This is an exemplary hardware structure diagram of an artificial intelligence-based urban spatiotemporal feature mining system in this application embodiment. Detailed Implementation

[0028] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0029] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0030] Figure 1This is a flowchart illustrating the urban spatiotemporal feature mining method based on artificial intelligence as described in this application, including the following steps: S101. Obtain the spatiotemporal correlation data of the traffic network and road connectivity within the target area, and construct a spatiotemporal data matrix.

[0031] The target area is typically defined as a specific geographic space requiring traffic signal control (e.g., a central urban business district or a key transportation hub); the traffic network refers to the topological entity system formed by the interweaving of multiple physical roads and intersections within this area. The spatiotemporal correlation data acquired for this network encompasses traffic operation indicators with both temporal and spatial dimensions, such as cross-sectional traffic flow and average queue length at each intersection over consecutive time steps; road connectivity essentially characterizes the distribution of directly connected physical lanes between intersections in the physical network. Based on this prior information, a spatiotemporal data matrix can be constructed using tensor alignment processing of the multidimensional heterogeneous data to serve as the carrier for subsequent calculations.

[0032] An AI-based urban spatiotemporal feature mining system collects raw, multi-source traffic data through detection terminals deployed within the target area. To eliminate noise in the underlying data, the system first performs missing value imputation and outlier smoothing filtering on the acquired raw data, and then resamples the time series data at a uniform clock frequency to ensure that the data from heterogeneous terminals maintain consistency in time steps. After cleaning and alignment, the system extracts the traffic flow parameters of each intersection as feature dimensions, and uses the intersection number and time slice as spatial and temporal dimensions, respectively. It then expands these dimensions and uses tensor concatenation operations to generate a two-dimensional spatiotemporal data matrix representing the three-dimensional attributes.

[0033] The construction process of the spatiotemporal data matrix can be flexibly implemented in various ways depending on the type of underlying data source. For example, in one specific implementation, the system can read microscopic vehicle trajectory data and perform spatiotemporal clustering analysis on it to extract the macroscopic traffic flow density at the road segment level. Then, this macroscopic traffic flow density is combined with the static coordinates of the road network and mapped to a multidimensional tensor to output the spatiotemporal data matrix. In another implementation, the system can also directly retrieve the static topology table of the road network, perform feature aggregation on the cross-sectional traffic flow collected by microwave radar according to fixed time window slices, and then fill the pre-initialized zero matrix with the aggregated traffic flow as node attribute features to complete the construction. It is understood that other methods can also be used to construct the spatiotemporal data matrix, which are not limited here.

[0034] S102. Map intersections in the traffic network to graph nodes and road connectivity to graph edges to construct a dynamic road network graph. Dynamically determine the graph convolution receptive field of each graph node based on its connectivity degree and local clustering coefficient.

[0035] An intersection refers to a vehicle diversion hub where multiple roads intersect in a physical road network; a graph node is an abstract operational unit representing a single intersection in a graph theory mathematical model; a graph edge is a line segment connecting two graph nodes in a graph theory mathematical model to represent the connectivity attribute of physical lanes; a dynamic road network graph refers to a graph structure that updates the node and edge attributes as the road network impedance or connectivity state changes; the connectivity degree is used to represent the number of neighboring nodes directly connected to a single graph node; the local clustering coefficient is used to represent the probability density of mutual connectivity between adjacent nodes of a graph node; the receptive field of a graph convolution refers to the maximum number of topological hops that a single central node can perceive and aggregate the features of its surrounding neighboring nodes when performing graph neural network aggregation operations.

[0036] The system establishes an adjacency matrix for the graph structure based on the acquired road connectivity relationships. For each graph node in the constructed dynamic road network graph, the system traverses the adjacency matrix to calculate its connectivity degree and calculates the local clustering coefficient based on the ratio of the actual number of edges between the node's neighbors to the maximum possible number of edges. Subsequently, the system introduces a linear weighting formula or a nonlinear mapping function, using the connectivity degree and the local clustering coefficient as input variables, to calculate the corresponding continuous topological complexity evaluation value. This evaluation value is then rounded up to serve as the hop count for the graph convolution operation, thereby limiting the receptive field of each graph node's graph convolution.

[0037] In some embodiments, the dynamic determination of the receptive field of graph convolution can be achieved in several ways: Optionally, the connectivity degree and local clustering coefficient of a single node are calculated based on the graph structure; static weights are set for both and a weighted sum is performed to obtain a baseline value; the baseline value is compared with a ladder hop count mapping table, and the corresponding discrete hop count is output as the receptive field range; Optionally, the reciprocal of the local clustering coefficient is extracted to evaluate the local divergence of the node; the connectivity degree is multiplied by the divergence index to obtain the global information demand of the node; the global information demand is smoothly mapped to the corresponding network perception layer number by taking the logarithm, which is the final receptive field range. It is understood that other methods can also be used to dynamically determine the receptive field range of graph convolution, which are not limited here.

[0038] To more intuitively illustrate the quantization process of mapping the physical road network to the dynamic road network map, please refer to [link to relevant documentation]. Figure 2 . Figure 2 This is a schematic diagram illustrating the dynamic road network graph construction and edge weight calculation in an embodiment of this application. For example... Figure 2As shown, intersections in the physical road network are abstracted as graph nodes N1 to N8, and road connectivity is mapped as directional graph edges. The thickness of the edges intuitively reflects the calculated comprehensive weight. Taking the graph edge from node N1 to node N2 as an example, assuming the system extracts that the directly connected road belongs to an urban arterial road, the corresponding road grade coefficient is quantified as 0.8, and the historical traffic flow transmission rate calculated from historical business data is 0.5. The system performs a weighted summation calculation according to a preset ratio (e.g., road grade coefficient accounts for 0.6, historical traffic flow transmission rate accounts for 0.4), i.e.: edge weight = 0.6 × 0.8 + 0.4 × 0.5 = 0.68. This calculation result of 0.68 is used as the final weight value assigned to the graph edge. In this way, this application enables the underlying graph topology structure to be deeply bound to the physical infrastructure attributes and objective traffic flow patterns of the real road network, avoiding ineffective feature propagation by the algorithm under equal weights that are detached from physical reality.

[0039] In situations where heavy rainstorms cause flooding at multiple underpasses, resulting in physical disconnections in the road network and the emergence of numerous isolated subgraphs, a virtual bridging node reconstruction mechanism can be introduced. The system calculates the number of isolated subgraphs in the dynamic road network graph caused by physical disconnections. When the number of isolated subgraphs exceeds the normal topological connectivity, the system extracts the geographic straight-line distance between the boundary nodes of the isolated subgraphs and the boundary nodes of the backbone network. Virtual edge weights are constructed using the geographic straight-line distance and the corresponding municipal emergency diversion road level as parameters, and virtual bridging graph edges are injected between the isolated subgraphs and the backbone network. By introducing this virtual bridging node reconstruction mechanism, feature collapse caused by adjacency matrix singularity in graph convolution operations under extreme flooding and isolated scenarios is prevented, maintaining the connectivity of global feature transmission and improving the accuracy of urban traffic hub signal control strategies in non-steady-state environments.

[0040] When mapping road network connectivity to an abstract mathematical graph structure, the edge weights of the graph can be initialized by combining road grade and historical traffic flow transmission rate, thereby improving the rationality of the physical representation of the underlying road network graph.

[0041] To address the edge weight determination process in dynamic road network map construction, the system acquires road grade coefficients and historical traffic flow transmission rates corresponding to road connectivity relationships. Specifically, the system accesses a geographic information database to extract basic road network attribute parameters of directly connected roads between adjacent intersections and quantifies them into static road grade coefficients. This introduces a physical hard constraint on the road design capacity limit. Simultaneously, the system retrieves historical business data to calculate the probability distribution of traffic flow transfer between adjacent nodes and outputs a dynamic historical traffic flow transmission rate. This operation supplements the soft dynamic characteristics of driver travel preferences and the normal tidal patterns of the road network. Subsequently, the system performs a weighted summation calculation of the road grade coefficients and historical traffic flow transmission rates according to a preset ratio. This preset ratio is a pre-defined hyperparameter weight allocation derived by those skilled in the art based on the traffic flow transmission distribution patterns of different road network grades at different historical time periods through multiple linear regression. The system multiplies the normalized road grade coefficients and historical traffic flow transmission rates by the corresponding factors of this preset ratio and adds the values ​​to obtain a floating-point value representing the comprehensive information transmission intensity of the physical road segment, which serves as the weight of the graph edge to complete the construction of the dynamic road network map. This weighted summation mechanism allows graph edge attributes to be extended beyond simple Boolean connectivity markers to include continuous mathematical variables containing multidimensional prior traffic knowledge. This provides a more accurate information flow channel for subsequent graph network models to aggregate neighbor node features, reflecting the actual physical carrying capacity of road networks.

[0042] By integrating static road grades reflecting physical carrying capacity with dynamic traffic flow transmission rates characterizing traffic operation patterns, the graph edge weights are established, enabling the graph network topology to balance infrastructure constraints and traffic flow preferences. This standardizes the information transmission direction when aggregating node features in the graph network, improving the rationality of subsequent spatial feature extraction.

[0043] S103. Based on the spatiotemporal data matrix and the receptive field of graph convolution, extract the spatial feature vector of the dynamic road network map, and perform time series calculation on the spatiotemporal data matrix to obtain the temporal feature vector.

[0044] Spatial feature vectors refer to low-dimensional representations of spatial distribution obtained by aggregating the attribute information of adjacent nodes in the road network topology; temporal computation refers to the computational process of extracting dependencies from data arranged along the time axis using a network model with memory units; temporal feature vectors refer to low-dimensional mathematical representations of the fluctuation patterns and trends of traffic flow in the temporal evolution dimension.

[0045] The system constructs a dual-stream feature extraction architecture consisting of a spatial graph convolutional neural network and a temporal series recurrent neural network. Spatially, the system inputs the spatiotemporal data matrix into the graph convolutional network, aggregating features from neighboring nodes layer by layer according to the receptive field of each graph node, and calculating the output spatial feature vector through matrix multiplication and a nonlinear activation function. Temporally, the system transforms the spatiotemporal data matrix into a time-series format and inputs it into a long short-term memory network or a gated recurrent unit. Through an internal computational gating mechanism, redundant historical information is forgotten and the current cell state is updated, mapping and extracting the latent state vectors at different time steps into temporal feature vectors.

[0046] In some embodiments, spatial and temporal feature vectors can be extracted in various ways: Optionally, a graph attention network mechanism can be applied to assign attention coefficients to spatial neighbor nodes; adjacency features with attention coefficients can be aggregated based on the receptive field of the graph convolution to generate spatial feature vectors; a one-dimensional dilated convolution operation of a temporal convolution network can be used to extract temporal feature vectors in the temporal dimension; Optionally, a spectral domain graph convolution operator based on Chebyshev polynomials can be constructed; the polynomial order can be set in conjunction with the dynamically determined receptive field to calculate spatial feature vectors; and historical time step sequences can be read forward and backward through a bidirectional recurrent neural network, and the hidden state vectors can be concatenated to output temporal feature vectors. It is understood that other methods can also be used to extract feature vectors, which are not limited here.

[0047] S104. Perform weighted fusion calculation on the spatial feature vector and the temporal feature vector to obtain the initial spatiotemporal feature vector.

[0048] Weighted fusion computation refers to the operation method of concatenating or aggregating feature vectors of different modalities or different dimensions after assigning numerical proportional weights to them; the initial spatiotemporal feature vector refers to the basic spatiotemporal joint mathematical representation that has not been modified by secondary intervention of external environmental parameters or scene information.

[0049] The system initializes the spatial and temporal attention weight matrices and calculates the cross-covariance between the spatial and temporal feature vectors. Based on the numerical distribution of the cross-covariance, the values ​​of the spatial and temporal attention weight matrices are dynamically adjusted. Subsequently, the system performs a dot product between the adjusted spatial attention weight matrix and the spatial feature vector, and a dot product between the adjusted temporal attention weight matrix and the temporal feature vector. The results of these dot products are then element-wise added and processed through a batch normalization layer to output the initial spatiotemporal feature vector.

[0050] In some embodiments, weighted fusion computation can be implemented in several ways: Optionally, a fixed ratio of spatial fusion coefficients and temporal fusion coefficients is set; the two types of coefficients are multiplied by their corresponding feature vectors respectively; the multiplied vectors are concatenated along the feature channel dimension, and the initial spatiotemporal feature vector is output through dimensionality reduction using a fully connected layer; Optionally, the spatial feature vector and temporal feature vector are input into a multilayer perceptron to obtain a fusion gate value; the fusion gate value is used to control the inflow ratio of the temporal feature vector, and the complement of the fusion gate value is used to control the inflow ratio of the spatial feature vector; the two sets of vectors after flow limiting are added to obtain the initial spatiotemporal feature vector. It is understood that other methods can also be used to implement weighted fusion computation, which are not limited here.

[0051] To illustrate the computational data flow of the underlying network model described above, please refer to [link / reference]. Figure 3 , Figure 3 This is a schematic diagram of the hierarchical structure of the spatiotemporal feature extraction and fusion model in an embodiment of this application. For example... Figure 3 As shown, the system uses preprocessed spatiotemporal data tensors as a unified input. In the spatial processing branch, the model first constructs a dynamic road network graph and dynamically establishes the receptive field of the graph convolution (i.e., dynamically determines the number of perception layers for topological neighbors) based on the connectivity and clustering coefficients of each node, thereby guiding the spatial feature layer (graph convolutional network) to extract spatial feature vectors. Simultaneously, in the temporal processing branch, the spatiotemporal data tensors are synchronously input into the temporal feature layer, which uses a bidirectional temporal network (such as Bi-LSTM) to extract sequence dependencies along both forward and reverse time dimensions, outputting temporal feature vectors. Subsequently, the spatial and temporal feature vectors are jointly fed into an attention-weighted fusion layer, where attention weights are dynamically allocated by calculating cross-covariance and dot product and element-wise addition operations are performed, ultimately outputting an initial spatiotemporal feature vector that takes into account both dimensions. This dual-stream architecture design ensures the independence and complementarity of spatial topology perception and temporal evolution inference.

[0052] S105. Determine the current traffic scenario category based on the macroscopic statistical features of the spatiotemporal data matrix and the initial spatiotemporal feature vector, and dynamically adjust the feature weight elements in the weighted fusion calculation according to the traffic scenario category to recalculate the target spatiotemporal feature vector.

[0053] Macro-statistical characteristics refer to conventional statistical indicators that reflect the overall average operating level of the road network, such as the average congestion index or regional vehicle density; the current traffic scenario category refers to the qualitative classification of the current macro state of the road network; the feature weight element refers to the floating-point value that determines the proportion of spatiotemporal features in the weighted fusion operation; the target spatiotemporal feature vector refers to the final spatiotemporal joint representation with high-dimensional generalization ability output after scenario category perception and adaptive weight optimization.

[0054] The system extracts macroscopic statistical features such as variance and mean from the spatiotemporal data matrix, concatenates them with the initial spatiotemporal feature vector, and inputs them into a pre-trained scene classifier. It calculates the probability distribution of each scene type using a normalized exponential function and selects the category corresponding to the highest probability as the current traffic scene category. Subsequently, the system retrieves a pre-stored scene-weight mapping dictionary and reads the spatial dimension adjustment compensation factor and temporal dimension adjustment compensation factor corresponding to the traffic scene category. The system updates the weights by multiplying the feature weight elements with the corresponding adjustment compensation factors, and then recalculates the feature combination on the spatial and temporal feature vectors using the updated weights to generate the target spatiotemporal feature vector.

[0055] In some embodiments, the recalculation of the target spatiotemporal feature vector can be achieved in several ways: Optionally, macroscopic statistical indicators obtained from physical monitoring are extracted and input into a feedforward neural network; the current traffic scene category is jointly determined as morning / evening peak or off-peak by combining the initial spatiotemporal feature vector; a rule engine is introduced to determine the current category, and the corresponding weight penalty variable is called according to different categories. The penalty variable is then applied to update the fusion parameters to calculate the output target spatiotemporal feature vector; Optionally, principal component analysis is performed on the spatiotemporal data matrix to extract macroscopic statistical features; a support vector machine is used to map the statistical features and the initial feature vector to a high-dimensional space to classify the current traffic scene category; a category-based weight gain bias is adaptively generated based on back gradient propagation, and the gain bias is injected into the feature fusion function to output the target spatiotemporal feature vector. It is understood that other methods can also be used to recalculate the target spatiotemporal feature vector, which are not limited here.

[0056] In some embodiments, when a large-scale power outage at communication base stations leads to a mass loss of macroscopic statistical features in the spatiotemporal data matrix transmitted from physical sensors to the central node, a generative adversarial completion mechanism can be introduced. The system scans the spatiotemporal data matrix to locate the masking locations of missing data holes in the macroscopic statistical features; it retrieves a generative adversarial network (GAN) model trained on a complete historical traffic state set; the incomplete matrix with missing hole masks and the prior structure of the road network's geographical topology are input into the GAN's generator, which generates forged macroscopic features to fill the holes based on the spatiotemporal fluctuation features of adjacent normal base stations through transposed convolution operations; a discriminator iteratively judges the reasonableness of the distribution of the forged macroscopic features using a game theory approach, and after the game converges, outputs high-confidence completed macroscopic statistical features to participate in subsequent scene classification. By introducing the generative adversarial completion mechanism, the algorithm is prevented from collapsing when data is physically disconnected due to external disasters, ensuring the continuity of scene recognition under sudden network outage conditions and improving the accuracy of urban traffic hub signal control strategies in non-steady-state environments.

[0057] In some embodiments, when faced with sudden situations where the collected data deteriorates due to interference from harsh environments on the front-end physical sensors, the explicit statistical features can be anomaly detected and masked before being combined with the implicit spatiotemporal features for cross-dimensional classification, thus avoiding noise pollution of the scene classification model from the underlying data.

[0058] To determine the current traffic scenario category, the system extracts low-dimensional macroscopic traffic and environmental statistical elements from the spatiotemporal data matrix, obtaining a macroscopic statistical feature matrix representing the explicit physical state. This operation introduces objective physical boundary information, compensating for the lack of intuitive physical constraints in single implicit neural network features. Furthermore, the system extracts multidimensional spatiotemporal fluctuation evaluation parameters from this macroscopic statistical feature matrix and compares them with a set modal confidence safety threshold. This threshold is a reference critical value predetermined by those skilled in the art based on the variance statistical distribution of historical normal traffic operation data through multiple sets of comparative experiments. When the evaluation parameter exceeds this threshold, the system determines that the currently collected explicit physical data has undergone modal distortion due to external environmental interference. This determination mechanism can identify and intercept degraded front-end data before heterogeneous feature fusion. After confirming modal distortion, the system calculates the feature space mapping difference between the distorted macroscopic statistical feature matrix and the initial spatiotemporal feature vector to quantify the degree of cross-modal feature conflict between explicit and implicit features. Quantifying the difference provides a mathematical basis for subsequent feature adjustment amplitude. Based on the calculated feature conflict level, the system constructs a dynamically adaptive feature isolation mask. This mask is used to adaptively suppress high-noise conflict channels with numerical anomalies in the macroscopic statistical feature matrix. This mask suppression can reduce signal distortion caused by environmental noise while preserving usable macroscopic data. Subsequently, the system performs cross-dimensional feature channel concatenation between the weighted macroscopic statistical feature matrix and the initial spatiotemporal feature vector to generate a joint representation vector that integrates explicit and implicit features. The system inputs the joint representation vector into a preset feature classification algorithm for classification and identification to determine the current traffic scene category. This preset feature classification algorithm is a classifier model that has been pre-trained offline and converged by those skilled in the art using a labeled sample set containing various historical traffic states through supervised learning.

[0059] By introducing a distortion detection mechanism before cross-dimensional feature stitching and using a feature isolation mask to suppress the weights of high-noise channels that experience cross-modal conflicts, the direct interference of front-end physical distortion data on the underlying implicit spatiotemporal features is blocked, thereby ensuring the robustness of the algorithm model in scene classification under non-steady-state environmental interference.

[0060] S106. Input the target spatiotemporal feature vector into the pre-trained timing parameter decoding network for mapping, dimensionality reduction and quantization calculation, and output the corresponding traffic signal timing strategy.

[0061] A pre-trained timing parameter decoding network refers to a machine learning structure used to convert high-dimensional abstract features into specific physical instructions. This network is pre-trained by those skilled in the art based on historical real-world manual timing strategies and corresponding traffic state feature samples, using a backpropagation algorithm to iteratively train the network model in an offline environment until the loss function converges. Mapping dimensionality reduction refers to the process of compressing a high-dimensional tensor into a low-dimensional scalar or short vector through mathematical transformation. Quantization calculation refers to the calculation steps of converting continuous floating-point variables into discrete engineering parameters according to physical device execution constraints. Traffic signal timing strategy refers to the control scheme used to send to the front-end intersection signal controller to directly drive the switching of traffic light states, including signal period, green ratio, and phase difference.

[0062] The system inputs the target spatiotemporal feature vector into the multi-layer fully connected layers of the timing parameter decoding network. Through matrix multiplication and activation calculations in the fully connected layers, the system maps and reduces the dimensionality of the target spatiotemporal feature vector into a continuous logical vector representing different control dimension parameters. Next, the system introduces physical safety constraint thresholds such as the minimum green light duration and maximum cycle limit at the intersection as boundary conditions. It then performs scaling and rounding quantization calculations on the continuous logical vector, converting it into specific integer values ​​for phase green light duration and phase difference. Finally, it packages and encapsulates these integrated parameters to output the traffic signal timing strategy.

[0063] In some embodiments, the output of traffic signal timing strategies can be implemented in several ways: Optionally, the target spatiotemporal feature vector is input into the periodic decoding branch and the green ratio decoding branch in the timing parameter decoding network; floating-point periodic values ​​and floating-point green ratio allocation vectors are output through linear activation units respectively; the floating-point values ​​are rounded and the travel time constraints are verified to reconstruct a traffic signal timing strategy containing specific durations; Optionally, a reinforcement learning-based action evaluation decoding network is constructed as the timing parameter decoding network; the target spatiotemporal feature vector is mapped to the timing adjustment amplitude action probability in the discrete action space; the adjustment amplitude corresponding to the maximum probability is selected to quantize and accumulate the baseline timing parameters of the previous period, and the final traffic signal timing strategy is output. It is understood that other methods can also be used to implement the output of traffic signal timing strategies, which are not limited here.

[0064] In other embodiments of this application, when a sudden local blockage occurs in the road network, feature distortion may result from the spread of abnormal data from the failed node along the graph network. Using the AI-based urban spatiotemporal feature mining method provided in this application, the node can be isolated, and a temporary graph edge can be established to reconstruct the road network according to the detour direction, extracting mutated features for compensation and correction.

[0065] like Figure 4The diagram shown is another flowchart illustrating the artificial intelligence-based urban spatiotemporal feature mining method provided in this application embodiment, including the following steps: S201. Obtain the spatiotemporal correlation data of the traffic network and road connectivity within the target area, and construct a spatiotemporal data matrix.

[0066] S202. Map intersections in the traffic network to graph nodes and road connectivity to graph edges to construct a dynamic road network graph. Dynamically determine the graph convolution receptive field of each graph node based on the connectivity degree and local clustering coefficient of each graph node.

[0067] S203. Based on the spatiotemporal data matrix and the receptive field of graph convolution, extract the spatial feature vector of the dynamic road network map, and perform time series calculation on the spatiotemporal data matrix to obtain the temporal feature vector.

[0068] In some embodiments, when performing feature extraction along the time dimension, a bidirectional network structure can be introduced to extract and fuse temporal features along both the positive and negative dimensions of the time axis, thereby comprehensively capturing the bidirectional fluctuation patterns of traffic flow.

[0069] For the extraction of temporal features from the spatiotemporal data matrix, the system inputs the spatiotemporal data matrix into a pre-defined bidirectional temporal neural network. This pre-defined bidirectional temporal neural network is a dual-stream computing architecture generated by those skilled in the art through pre-iterative training with a large number of offline traffic flow sequences, based on fundamental networks with temporal memory capabilities such as Long Short-Term Memory (LSTM). In specific computation, the system first drives the network along the forward computational dimension of the time series, reading the traffic flow data frame by frame in chronological order, and extracting the historical contextual features of the spatiotemporal data matrix using its internal hidden state memory mechanism. This forward computational operation captures the forward cumulative logic of traffic flow evolution from the past to the present, providing basic evolutionary trend support. Simultaneously, the system drives the network along the reverse computational dimension of the time series, inputting the matrix data in reverse order from the end to the beginning of the time window, extracting reverse temporal evolution features. Introducing the reverse dimension of deduction supplements the boundary constraints of subsequent data on preceding data that are easily missed in unidirectional computation, giving the model a broader contextual perspective. Subsequently, at the network end, the system performs cross-channel concatenation calculations on historical dependent context features and reverse temporal evolution features. By integrating two state tensors of opposite dimensions, a time feature vector representing the bidirectional fluctuation pattern of traffic flow is generated, thereby providing a more three-dimensional temporal mathematical representation for downstream processing.

[0070] By utilizing a bidirectional structure to extract sequence information in parallel along both positive and negative time dimensions, the shortcomings of traditional unidirectional temporal networks, such as information forgetting and limited perspective caused by unidirectional transmission, are overcome. The concatenation of positive historical accumulation and negative constraint features enables the output vector to possess a comprehensive dependency relationship of the global time window, thereby improving the representational ability of the underlying temporal features.

[0071] S204. Perform weighted fusion calculation on the spatial feature vector and the temporal feature vector to obtain the initial spatiotemporal feature vector.

[0072] S205. Based on the macroscopic statistical features of the spatiotemporal data matrix and the initial spatiotemporal feature vector, determine the current traffic scenario category, and dynamically adjust the feature weight elements in the weighted fusion calculation according to the traffic scenario category to recalculate the target spatiotemporal feature vector.

[0073] Steps S201-S205 and Figure 1 Steps S101-S105 in the illustrated embodiment are similar and can be found in the descriptions of steps S101-S105, which will not be repeated here.

[0074] S206. Extract the real-time micro-anomaly distribution features of traffic flow at intersections that have experienced abnormal changes from the spatiotemporal data matrix, and locate the topological failure node where severe traffic congestion and deadlock occur based on the real-time micro-anomaly distribution features.

[0075] An altered intersection refers to a physical intersection where the traffic flow pattern breaks the normal historical cycle; real-time microscopic anomaly distribution features are used to represent the irregular dispersion state of vehicles in the spatial location of the underlying layer, such as features like sudden braking aggregation and discontinuous lane changes; chaotic traffic backlog refers to the congestion and queuing phenomenon caused by vehicles in a disordered state; topological failure nodes refer to graph node units in the road network structure that have lost their normal traffic diversion capacity, resulting in traffic function paralysis.

[0076] An AI-based urban spatiotemporal feature mining system retrieves the spatiotemporal data matrix within the current time window and calculates the rate of change of vehicle density gradient between adjacent nodes along the spatial dimension. The system performs statistical analysis on microscopic vehicle speed and yaw angle indices in the matrix, extracting numerical combinations with local variance exceeding the normal range to construct real-time microscopic anomaly dispersion features. For example, when a traffic accident at an intersection causes the main lane to be occupied, vehicles may exhibit abnormal continuous lane changes, abnormal speed reductions, and overflow trajectories into adjacent auxiliary lanes. The system extracts numerical features such as an increase in the average yaw angle and a surge in negative acceleration as anomaly dispersion features. Based on these real-time microscopic anomaly dispersion features, the system uses a density-based spatial clustering algorithm to evaluate the physical spatial clustering degree of the anomaly features, identifies the physical coordinate mapping region with the highest clustering density, and locates the graph nodes corresponding to this region as topologically invalid nodes.

[0077] In some embodiments, the steps of extracting features and locating nodes from a spatiotemporal data matrix can be implemented in various ways: Optionally, extract the vehicle speed time series of each node in the spatiotemporal data matrix; calculate the vehicle speed decrease slope and number of stops for a single node within five consecutive time steps to form anomaly dispersion features; set a slope threshold to extract the coordinates of nodes exceeding the threshold, and output them as topology failure nodes; Optionally, read the second derivative information representing the difference in adjacent traffic flow density in the matrix to form anomaly dispersion features; apply a support vector machine to perform a binary classification operation on the dispersion features to define congestion anomaly class features; map the center point of the congestion anomaly class features to the road network base map, and mark the intersection closest to the center point as a topology failure node. It is understood that other methods can also be used to extract real-time microscopic anomaly dispersion features and locate topology failure nodes, which are not limited here.

[0078] S207. Perform topology isolation processing on the graph edge weights between the topology-invalidated node and its adjacent connected nodes in the dynamic road network graph, so as to perform a local graph topology disconnection operation in the dynamic road network graph.

[0079] Adjacent connected nodes refer to direct neighbor nodes in the adjacency matrix of the original dynamic road network graph that have non-zero weighted edges with topologically unreachable nodes; topological isolation processing refers to the process of forcibly adjusting mathematical variables representing physical connectivity attributes to represent unreachable states; graph topological disconnection operation is used to represent the structural change action of artificially cutting off the information interaction path between local nodes and the global network in the mathematical graph.

[0080] The system accesses the underlying adjacency matrix and weight feature matrix, which are essential for constructing the dynamic road network graph, in memory. Using the index of the failed node as a reference, the system traverses the rows and columns corresponding to that index in the adjacency matrix to find all adjacent connected nodes. The system then performs a zeroing operation on the graph edge weights of the located rows and columns involving adjacent connected nodes, overwriting the positive real numbers representing the weighted values ​​of road grade coefficients and historical traffic flow transmission rates with zero. By zeroing out the graph edge weights, the algorithm model will be unable to propagate or receive any feature tensors from the failed node during subsequent matrix multiplication operations, thus implementing the graph topology disconnection operation within the graph data structure.

[0081] In some embodiments, graph topology disconnection operations can be implemented in several ways: Optionally, retrieve all edge entities containing identifiers of topologically invalid nodes from the graph database; call the data update interface to change the connectivity boolean attribute of these edge entities from true to false; synchronously refresh the graph structure instance in memory to sever the topological connection; Optionally, create a unity mask matrix with the same dimensions as the original adjacency matrix; replace the elements in the rows and columns associated with the invalid nodes in the mask matrix with zero; perform a Hadamard product calculation on the original adjacency matrix and the mask matrix, and output a new adjacency matrix with the isolation processed. It is understood that other methods can also be used to implement graph topology disconnection operations, which are not limited here.

[0082] In extreme cases where a hazardous chemical transport vehicle leaks at a topological failure node, causing harmful gases to diffuse radially in all directions, a reverse impedance topological constraint mechanism can be implemented. The system acquires the diffusion direction and real-time diffusion radius of the harmful gas; calculates the distance difference between each adjacent connected node and the boundary of the pollution center; and constructs an impedance attenuation function based on these distance differences, assigning extreme negative attenuation values ​​to the graph edge weights from the failure node to the opposite direction of entry, rather than simply setting them to zero. By introducing a reverse repulsion field topological constraint mechanism, not only is information transmission at congested intersections disrupted, but the negative weights also guide the graph network to actively reject characteristic responses leading to the pollution zone during feature aggregation, improving the accuracy of urban traffic hub signal control strategies in non-steady-state environments.

[0083] S208. Based on the abnormal detour direction of the traffic flow characterized by the real-time micro-anomaly distribution features, determine the surrounding secondary diversion nodes that receive the abnormal detour traffic flow, and establish temporary graph edges between the adjacent nodes of the topology failure node and the surrounding secondary diversion nodes.

[0084] Abnormal detour direction of traffic flow refers to the vector orientation of the deviated route from the normal path and the target of transfer caused by the obstruction of the original path; the surrounding secondary diversion node refers to the intersection that is spatially close to the failure area and has a lower road network level, and begins to passively accept the overflow traffic flow; temporary map edge refers to the virtual connection dynamically added to the map to reflect the transient transfer path of traffic flow under abnormal conditions.

[0085] The system performs vector summation calculations on the microscopic vehicle displacement travel angles contained in the real-time microscopic anomaly dispersion features, extracting the direction vector representing the macroscopic transfer trend as the abnormal detour direction of the traffic flow. The system establishes a search space centered on the topological failure node, matching nodes along the vector cone angle of this abnormal detour direction. Low-level intersections that meet a preset straight-line distance threshold are marked as surrounding secondary diversion nodes; this preset straight-line distance threshold is a one-kilometer search range indicator set by those skilled in the art based on the average spacing constant of urban road grids. In the graph data structure, the system pairs the originally severed adjacent nodes of the topological failure node with the newly determined surrounding secondary diversion nodes, assigning new index positions to these two originally non-adjacent nodes in the adjacency matrix and logically connecting them, thereby establishing temporary graph edges in the mathematical model.

[0086] In some embodiments, the determination of surrounding secondary tributary nodes and the establishment of temporary graph edges can be achieved in several ways: Optionally, the actual deflection heading angle of individual vehicles with abnormal dispersion is extracted; the heading angle is clustered based on the K-Means algorithm to determine the mainstream detour angle; within this angle range, the three closest non-arterial road intersections are searched as secondary tributary nodes, and new edge objects are injected into the road network object set to establish temporary graph edges; Optionally, the abnormal detour origin and destination mapping records provided by the map navigation interface are obtained; the coordinate overlap between the records and the bottom-level nodes is compared to find the tributary node receiving the flow direction; the identifier "1" is assigned to the corresponding position of the adjacency tensor to achieve forced connection between the two nodes in the graph theory dimension. It is understood that other methods can also be used to implement the operation of establishing temporary graph edges, which are not limited here.

[0087] S209. Obtain the real-time traffic flow transfer load parameters corresponding to the temporary map edges, dynamically assign edge weights based on the real-time traffic flow transfer load parameters, and update the dynamic road network map to a transient reconstructed road network map.

[0088] Real-time traffic flow transfer load parameter refers to the proportion of abnormal traffic flow entering the physical road segment corresponding to the temporary graph edge within a specific time window; transient reconstructed road network graph is used to represent an abnormal mathematical road network model with strong timeliness formed during a disaster due to node disconnection and edge weight overwriting reconstruction.

[0089] The system uses monitoring equipment or trajectory mapping technology to count the number of standard vehicles equivalent per minute flowing into the new path along the abnormal detour direction of the traffic flow, and obtains the theoretical capacity parameter of the physical road segment indicated by the temporary graph edge. The system divides the number of standard vehicles equivalent per minute by the theoretical capacity parameter to calculate and output the real-time traffic flow transfer load parameter. The system substitutes this load parameter into a nonlinear normalization function for numerical scaling, and assigns the resulting smoothed value as the new edge weight to the temporary graph edge position in the adjacency matrix. By introducing weights representing the characteristics of unconventional congestion transfer, the original dynamic road network map undergoes a fundamental change in topological connectivity and information transmission strength, thereby completing the update to a transiently reconstructed road network map.

[0090] In some embodiments, the dynamic assignment of edge weights and the updating of the road network map can be achieved in several ways: Optionally, the traffic flow records captured by the checkpoint cameras installed at the entrances of the diversion sections are read; the traffic density within the time slice is statistically analyzed and used as a load transfer parameter; this parameter is logarithmically scaled and directly used as the edge weight, overwriting the original graph structure to form a transient reconstructed road network map; Optionally, real-time floating car GPS trajectories from ride-hailing platforms are received; the penetration rate of the number of trajectories in the entire road network is calculated, and the traffic load value on the entire temporary edge is inferred based on the penetration rate; the inferred value is input into the impedance formula derived based on queuing theory to obtain the edge weight tensor, and the network model is updated to generate the transient reconstructed road network map. It is understood that other methods can also be used to update the transient reconstructed road network map, which are not limited here.

[0091] In situations where a sudden, widespread power outage causes extensive disconnection and network loss for roadside front-end microwave radar and video surveillance cameras, making it impossible to directly obtain physical traffic flow data, the system can still utilize mobile operator cellular base station signaling drift data to infer load transfer. The system connects to the telecommunications operator's signaling gateway to extract the handover sequences of mobile terminal base stations located around isolated road network areas; calculates the group frequency drift of terminals switching from main road base stations to secondary diversion road segment base stations within a specific time window; and uses historical road network vehicle-passenger ratio coefficients to convert this signaling drift into a pseudo-traffic flow load transfer matrix and assign it temporary graph edges. By introducing cross-modal load inference using signaling data, the system solves the problem of algorithm feature interruption caused by the paralysis of front-end visual perception equipment, improving the accuracy of urban traffic hub signal control strategies in non-steady-state environments.

[0092] In some embodiments, when faced with sudden abnormal scenarios and adverse weather conditions that impair the road network's traffic capacity, environmental disaster assessment can be introduced and combined with a topology pruning mechanism to dynamically constrain the graph edge weights, preventing traffic flow from being incorrectly guided to high-risk road sections by the algorithm.

[0093] Regarding the dynamic road network map update process, when a sudden nonlinear variation scenario is detected and the real-time meteorological monitoring values ​​exceed a preset disaster warning threshold, the system acquires spatial physical environment data and real-time meteorological evolution sequences for the surrounding secondary diversion nodes and the corresponding areas of temporary map edges. The preset disaster warning threshold is a physical quantity critical value pre-set by those skilled in the art through statistical analysis based on historical meteorological disaster records and road paralysis standards. This operation introduces objective physical environment prior constraints into the abstract network, enabling the model to possess meteorological perception capabilities. Subsequently, the system calculates traffic loss due to environmental disasters based on the spatial physical environment data and real-time meteorological evolution sequences. Specifically, the system obtains predicted values ​​of rainfall or water depth evolution by calling the API interface of the urban meteorological monitoring platform, combines this with the low-lying elevation attributes of physical road sections, and calculates the dynamic traffic loss rate of the physical road section corresponding to the temporary map edge using a preset water accumulation-vehicle speed attenuation model. The system further constructs a traffic constraint penalty mechanism based on the dynamic traffic loss rate, calculating the environmental impedance coefficient for the temporary map edge using methods such as an inverse proportional attenuation function; this calculation transforms abstract meteorological hazards into quantifiable traffic attenuation indicators. Next, the system uses the environmental impedance coefficient to perform attenuation correction calculations on the real-time traffic flow transfer load parameters, obtaining restricted transfer load parameters that incorporate environmental constraints. This correction process, by reducing the weight of affected road sections, reflects the physical blocking effect of severe weather at the data level. The system compares the restricted transfer load parameters with a set topological connectivity validity threshold; this threshold is a weight lower limit constant determined based on the minimum safe passage requirements of roads. If the restricted transfer load parameter is lower than this topological connectivity validity threshold, it indicates that the corresponding road section no longer has diversion conditions, and the system performs secondary topological pruning on the temporary graph edge, removing the connection in the graph structure to cut off computational power transmission; if the restricted transfer load parameter is not lower than the topological connectivity validity threshold, the restricted transfer load parameter is assigned as the final weight to the temporary graph edge, completing the transient reconstruction of the road network map update operation.

[0094] By introducing meteorological evolution calculations and environmental impedance penalties, physical disaster constraints are transformed into numerical attenuation of graph edge weights. Secondary topology pruning, combined with quantization thresholds, eliminates logical connections leading to high-risk road sections, preventing the model from blindly guiding traffic flow to disaster-prone areas and ensuring the security, compliance, and practical rationality of the underlying network reconstruction.

[0095] S210. Substitute the updated transient reconstructed road network map back into the graph convolutional network to perform secondary spatial feature extraction, and obtain the abrupt spatial feature vector that reflects the extreme detour diffusion law of traffic flow.

[0096] Secondary spatial feature extraction refers to the process of performing a supplementary operation of graph feature aggregation again using a graph structure with new topological parameters, in addition to the basic feature extraction operation; extreme detour shock wave law is used to represent the physical phenomenon of traffic flow surging into and spreading to secondary roads after encountering a breakpoint and forming a back propagation of queuing oscillations; variant spatial feature vector refers to a low-dimensional mathematical spatial representation vector containing abnormal congestion state captured by a new topological structure.

[0097] The system saves previously extracted spatial features and retrieves a pre-built spatial graph convolutional neural network model instance from S203. The system inputs the transient reconstructed road network data (updated with updated bottom-level adjacency matrices and weight parameters) and related node features into the input layer of the graph convolutional network. The graph convolutional network executes a message-passing mechanism according to the temporary graph edge paths in the transient reconstructed road network, aggregating the features of locally clustered nodes layer by layer along the new connectivity relationships to secondary nodes. After matrix multiplication operations in the forward propagation formula and dimensionality compression in the graph pooling layer, the network outputs a variant spatial feature vector representing the current irregular spatial oscillation distribution.

[0098] In some embodiments, secondary spatial feature extraction can be implemented in several ways: Optionally, the updated angle matrix and reconstructed adjacency matrix are extracted; the reconstructed adjacency matrix is ​​symmetrically normalized using the degree matrix according to the forward propagation formula of the graph convolutional network layer; the processed matrix is ​​multiplied by the node feature matrix, and the variant spatial feature vector is output after nonlinear activation; Optionally, a deep graph attention network with a skip connection mechanism is constructed; the transient reconstructed road network graph is fed into the network to calculate the attention allocation coefficients between temporary graph edge connection nodes; and the variant spatial feature vector representing the detour pattern is extracted by weighted summation and aggregation of neighborhood information. It is understood that other methods can also be used to implement the operation of extracting variant spatial feature vectors, which are not limited here.

[0099] S211. The variation space feature vector is used as a compensation term and fused with the initial spatiotemporal feature vector to obtain a deep-corrected spatiotemporal feature vector for sudden nonlinear variation scenarios. The target spatiotemporal feature vector is then replaced with the deep-corrected spatiotemporal feature vector and input into the timing parameter decoding network.

[0100] Compensation terms refer to supplementary calculation factors introduced to correct errors when the basic mathematical model deviates due to constraints; feature fusion refers to the operation of merging data representations from multiple sources into a single high-dimensional comprehensive representation through splicing or combination; deep-corrected spatiotemporal feature vectors refer to robust panoramic traffic mathematical representations generated after absorbing the variables of detour impact and disaster disconnection on the normal spatiotemporal feature base.

[0101] The system extracts the variability space feature vector and the initial spatiotemporal feature vector obtained in S104 within the tensor operation space. The system assigns a compensation weight scalar to this variability space feature vector, and performs a bitwise addition operation on the feature channel dimension between the variability space feature vector multiplied by the scalar and the initial spatiotemporal feature vector, integrating the shock wave diffusion properties contained therein into the original spatiotemporal base to calculate and generate a depth-corrected spatiotemporal feature vector. To ensure the continuity of data flow in subsequent processes, the system overwrites the pointer address of this depth-corrected spatiotemporal feature vector to the original target spatiotemporal feature vector object at the memory reference level, allowing subsequent steps to directly read and process the new vector containing the correction information.

[0102] In some embodiments, feature fusion and replacement can be implemented in various ways: Optionally, a feature fusion module including a multilayer perceptron is constructed; the mutation space feature vector and the initial spatiotemporal feature vector are concatenated into a long vector along the last dimension of the tensor; a depth-corrected spatiotemporal feature vector is mapped through dimensionality reduction calculation of the perceptron layer, and variable name reassignment is performed in the algorithm pipeline to replace the target spatiotemporal feature vector; Optionally, the root mean square norm of the mutation space feature vector is calculated to adaptively generate a dynamic adjustment ratio; mutation feature elements are proportionally added to the initial spatiotemporal feature vector according to this ratio to generate a depth-corrected spatiotemporal feature vector; when the data is passed to the timing parameter decoding network interface, a cached instance of the corrected vector is directly passed to complete the functional physical replacement. It is understood that other methods can also be used to implement the feature fusion and replacement steps, which are not limited here.

[0103] S212. Input the target spatiotemporal feature vector into the pre-trained timing parameter decoding network for mapping, dimensionality reduction and quantization calculation, and output the corresponding traffic signal timing strategy.

[0104] Step S212 and Figure 1 Step S106 in the illustrated embodiment is similar and can be found in the description of step S106, which will not be repeated here.

[0105] In the above embodiments, by severing the transmission of abnormal information from chaotic congestion sources to the global adjacency network at the graph structure level, the extracted spatiotemporal features are prevented from being offset by locally paralyzed data. The reconstructed transient topology-guided feature aggregation process follows the physical diffusion shock wave law of traffic flow under blocking conditions, endowing the algorithm model with dynamic adaptive capability in the face of sudden physical blockages, and improving the robustness of the signal control strategy generated in non-steady-state traffic environments.

[0106] The following describes an exemplary AI-based urban spatiotemporal feature mining system 300 provided in an embodiment of this application. Figure 5 This is an exemplary hardware structure diagram of an artificial intelligence-based urban spatiotemporal feature mining system 300 provided in this application embodiment.

[0107] In some embodiments, the AI-based urban spatiotemporal feature mining system 300 is a computer device or includes a computer device. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores data. The network interface of the computer device is used to communicate with other external terminals or servers via a network connection. In some embodiments, the network interface can be a wired network interface; in some embodiments, the network interface can also be a wireless network interface. When the computer program is executed by the processor, it implements the methods in the embodiments of this application.

[0108] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0109] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0110] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0111] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can 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 can 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 can 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.

[0112] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for mining urban spatiotemporal features based on artificial intelligence, characterized in that, include: Acquire spatiotemporal correlation data of the traffic network and road connectivity within the target area, and construct a spatiotemporal data matrix; The intersections in the traffic network are mapped as graph nodes, and the road connectivity is mapped as graph edges to construct a dynamic road network graph. The graph convolution receptive field of each graph node is dynamically determined based on the connectivity degree and local clustering coefficient of each graph node. Based on the spatiotemporal data matrix and the receptive field range of the graph convolution, the spatial feature vector of the dynamic road network map is extracted, and the temporal feature vector is obtained by performing time-series calculation on the spatiotemporal data matrix. The spatial feature vector and the temporal feature vector are weighted and fused to obtain an initial spatiotemporal feature vector. Based on the macroscopic statistical characteristics of the spatiotemporal data matrix and the initial spatiotemporal feature vector, the current traffic scenario category is determined, and the feature weight elements in the weighted fusion calculation are dynamically adjusted according to the traffic scenario category to recalculate the target spatiotemporal feature vector. The target spatiotemporal feature vector is input into a pre-trained timing parameter decoding network for mapping, dimensionality reduction, and quantization calculation, and the corresponding traffic signal timing strategy is output.

2. The method according to claim 1, characterized in that, Before inputting the target spatiotemporal feature vector into a pre-trained timing parameter decoding network for mapping, dimensionality reduction, and quantization calculations, the process also includes: The real-time micro-anomaly distribution features of traffic flow at intersections that have experienced abnormal changes are extracted from the spatiotemporal data matrix, and the topology failure nodes that have caused severe traffic congestion and deadlock are located based on the real-time micro-anomaly distribution features. Topology isolation processing is performed on the graph edge weights between the topology-invalidated node and its adjacent connected nodes in the dynamic road network graph, so as to perform a local graph topology disconnection operation in the dynamic road network graph. Based on the abnormal detour direction of the traffic flow characterized by the real-time micro-anomaly dispersion features, the surrounding secondary diversion nodes that receive the abnormal detour traffic flow are determined, and temporary graph edges are established between the adjacent nodes of the topology failure node and the surrounding secondary diversion nodes. Obtain the real-time traffic flow transfer load parameters corresponding to the temporary graph edges, dynamically assign edge weights based on the real-time traffic flow transfer load parameters, and update the dynamic road network map into a transient reconstructed road network map; The updated transient reconstructed road network map is then substituted back into the graph convolutional network to perform secondary spatial feature extraction, resulting in a mutation spatial feature vector that reflects the extreme detour diffusion pattern of traffic flow. The mutation space feature vector is used as a compensation term and fused with the initial spatiotemporal feature vector to obtain a deep-corrected spatiotemporal feature vector for sudden nonlinear mutation scenarios. The target spatiotemporal feature vector is then replaced with the deep-corrected spatiotemporal feature vector and input into the timing parameter decoding network.

3. The method according to claim 2, characterized in that, The step of dynamically assigning edge weights based on the real-time traffic flow transfer load parameters and updating the dynamic road network map to a transient reconstructed road network map specifically includes: When the real-time meteorological monitoring values ​​associated with the sudden nonlinear variation scenario are identified as exceeding the preset disaster warning threshold, the spatial physical environment data and real-time meteorological evolution sequence of the surrounding secondary diversion nodes and the corresponding areas of the temporary map edges are obtained. Based on the spatial physical environment data and the real-time meteorological evolution sequence, the traffic loss due to environmental disasters is calculated, and the dynamic traffic loss rate of the physical road segment corresponding to the temporary map edge is obtained. A passage constraint penalty mechanism is constructed based on the dynamic passage loss rate, and the environmental impedance coefficient for the temporary graph edge is calculated. The environmental impedance coefficient is used to attenuate and correct the real-time traffic flow transfer load parameter to obtain a constrained transfer load parameter that incorporates environmental constraints. The restricted transfer load parameter is compared with the set topology connectivity validity threshold. If the restricted transfer load parameter is lower than the topology connectivity validity threshold, then secondary topology pruning is performed on the temporary graph edges. If the restricted transfer load parameter is not lower than the topological connectivity validity threshold, then the restricted transfer load parameter is assigned as the final weight to the temporary graph edge, thus completing the update of the transient reconstructed road network graph.

4. The method according to claim 1, characterized in that, The determination of the current traffic scenario category based on the macroscopic statistical features of the spatiotemporal data matrix and the initial spatiotemporal feature vector specifically includes: Low-dimensional macroscopic traffic and environmental statistical elements are extracted from the spatiotemporal data matrix to obtain a macroscopic statistical feature matrix representing the explicit physical state as the macroscopic statistical features. The macroscopic statistical feature matrix and the initial spatiotemporal feature vector are concatenated across the feature channels to generate a joint representation vector that integrates explicit and implicit features. The joint representation vector is input into a preset feature classification algorithm for classification and recognition to determine the current traffic scene category.

5. The method according to claim 4, characterized in that, The step of concatenating the macroscopic statistical feature matrix with the initial spatiotemporal feature vector across dimensional feature channels to generate a joint representation vector that integrates explicit and implicit features specifically includes: Extract the multidimensional spatiotemporal fluctuation evaluation parameters of the macroscopic statistical feature matrix; If the multidimensional spatiotemporal fluctuation assessment parameter exceeds the set modal confidence safety threshold, it is determined that the current explicit physical data is subject to modal distortion due to environmental interference. Calculate the feature space mapping difference between the macroscopic statistical feature matrix where the modal distortion occurs and the initial spatiotemporal feature vector to quantify the degree of cross-modal feature conflict between explicit and implicit features; Based on the degree of cross-modal feature conflict, a dynamic adaptive feature isolation mask is constructed, and the feature isolation mask is used to adaptively suppress the high-noise conflict channels in the macro-statistical feature matrix. The macroscopic statistical feature matrix after weight suppression is concatenated with the initial spatiotemporal feature vector through cross-dimensional feature channels to generate the joint representation vector.

6. The method according to claim 1, characterized in that, The step of mapping intersections in the traffic network as graph nodes and mapping road connectivity relationships as graph edges to construct a dynamic road network graph specifically includes: Obtain the road grade coefficient and historical traffic flow transmission rate corresponding to the road connectivity relationship; The road grade coefficient and the historical traffic flow transmission rate are weighted and summed according to a preset ratio to obtain the weight of the graph edges, so as to complete the construction of the dynamic road network map.

7. The method according to claim 1, characterized in that, The step of performing time-series calculations on the spatiotemporal data matrix to obtain the time feature vector specifically includes: The spatiotemporal data matrix is ​​input into a preset bidirectional temporal neural network; Historical dependency context features of the spatiotemporal data matrix are extracted along the forward computation dimension of the time series, and reverse temporal evolution features are extracted along the reverse computation dimension of the time series. The historical dependent context features and the reverse time-series evolution features are concatenated and calculated to obtain the time feature vector representing the bidirectional fluctuation pattern of traffic flow.

8. An artificial intelligence-based urban spatiotemporal feature mining system, characterized in that, The AI-based urban spatiotemporal feature mining system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to cause the AI-based urban spatiotemporal feature mining system to perform the method as described in any one of claims 1-7.

9. A computer program product containing instructions, characterized in that, When the computer program product is run on the AI-based urban spatiotemporal feature mining system, the AI-based urban spatiotemporal feature mining system performs the method as described in any one of claims 1-7.

10. A computer-readable storage medium comprising instructions, characterized in that, When the instructions are run on an AI-based urban spatiotemporal feature mining system, the AI-based urban spatiotemporal feature mining system performs the method as described in any one of claims 1-7.

Images