An urban commuting traffic situation intelligent early warning method based on multi-source data fusion

Abstract

The application relates to the technical field of traffic situation early warning, and particularly discloses an urban commuting traffic situation intelligent early warning method based on multi-source data fusion. The application constructs a road network space quantization grid, respectively establishes a physical situation field and a semantic situation field, forms a grid unit group according to preset division attributes and levels, calculates the physical situation deviation degree and the semantic situation convergence degree of each grid unit in the group by taking other grid units in the group as references, takes the difference between the two as a field situation cooperative deviation, generates a field situation cooperative deviation field based on a gradient, and inputs the field situation cooperative deviation field into a situation early warning discriminator to obtain an early warning result. The application realizes deep cooperation of physical perception data and semantic description data, and improves the accuracy and real-time performance of urban commuting traffic situation early warning.

Description

Technical Field

[0001] This invention relates to the field of traffic situation early warning technology, and in particular to an intelligent early warning method for urban commuter traffic situation based on multi-source data fusion. Background Technology

[0002] With the increasing severity of urban traffic congestion, traffic situation warning during peak commuting hours has become one of the core requirements of intelligent transportation systems. Existing traffic warning technologies are typically based on multi-source data fusion methods. This involves collecting physical sensing data such as speed and flow rate from the road network, combining it with semantic descriptions obtained through image recognition or text parsing, and then using feature stitching or model fusion to comprehensively assess traffic conditions. For example, Chinese invention patent CN120032522B, entitled "Traffic Incident Violation Warning System Based on AI and Multi-Source Data Fusion," uses timestamp alignment and difference analysis between multi-camera video and external traffic data to achieve early warning of traffic violations. Another Chinese invention patent CN121483084A, entitled "AI-Based Multi-Source Heterogeneous Traffic Big Data Accident Warning and Intervention System," uses edge computing and LSTM prediction models to extract multi-dimensional features such as traffic flow speed variance to dynamically assess accident risk probability. These solutions, to a certain extent, achieve comprehensive utilization of multi-source data, providing a technological foundation for traffic situation warning.

[0003] However, existing technologies generally splice or fuse physical perception data and semantic description data in the same feature space before inputting them into the prediction model, ignoring the fundamental differences in data form and information hierarchy between physical data (continuous numerical type) and semantic data (discrete symbolic type). This results in the heterogeneity of the two types of data being weakened or lost during the fusion process. Furthermore, existing methods often rely on globally fixed thresholds or simple partitioning for early warning, lacking an adaptive reference system based on horizontal comparisons within the same attribute group. This makes it difficult to accurately identify abnormal situational changes in similar road network areas, leading to insufficient accuracy and targeting of early warnings.

[0004] Therefore, this invention proposes an intelligent early warning method for urban commuter traffic situation based on multi-source data fusion. Summary of the Invention

[0005] This invention provides an intelligent early warning method for urban commuting traffic situation based on multi-source data fusion. By achieving deep collaboration between physical and semantic data through dual-field separation representation and horizontal comparison within groups, it overcomes the shortcomings of information loss in heterogeneous data fusion and lack of adaptability in fixed threshold discrimination in existing technologies, and significantly improves the accuracy and adaptability of traffic situation early warning.

[0006] This invention provides an intelligent early warning method for urban commuter traffic conditions based on multi-source data fusion, comprising: Acquire multi-source sensing data from the urban commuter road network. The multi-source sensing data includes physical sensing data and semantic description data. Construct a spatial quantization grid for the road network, map physical sensing data to each grid cell and extract physical state parameters, and construct a physical situation field based on the physical state parameters of each grid cell. The physical situation field is a continuous numerical vector field. Semantic event types are extracted from semantic description data, and a semantic situation field is constructed based on the semantic event types. The semantic situation field is a discrete symbol field. All grid units are divided according to the preset grid unit division attributes and division levels to obtain grid unit groups under each division attribute and division level; the preset grid unit division attributes include road level attributes, functional area type attributes, and historical congestion level attributes. Based on the physical situation field and the semantic situation field, within each grid cell group, with other grid cells in the group as references, the physical situation deviation and semantic situation convergence of each grid cell are calculated respectively. The difference between the physical state deviation and the semantic state convergence of each grid cell is used as the field state coordination deviation of each grid cell. The field state coordination deviation field is generated based on the gradient of the field state coordination deviation on the road network spatial quantization grid. The field situational coordination deviation field is used as the current situational awareness data of the road network space and input into the situational warning discriminator to obtain the warning result.

[0007] Furthermore, before dividing all grid cells according to the preset grid cell division attributes and division hierarchy, an adaptive construction step of the group reference system is also included: Obtain the time series sequence of physical state parameters and semantic event type of each grid cell in the road network spatial quantization grid within the historical time window; The physical state parameters of each grid cell are used to extract the physical situation time-series feature vector of each grid cell. The physical situation time-series feature vector contains the time sequence of the average velocity, time occupancy rate, spatial density and velocity variance of each grid cell at each historical moment. Based on the semantic event type time series of each grid cell, the semantic situation frequency feature vector of each grid cell is statistically analyzed. The semantic situation frequency feature vector contains the normalized value of the occurrence frequency of each semantic event type in each grid cell within the historical time window. Density-based clustering algorithm is used to cluster the time series feature vectors of each physical situation to obtain multiple initial physical groups. Each initial physical group is identified by the group center vector of the time series feature vector of the physical situation. A density-based clustering algorithm is used to cluster the semantic situation frequency feature vectors to obtain multiple initial semantic groups. Each initial semantic group is identified by the group center vector of the semantic situation frequency feature vector. The spatial intersection grid cells of the initial physical group and the initial semantic group are taken as hybrid group cells. The group center vector of the physical situation temporal feature vector and the group center vector of the semantic situation frequency feature vector in the hybrid group cell are concatenated to obtain the hybrid group center vector. The hybrid group center vector is used as the initial reference benchmark for determining the group affiliation of grid cells under the preset partitioning attributes and partitioning levels. Based on the similarity between the physical situation temporal feature vector and the semantic situation frequency feature vector of each grid cell and the hybrid group center vector, the initial group affiliation of each grid cell is completed.

[0008] Furthermore, the physical state deviation of each grid cell within each grid cell group is calculated, including: Obtain the physical state feature vector sample set of all grid cells in the grid cell group under the current time window. Each physical state feature vector contains four dimensions: average velocity, time occupancy rate, spatial density, and velocity variance. The covariance matrix and mean vector of the physical state feature vector sample set are calculated to obtain the covariance matrix and mean vector of the physical state feature vectors of all grid cells in the grid cell group. Calculate the Mahalanobis distance between the physical state feature vector and the mean vector of each grid cell, and take the ratio of the Mahalanobis distance to the maximum value of the Mahalanobis distance of all grid cells in the grid cell group as the physical state deviation of each grid cell.

[0009] Furthermore, the semantic situation convergence degree of each grid cell within each grid cell group is calculated, including: Obtain the semantic event type encoding sequence of all grid cells within the corresponding grid cell group under the current time window; The frequency of occurrence of each semantic event type encoding value within the grid cell group is statistically analyzed. A frequency distribution vector is constructed using the frequency of occurrence of all semantic event type encoding values. The frequency distribution vector is used as the baseline probability value of each semantic event type encoding value. Calculate the cross-entropy between the semantic symbol vector and the baseline probability value of each grid cell, normalize the cross-entropy by the maximum and minimum values ​​within the grid cell group, and then invert it to obtain the semantic situation convergence degree of each grid cell.

[0010] Furthermore, after constructing the physical state field, the process also includes the step of extracting the local field gradient vector from the physical state field: The first-order partial derivatives of the physical state parameters of each grid cell are calculated in the row and column directions of the road network spatial quantization grid to obtain the row and column gradient components of each grid cell in each physical state parameter dimension. The row and column gradient components of each grid cell in each physical state parameter dimension are vectorized to obtain the local field gradient vector of each grid cell. Using the field situational coordination deviation field as the current situational awareness data of the road network space, and combining it with the local field gradient vector, the data are input into the situational warning discriminator to obtain better warning results.

[0011] Furthermore, the field situation coordination deviation field is generated in the following way: The field situation coordination deviation of each grid cell is used as a scalar value; The gradient components of the scalar values ​​of each grid cell in the row and column directions of the quantized grid in the road network space are used as the field vectors of the field-situational coordination deviation field in the corresponding grid cell to obtain the field-situational coordination deviation field.

[0012] Furthermore, it also includes field line tracking and anomaly region geometric calibration steps: The field vector of the field situation coordination deviation field is traced along the field vector direction at each grid cell to obtain multiple field lines of the field situation coordination deviation field. Based on the curvature of each field line and the convergence between field lines, the abnormal convergence region and abnormal divergence region of the field situation coordination deviation field are identified. The calibration results of the anomaly convergence region and the anomaly divergence region are superimposed with the magnitude of the field vector at each grid cell of the field-situation coordination deviation field to generate an enhanced field-situation coordination deviation field containing the region boundary markers. The enhanced field situational coordination deviation field input situational early warning discriminator is used to obtain enhanced early warning results.

[0013] Furthermore, the field vector of the field-state cooperative deviation field is traced along the field vector direction at each grid cell to obtain multiple field lines of the field-state cooperative deviation field, including: Starting from the boundary grid cells of the road network spatial quantization grid, the field vector direction of each boundary grid cell is used as the initial tracking direction, and the position coordinates of the next tracking point are calculated sequentially with a preset step size; The field vector direction at each tracking point is obtained by bilinear interpolation, and the corresponding direction is used as the tracking direction for the next tracking step. At each tracking point location, the field vector divergence value at the corresponding point is calculated synchronously. When the sign of the field vector divergence value of consecutive tracking points changes or the tracking point exceeds the boundary of the road network spatial quantization grid, the current tracking is terminated, and all tracking points on the current tracking path are connected in sequence to form a single field line. The tracking process is repeated for all boundary grid cells to obtain multiple field lines of the field state cooperative deviation field.

[0014] Furthermore, it also includes a time-domain adaptive calibration step for the group statistical reference baseline within the current time window: Obtain the physical and semantic situation fields of each grid cell in the road network spatial quantization grid within the same historical time window; Based on the physical and semantic situation fields of each grid cell within the historical time window, the mean physical situation parameters and the probability of occurrence of semantic event types of each grid cell in the historical period are calculated. The historical baseline physical and semantic situation fields are constructed using the mean physical situation parameters and the probability of occurrence of semantic event types of each grid cell. Based on the historical contemporary benchmark physical situation field and the historical contemporary benchmark semantic situation field, the situation coordination deviation of the historical contemporary benchmark field is calculated. The physical situation field of each grid cell in the current time window is compared with the historical benchmark physical situation field in the same period by performing a grid-by-grid difference calculation to obtain the physical situation time domain offset of each grid cell. The semantic situation field of each grid cell in the current time window is compared with the historical benchmark semantic situation field by performing grid-by-grid relative entropy calculation to obtain the temporal offset of the semantic situation of each grid cell. The physical state time-domain offset and semantic state time-domain offset of each grid cell are concatenated to obtain the time-domain offset vector of each grid cell. The time-domain offset vector of each grid cell is divided by the historical standard deviation of the corresponding dimension to obtain the standardized time-domain offset vector of each grid cell. The magnitude of the standardized time-domain offset vector is used as the time-domain offset degree of each grid cell. Using the temporal offset of each grid cell as the update weight of the group statistical reference baseline of the grid cell group within the current time window, the calculation reference values ​​of the physical situation deviation and the semantic situation convergence within the current time window are weighted and updated to obtain the group statistical reference baseline after temporal calibration. Using the time-domain calibrated group statistical reference baseline, the physical state deviation and semantic state convergence of each grid cell within the current time window are corrected, and the corrected physical state deviation and semantic state convergence are used for the field state coordination deviation calculation in the current time window. Meanwhile, the time-domain calibrated group statistical reference baseline is used as the calculation reference value for the next time window, which is used to calculate the physical situation deviation and semantic situation convergence in the next time window.

[0015] Furthermore, it also includes steps for dynamic deviation index calculation, adaptive dynamic threshold calibration, and field vector weighted modulation: The difference between the field situation coordination deviation of each grid cell and the historical reference field situation coordination deviation of the same grid cell is calculated to obtain the reference deviation of each grid cell. The historical reference field situation coordination deviation is calculated based on the average physical situation deviation and the average semantic situation convergence of each grid cell within the historical time window. The absolute value of the reference deviation of each grid cell is multiplied by the time-domain offset to obtain the field situation dynamic deviation index of each grid cell. Based on the spatial distribution of the dynamic deviation index of the field situation in each grid cell on the road network spatial quantization grid, a dynamic deviation index field is constructed. Calculate the mean and standard deviation of the field state dynamic deviation index for all grid cells, and use the product of the mean plus the standard deviation and the preset multiplication factor as the dynamic threshold. The ratio of the field situation dynamic deviation index to the dynamic threshold of each grid cell is used as the actual weight coefficient of each grid cell. The field vector of the field situation cooperative deviation field is weighted and modulated using the actual weight coefficient to obtain the weighted and modulated field situation cooperative deviation field. The weighted modulated field situation collaborative deviation field is input into the situation warning discriminator to obtain the weighted warning result.

[0016] The beneficial effects of this invention compared to existing technologies are as follows: Existing traffic situation warning technologies simply splice and fuse physical perception data (continuous numerical type) and semantic description data (discrete symbolic type) in the same feature space, ignoring the essential differences in information form between the two data types, resulting in the feature representation after fusion failing to fully retain their respective heterogeneous information; at the same time, the warning judgment of existing methods mostly rely on globally fixed thresholds and lack an adaptive reference system based on horizontal comparison within the same attribute group, making it difficult to accurately identify situational anomalies in similar road network areas. This invention constructs a dual-field separation representation architecture for physical and semantic situation fields, maintaining the independent form of the two types of data. It then performs horizontal comparisons within grid cell groups divided by road level, functional zone type, and historical congestion level, calculating the physical situation deviation and semantic situation convergence of each grid cell. The difference between the two is used as the field situation coordination deviation, and a coordination deviation field is generated based on gradients. Finally, this field is input into the early warning discriminator, achieving deep coordination between physical perception data and semantic description data. This eliminates the loss of fusion information caused by differences in data form, enabling early warning discriminator to adapt to dynamic reference benchmarks of different groups, thus improving the accuracy and relevance of urban commuter traffic situation early warning.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the intelligent early warning method for urban commuter traffic situation based on multi-source data fusion in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the adaptive construction process of the group reference system in this embodiment of the invention. Figure 3 This is a flowchart of the time-domain adaptive calibration and dynamic deviation exponential weighted modulation in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] like Figure 1 As shown, this invention provides an embodiment of an intelligent early warning method for urban commuter traffic conditions based on multi-source data fusion, comprising: Acquire multi-source sensing data from the urban commuter road network. The multi-source sensing data includes physical sensing data and semantic description data. Construct a spatial quantization grid for the road network, map physical sensing data to each grid cell and extract physical state parameters, and construct a physical situation field based on the physical state parameters of each grid cell. The physical situation field is a continuous numerical vector field. Semantic event types are extracted from semantic description data, and a semantic situation field is constructed based on the semantic event types. The semantic situation field is a discrete symbol field. All grid units are divided according to the preset grid unit division attributes and division levels to obtain grid unit groups under each division attribute and division level; the preset grid unit division attributes include road level attributes, functional area type attributes, and historical congestion level attributes. Based on the physical situation field and the semantic situation field, within each grid cell group, with other grid cells in the group as references, the physical situation deviation and semantic situation convergence of each grid cell are calculated respectively. The difference between the physical state deviation and the semantic state convergence of each grid cell is used as the field state coordination deviation of each grid cell. The field state coordination deviation field is generated based on the gradient of the field state coordination deviation on the road network spatial quantization grid. The field situational coordination deviation field is used as the current situational awareness data of the road network space and input into the situational warning discriminator to obtain the warning result.

[0022] In this embodiment, the urban commuter road network refers to the road network within an urban area that is primarily driven by commuter travel demand. It includes roads of different levels, such as expressways, arterial roads, secondary arterial roads, and local roads, and the road network topology formed by their interconnections. This road network covers the main travel routes of the urban commuter population.

[0023] In this embodiment, multi-source sensing data from the urban commuter road network is acquired. The multi-source sensing data includes physical sensing data and semantic description data. This refers to the real-time collection or reception of road traffic status information from different data sources through various sensing devices, communication terminals, and information release platforms deployed in the urban commuter road network, providing a data foundation for the subsequent construction of the physical and semantic situation fields.

[0024] In this embodiment, physical perception data refers to time-series speed data collected by fixed detectors, trajectory point data collected by floating cars, road section cross-sectional flow data collected by radar, and lane occupancy rate data collected by roadside units. These data represent the physical traffic state of the road network in the form of numerical time series. Semantic description data refers to traffic incident text reports, traffic control announcements, internet traffic condition user reports, and structured traffic incident information extracted from the above texts through natural language processing technology. These data represent non-physical quantitative traffic incidents occurring in the road network in the form of discrete symbols.

[0025] In this embodiment, constructing a road network spatial quantization grid refers to dividing the urban commuter road network into a uniformly distributed two-dimensional grid cell array according to a preset spatial resolution. Each grid cell corresponds to a fixed geographical area in the road network in space. The grid covers the entire target road network range and is used to unify the dispersed multi-source sensing data into the same spatial reference system for subsequent processing.

[0026] In this embodiment, mapping physical sensing data to each grid cell and extracting physical state parameters involves: first, slicing the physical sensing data according to a preset time window; removing outliers and imputing missing values ​​in the speed and flow data within each time window slice; then projecting the imputed physical sensing data onto the corresponding grid cell in the road network spatial quantization grid using a road network matching algorithm, so that each grid cell obtains physical sensing data that matches its spatial location; and then extracting the values ​​of four dimensions—average speed, time occupancy rate, spatial density, and speed variance—from the physical sensing data of each grid cell as the physical state parameters of that grid cell.

[0027] In this embodiment, a physical state field is constructed using the physical state parameters of each grid cell. The physical state field is a continuous numerical vector field, which means that the physical state parameters of all grid cells in the road network spatial quantization grid are organized into a multi-dimensional numerical matrix according to the spatial arrangement of the grid. Each element in the matrix corresponds to a physical state parameter vector at the location of a grid cell. The matrix as a whole constitutes a continuous numerical vector field defined on the road network spatial quantization grid. The vector at the location of each grid cell consists of four components: the average velocity, time occupancy rate, spatial density, and velocity variance of the grid cell. Furthermore, the physical state parameter values ​​between adjacent grid cells change continuously in space.

[0028] In this embodiment, extracting the spatial variation characteristics of the physical situation field refers to calculating the physical state parameters of each grid cell in the road network spatial quantization grid, calculating the rate of change of physical state parameter values ​​between adjacent grid cells in the row and column directions respectively, and obtaining a composite vector at each grid cell by vector synthesis of the rate of change in the row and column directions. This composite vector is used to quantify the degree of spatial variation of the physical situation field at the location of the grid cell, and this composite vector is the spatial variation characteristic of the physical situation field.

[0029] In this embodiment, extracting semantic event types from semantic description data involves: first, performing natural language parsing on the semantic description data to identify the event subject, event action, and event object, forming event triples; then, matching the event triples with a predefined traffic semantic ontology library, which predefines five semantic event types—traffic accidents, road construction, temporary traffic control, severe weather, and special vehicle passage—along with their corresponding keywords and semantic rules; and finally, classifying the semantic description data into the matched semantic event type to complete the extraction of semantic event types.

[0030] In this embodiment, semantic event types include five types: traffic accidents, road construction, temporary traffic control, severe weather, and special vehicle passage. Traffic accidents refer to events such as vehicle collisions, rear-end collisions, and scrapes that occur in the road network and cause abnormal traffic flow. Road construction refers to events that occupy road space, such as road repairs, pipeline laying, and facility installation. Temporary traffic control refers to events that implement temporary traffic control on specific road sections due to large-scale events, emergencies, or security needs. Severe weather refers to meteorological events such as rain, snowfall, fog, and freezing that affect road traffic conditions. Special vehicle passage refers to events where emergency vehicles such as ambulances, fire trucks, and police cars or special transport vehicles pass through the road network.

[0031] In this embodiment, a semantic situation field is constructed based on semantic event types. The semantic situation field is a discrete symbol field, which means that the semantic event type information of all grid cells in the road network spatial quantization grid is organized into a discrete symbol matrix according to the spatial arrangement of the grid. Each element in the matrix corresponds to the semantic event type encoding value at the location of a grid cell. The matrix as a whole constitutes a discrete symbol field defined on the road network spatial quantization grid. Each grid cell location uses a pre-set numerical encoding to represent the semantic event type that occurs within the current time window at that location. When a grid cell does not have any semantic events within the time window, the encoding value of that grid cell is zero.

[0032] In this embodiment, the discrete symbol field refers to the spatial data structure formed by the discrete distribution of semantic event type encoding values ​​on the road network spatial quantization grid according to the grid unit position. In contrast to the continuous numerical vector field, the value at each grid unit position in the discrete symbol field comes from a finite set of discrete symbols rather than a continuous numerical range. The symbol values ​​between adjacent grid units are not required to have mathematical continuity or differentiability. Its spatial variation reflects the spatial distribution pattern of semantic event types in the road network.

[0033] In this embodiment, all grid units are divided according to preset grid unit division attributes and division levels to obtain grid unit groups under each division attribute and division level. Specifically, grid units with the same road level attribute are grouped into the same road level group, grid units with the same functional area type attribute are grouped into the same functional area type group, and grid units with the same historical congestion level attribute are grouped into the same historical congestion level group. Each division attribute independently completes one division of all grid units. Each group generated by each division is a division level under that division attribute. The grid unit groups under each division level of each division attribute together form a multi-dimensional group division result.

[0034] In this embodiment, the road grade attribute refers to the grade classification of the road corresponding to the grid unit, including four grades: expressway, arterial road, secondary arterial road, and local road; the functional area type attribute refers to the urban functional area classification of the area corresponding to the grid unit, including five types: commercial area, residential area, office area, industrial area, and mixed area; the historical congestion grade attribute refers to the congestion level classification of the grid unit in the same historical period, including four levels: smooth traffic, light congestion, moderate congestion, and severe congestion. This attribute is based on the average speed data of each grid unit in the same historical time window and is classified according to the quartile method.

[0035] In this embodiment, based on the physical situation field and the semantic situation field, within each grid cell group, using other grid cells within the group as references, the physical situation deviation and semantic situation convergence of each grid cell are calculated respectively. Specifically: the physical situation deviation is calculated by using the mean vector and covariance matrix of the physical state parameter vectors of all grid cells in the group as the group statistical reference, and calculating the degree of deviation of the physical state parameter vector of the grid cell relative to the group statistical reference; the semantic situation convergence is calculated by using the frequency distribution of the semantic event type codes of all grid cells in the group as the group semantic reference, and calculating the degree of convergence of the semantic event type codes of the grid cell relative to the group semantic reference. All the above calculations are limited to grid cell groups with the same partitioning attribute and the same partitioning level, and do not compare across groups.

[0036] In this embodiment, the difference between the physical state deviation and semantic state convergence of each grid cell is used as the field state coordination deviation of each grid cell. Specifically, for each grid cell in the road network spatial quantization grid, the difference between the physical state deviation and the semantic state convergence of the grid cell is the field state coordination deviation of that grid cell. When the physical state deviation is greater than the semantic state convergence, the difference is positive, indicating that the abnormal change in the physical state at that grid cell cannot be explained by the semantic event. When the physical state deviation is less than the semantic state convergence, the difference is negative, indicating that the abnormal state indicated by the semantic event at that grid cell has not yet manifested in the physical state.

[0037] In this embodiment, the field situation coordination deviation field is generated based on the gradient of the field situation coordination deviation on the road network spatial quantization grid. This means taking the field situation coordination deviation value of each grid cell in the road network spatial quantization grid as a scalar value at that grid cell, and then calculating the spatial change rate of the scalar value in the row and column directions of the road network spatial quantization grid respectively. The spatial change rates in the row and column directions are combined into a two-dimensional vector. The direction of the two-dimensional vector represents the direction in which the field situation coordination deviation value changes most drastically in space, and the magnitude of the two-dimensional vector represents the rate of change in that direction. The two-dimensional vectors at each grid cell form a vector field, which is the field situation coordination deviation field.

[0038] In this embodiment, the field situation coordination deviation field is used as the current situational awareness data of the road network space and input into the situational warning discriminator to obtain the warning result. The field vectors of all grid cells in the field situation coordination deviation field are organized into a multidimensional tensor according to the spatial arrangement of the road network space quantized grid. This multidimensional tensor is used as the input data of the situational warning discriminator. After calculating the input data, the situational warning discriminator outputs the warning level and warning type.

[0039] In this embodiment, the situation warning discriminator is a multi-task output neural network. This multi-task output neural network includes a shared feature extraction layer, a warning level classification head, and a warning type classification head. The shared feature extraction layer is composed of a convolutional neural network and is used to extract spatial features from the input multidimensional tensor. The warning level classification head is composed of a fully connected layer and a Softmax normalized output layer and is used to output three discrete warning levels. The warning type classification head is composed of a fully connected layer and a Sigmoid activation output layer and is used to output warning type labels. The input of this multi-task output neural network is a multidimensional tensor of the field situation co-variance deviation field, and the output is the warning level and the warning type.

[0040] The beneficial effects of the above technical solutions are as follows: By constructing a dual-field separation representation architecture of physical situation field and semantic situation field, the continuous numerical characteristics of physical perception data and the discrete symbolic characteristics of semantic description data are preserved respectively, eliminating the information loss caused by simply splicing and fusing the two types of heterogeneous data; at the same time, by performing horizontal comparison within grid unit groups divided according to road level, functional area type and historical congestion level, and calculating the physical situation deviation and semantic situation convergence of each grid unit with other grid units in the group as references, the early warning judgment no longer depends on a globally fixed threshold, but adapts to the dynamic reference benchmark of different groups. The difference between the deviation and the convergence is used as the field situation collaborative deviation and the collaborative deviation field is generated based on the gradient and input into the early warning discriminator, realizing a complete technical closed loop from data acquisition to early warning output, effectively improving the accuracy and adaptability of urban commuter traffic situation early warning.

[0041] like Figure 2 As shown, in another embodiment of the present invention, before dividing all grid cells according to preset grid cell division attributes and division levels, an adaptive construction step of a group reference system is also included: Obtain the time series sequence of physical state parameters and semantic event type of each grid cell in the road network spatial quantization grid within the historical time window; The physical state parameters of each grid cell are used to extract the physical situation time-series feature vector of each grid cell. The physical situation time-series feature vector contains the time sequence of the average velocity, time occupancy rate, spatial density and velocity variance of each grid cell at each historical moment. Based on the semantic event type time series of each grid cell, the semantic situation frequency feature vector of each grid cell is statistically analyzed. The semantic situation frequency feature vector contains the normalized value of the occurrence frequency of each semantic event type in each grid cell within the historical time window. Density-based clustering algorithm is used to cluster the time series feature vectors of each physical situation to obtain multiple initial physical groups. Each initial physical group is identified by the group center vector of the time series feature vector of the physical situation. A density-based clustering algorithm is used to cluster the semantic situation frequency feature vectors to obtain multiple initial semantic groups. Each initial semantic group is identified by the group center vector of the semantic situation frequency feature vector. The spatial intersection grid cells of the initial physical group and the initial semantic group are used as hybrid group cells. The group center vector of the physical situation temporal feature vector and the group center vector of the semantic situation frequency feature vector in the hybrid group cell are concatenated to obtain the hybrid group center vector. The hybrid group center vector is used as the initial reference for determining the group affiliation of grid cells under the preset partitioning attributes and partitioning levels. The physical situation temporal feature vector and the semantic situation frequency feature vector of each grid cell are concatenated to obtain the joint feature vector of each grid cell. The cosine similarity between the joint feature vector of each grid cell and the hybrid group center vector is calculated. The initial group affiliation of each grid cell is completed based on the cosine similarity.

[0042] In this embodiment, the historical time window refers to a continuous time interval before the current moment. The length of the time window is the past 30 days. It is used to extract the long-term statistical patterns of the physical state parameters and semantic event types of each grid cell from historical data in order to construct an adaptive initial benchmark for the group reference system.

[0043] In this embodiment, obtaining the physical state parameter time series and semantic event type time series of each grid cell in the road network spatial quantization grid within a historical time window means that for each grid cell in the road network spatial quantization grid, the physical state parameters of the grid cell are collected in chronological order at each time unit within the historical time window to form the physical state parameter time series of the grid cell; at the same time, the semantic event type encoding values ​​of the grid cell are collected in chronological order at each time unit within the historical time window to form the semantic event type time series of the grid cell; each grid cell obtains its own two independent time series sequences.

[0044] In this embodiment, the physical state parameter time series of each grid cell is used to extract the physical situation time series feature vector of each grid cell. This means that the values ​​of the four dimensions of average velocity, time occupancy rate, spatial density and velocity variance of all historical moments in the physical state parameter time series of each grid cell are arranged into four time series in chronological order. Then, the first and last of the four time series are concatenated into a multi-dimensional vector. This multi-dimensional vector is the physical situation time series feature vector of the grid cell, and its length is equal to the number of time units contained in the historical time window multiplied by 4.

[0045] In this embodiment, the temporal arrangement of the average velocity, time occupancy rate, spatial density, and velocity variance of each grid cell at each historical moment refers to the following for each grid cell: the average velocity values ​​from the first moment to the Nth moment within the historical time window are arranged in chronological order to form an average velocity sequence; the time occupancy rate values ​​are arranged in chronological order to form a time occupancy rate sequence; the spatial density values ​​are arranged in chronological order to form a spatial density sequence; and the velocity variance values ​​are arranged in chronological order to form a velocity variance sequence. These four sequences together constitute the temporal arrangement of the grid cell.

[0046] In this embodiment, the semantic situation frequency feature vector of each grid cell is statistically analyzed based on the semantic event type time series of each grid cell. Specifically, for each grid cell, the number of occurrences of traffic accident code value, road construction code value, temporary traffic control code value, severe weather code value, and special vehicle passage code value in its semantic event type time series is counted. The occurrence counts of the above five semantic event types are divided by the total number of time units included in the historical time window to obtain five normalized frequency values. These five normalized frequency values ​​are concatenated into a five-dimensional vector, which is the semantic situation frequency feature vector of the grid cell.

[0047] In this embodiment, the normalized frequency of each semantic event type in each grid cell within the historical time window refers to the following normalized values ​​for each grid cell: Traffic accident normalized value: the number of occurrences of the code value belonging to traffic accidents in the semantic event type time sequence divided by the total length of the time sequence; Road construction normalized value: the number of occurrences of the code value belonging to road construction normalized value: the number of occurrences of the code value belonging to temporary control normalized value: the number of occurrences of the code value belonging to severe weather normalized value: the number of occurrences of the code value belonging to special vehicle passage ...

[0048] The density-based clustering algorithm uses the DBSCAN algorithm, with the neighborhood radius set to twice the center-to-center distance between adjacent grid cells in the road network spatial quantization grid, and the minimum number of samples set to 5 grid cells.

[0049] In this embodiment, density-based clustering is used to cluster the physical situation time-series feature vectors to obtain multiple initial physical groups. Each initial physical group is identified by the group center vector of the physical situation time-series feature vectors. The input sample set is the physical situation time-series feature vectors of all grid cells in the road network spatial quantization grid. The density-based clustering algorithm is used to cluster the input sample set. This algorithm divides the groups by finding dense regions of sample points in high-dimensional space, grouping sample points in dense regions into the same group, and marking sample points in sparse regions as noise points. After clustering, multiple initial physical groups are obtained, and each initial physical group contains multiple grid cells. Then, for each initial physical group, the mean vector of all physical situation time-series feature vectors in the group is calculated. This mean vector is used as the group center vector of the initial physical group to identify the center position of the physical situation features of the group.

[0050] In this embodiment, density-based clustering is used to cluster the semantic situation frequency feature vectors to obtain multiple initial semantic groups. Each initial semantic group is identified by the group center vector of the semantic situation frequency feature vector. The semantic situation frequency feature vectors of all grid cells in the road network spatial quantization grid are used as the input sample set. The density-based clustering algorithm is used to cluster the input sample set. After clustering, multiple initial semantic groups are obtained, and each initial semantic group contains multiple grid cells. Then, for each initial semantic group, the mean vector of all semantic situation frequency feature vectors in the group is calculated. This mean vector is used as the group center vector of the initial semantic group to identify the semantic situation feature center position of the initial semantic group.

[0051] In this embodiment, the spatial intersection grid cells of the initial physical group and the initial semantic group are regarded as hybrid group cells. This means that for each initial physical group and each initial semantic group, grid cells that belong to both the initial physical group and the initial semantic group are identified, and the set of these grid cells is the spatial intersection of the initial physical group and the initial semantic group. When a grid cell belongs to both an initial physical group and an initial semantic group, the grid cell is a hybrid group cell, and the grid cell has group affiliation information with both physical situation temporal characteristics and semantic situation frequency characteristics.

[0052] In this embodiment, the group center vector of the physical situation time-series feature vector and the group center vector of the semantic situation frequency feature vector in the hybrid group unit are concatenated to obtain the hybrid group center vector. Specifically, for each hybrid group unit, the group center vector of the initial physical group to which the hybrid group unit belongs is obtained, and the group center vector of the initial semantic group to which the hybrid group unit belongs is obtained. The two group center vectors are then concatenated end to end into a longer vector, which is the hybrid group center vector of the hybrid group unit. This longer vector integrates the group center information of both the physical situation and the semantic situation dimensions.

[0053] In this embodiment, the hybrid group center vector is used as the initial reference benchmark for dividing grid unit groups under the preset division attributes and division levels. This means that the set of hybrid group center vectors of all hybrid group units in the road network spatial quantization grid is used as the initial reference benchmark for subsequent group classification. This initial reference benchmark provides the standard position of each hybrid group unit in the physical-semantic joint feature space, which is used to measure the feature similarity between other grid units and each hybrid group unit, thereby providing a reference basis for group classification of grid units that have not yet been classified into any hybrid group.

[0054] In this embodiment, the preset division attributes and division levels refer to the pre-defined classification dimensions and their subdivision levels used to divide all grid units in the road network spatial quantification grid in multiple dimensions. The division attributes include three independent classification dimensions: road grade attributes, functional area type attributes, and historical congestion level attributes. The division levels refer to the specific classification values ​​under each division attribute. The division levels under the road grade attribute are four levels: expressway, arterial road, secondary arterial road, and local road. The division levels under the functional area type attribute are five levels: commercial area, residential area, office area, industrial area, and mixed area. The division levels under the historical congestion level attribute are four levels: smooth traffic, light congestion, moderate congestion, and severe congestion.

[0055] In this embodiment, based on the similarity between the physical situation temporal feature vector and semantic situation frequency feature vector of each grid cell and the center vector of the mixed group, an initial group assignment is performed on each grid cell. This means that for each grid cell in the road network spatial quantization grid that has not yet been assigned to any mixed group, the physical situation temporal feature vector and the semantic situation frequency feature vector of the grid cell are concatenated end to end to form a joint feature vector. Then, the cosine similarity between the joint feature vector and the center vector of each mixed group is calculated. The grid cell is assigned to the mixed group corresponding to the center vector of the mixed group with the largest cosine similarity, thus completing the initial group assignment of the grid cell. After performing the above operation on all grid cells that have not been assigned to any mixed group, all grid cells in the road network spatial quantization grid are assigned to the corresponding mixed group, thus completing the initial group assignment.

[0056] In this embodiment, the similarity between the physical situation temporal feature vector and semantic situation frequency feature vector of each grid cell and the hybrid group center vector refers to the cosine similarity between the joint feature vector and the hybrid group center vector after concatenating the physical situation temporal feature vector and the semantic situation frequency feature vector of each grid cell into a joint feature vector. The cosine similarity ranges from -1 to 1. The closer the value is to positive 1, the more consistent the feature direction of the grid cell and the hybrid group center vector are. The closer the value is to negative 1, the more opposite the feature direction is. Each grid cell selects the hybrid group corresponding to the hybrid group center vector with the largest cosine similarity as its belonging group.

[0057] The beneficial effects of the above technical solution are as follows: By extracting the physical situation time-series feature vector and semantic situation frequency feature vector of each grid unit from historical time window data, and using density-based clustering algorithms to cluster them to obtain initial physical groups and initial semantic groups, the spatial intersection of the two groups is used as a mixed group unit, and the physical and semantic group center vectors are concatenated to obtain the mixed group center vector. Then, based on the similarity between the physical situation time-series feature vector and semantic situation frequency feature vector of each grid unit and the mixed group center vector, the initial group classification is performed. This achieves the adaptive determination of the grid unit group classification reference benchmark, avoiding the subjectivity and inaccuracy caused by relying on manual experience or fixed rules to classify groups, and making subsequent intra-group horizontal comparisons have a data-driven and reasonable reference benchmark.

[0058] In another embodiment of the present invention, calculating the physical state deviation of each grid cell within each grid cell group includes: Obtain the physical state feature vector sample set of all grid cells in the grid cell group under the current time window. Each physical state feature vector contains four dimensions: average velocity, time occupancy rate, spatial density, and velocity variance. The covariance matrix and mean vector of the physical state feature vector sample set are calculated to obtain the covariance matrix and mean vector of the physical state feature vectors of all grid cells in the grid cell group. Calculate the Mahalanobis distance between the physical state feature vector and the mean vector of each grid cell. The ratio of the Mahalanobis distance to the maximum Mahalanobis distance of all grid cells in the grid cell group is taken as the physical state deviation of each grid cell. The physical state deviation ranges from 0 to 1. The larger the value, the greater the deviation of the physical state of the grid cell from other grid cells in the group.

[0059] In this embodiment, when the number of grid cells in the grid cell group is less than the dimension of the physical state feature vector, the covariance matrix is ​​a singular matrix. In this case, a regularized covariance matrix is ​​used to replace the original covariance matrix. A very small positive number is added to the diagonal elements of the covariance matrix to ensure that the matrix is ​​invertible, or the Euclidean distance is directly used to replace the Mahalanobis distance.

[0060] In this embodiment, the current time window refers to a continuous time interval before the current moment. The length of the time window is 5 minutes, which is used to limit the time range of physical perception data and semantic description data participating in the calculation of physical situation deviation and semantic situation convergence, so as to ensure that each grid unit performs horizontal comparison within the group at the same time scale.

[0061] In this embodiment, the physical state feature vector sample set of all grid cells in the grid cell group under the current time window is obtained. This means that for a certain grid cell group, the values ​​of the four dimensions of average velocity, time occupancy rate, spatial density and velocity variance of each grid cell in the group under the current time window are combined into a four-dimensional vector. The four-dimensional vectors of all grid cells in the group are gathered together to form the physical state feature vector sample set of the group.

[0062] In this embodiment, the average speed refers to the arithmetic mean of the speeds of all vehicles passing through the road segment corresponding to the grid unit within the current time window, in kilometers per hour; the time occupancy rate refers to the proportion of time that the road segment corresponding to the grid unit is occupied by vehicles within the current time window, with a value between 0 and 1; the spatial density refers to the number of vehicles per unit length in the road segment corresponding to the grid unit within the current time window, in vehicles per kilometer; and the speed variance refers to the variance of the speeds of all vehicles passing through the road segment corresponding to the grid unit within the current time window, used to characterize the degree of speed dispersion.

[0063] In this embodiment, the covariance matrix and mean vector are calculated for the physical state feature vector sample set to obtain the covariance matrix and mean vector of the physical state feature vectors of all grid cells in the corresponding grid cell group. This involves calculating the pairwise covariance values ​​between each of the four dimensions of each sample in the physical state feature vector sample set, arranging all the covariance values ​​into a four-row, four-column covariance matrix. The diagonal elements of this matrix are the variance values ​​of each dimension, and the off-diagonal elements are the covariance values ​​between different dimensions. At the same time, the average value of all samples in the physical state feature vector sample set in each dimension is calculated, and the average values ​​of the four dimensions are combined into a four-dimensional mean vector.

[0064] In this embodiment, the Mahalanobis distance between the physical state feature vector and the mean vector of each grid cell is calculated by subtracting the mean vector of the group to which the physical state feature vector belongs from the physical state feature vector of each grid cell to obtain the difference vector. Then, the transpose of the difference vector is multiplied by the inverse of the covariance matrix of the group to which it belongs, and then multiplied by the difference vector again. Finally, the square root of the calculation result is taken to obtain the Mahalanobis distance between the physical state feature vector and the mean vector of the grid cell.

[0065] In this embodiment, the ratio of the Mahalanobis distance to the maximum Mahalanobis distance of all grid cells within the same grid cell group is used as the physical state deviation of each grid cell. Specifically, after calculating the Mahalanobis distance between the physical state feature vector and the mean vector of all grid cells in the group, the maximum Mahalanobis distance value among all Mahalanobis distances in the group is found. Then, the Mahalanobis distance of each grid cell is divided by this maximum distance value, and the resulting ratio is the physical state deviation of that grid cell. The value of this physical state deviation ranges from 0 to 1; the closer the value is to 1, the greater the deviation of the physical state of that grid cell relative to other grid cells in the group.

[0066] The beneficial effects of the above technical solution are as follows: by obtaining a sample set of physical state feature vectors of all grid cells in the group, the covariance matrix and mean vector of the sample set are calculated as statistical references for the group. Then, the Mahalanobis distance between the physical state feature vector of each grid cell and the mean vector is calculated, and the ratio of this distance to the maximum Mahalanobis distance in the group is used as the physical situation deviation. This makes the calculation of the deviation fully consider the correlation between the four dimensions of average velocity, time occupancy rate, spatial density and velocity variance, as well as the different dimensional scales of each dimension. Thus, it is possible to more accurately quantify the comprehensive deviation of the physical situation of each grid cell relative to other grid cells in the group.

[0067] In another embodiment of the present invention, calculating the semantic situation convergence degree of each grid cell within each grid cell group includes: Obtain the semantic event type encoding sequence of all grid cells within the corresponding grid cell group under the current time window; Construct semantic symbol vectors for each grid cell. The semantic symbol vectors are one-hot encoded vectors. The dimension of the semantic symbol vector is equal to the total number of semantic event type encoding values ​​plus one. If a semantic event occurs in the grid cell within the current time window, the dimension corresponding to the semantic event type encoding value in the semantic symbol vector is set to 1, and the other dimensions are set to 0. If no semantic event occurs in the grid cell within the current time window, the reserved no-event dimension in the semantic symbol vector is set to 1, and the other dimensions are set to 0. The frequency of occurrence of each semantic event type encoding value within the grid cell group is statistically analyzed. A frequency distribution vector is constructed using the frequency of occurrence of all semantic event type encoding values. The frequency distribution vector is used as the baseline probability value of each semantic event type encoding value. Calculate the cross-entropy between the semantic symbol vector and the baseline probability value of each grid cell. Normalize the cross-entropy to the maximum and minimum values ​​within the grid cell group and then invert it to obtain the semantic situation convergence degree, which ranges from 0 to 1. The larger the semantic situation convergence degree, the higher the degree of convergence between the semantic situation of the grid cell and the semantic situation of other grid cells in the group. The maximum and minimum value normalization is calculated with the maximum value of the cross-entropy of all grid cells in the current grid cell group as the upper bound and the minimum value as the lower bound.

[0068] In this embodiment, obtaining the semantic event type encoding sequence of all grid cells within the grid cell group under the current time window means that for a certain grid cell group, the latest semantic event type occurring in each grid cell within the current time window is represented by the corresponding numerical encoding. Each grid cell corresponds to a encoding value. The encoding values ​​of all grid cells in the group are collected together to form the semantic event type encoding sequence of the group. If no semantic event occurs in a certain grid cell within the current time window, the encoding value corresponding to that grid cell is 0.

[0069] In this embodiment, the semantic symbol vector of each grid cell refers to the vector obtained by one-hot encoding the latest semantic event type that occurs in each grid cell within the current time window; the dimension of the semantic symbol vector is equal to the total number of semantic event type encoding values ​​plus one; if a semantic event occurs in the grid cell within the current time window, the dimension corresponding to the semantic event type encoding value in the semantic symbol vector is 1, and the other dimensions are 0; if no semantic event occurs in the grid cell within the current time window, the reserved no-event dimension in the semantic symbol vector is 1, and the other dimensions are 0.

[0070] In this embodiment, the frequency of occurrence of each semantic event type encoding value within the grid cell group is counted. This means that for each encoding value in the semantic event type encoding sequence, the number of times the encoding value appears in the sequence is counted, and the number of occurrences is divided by the total number of encoding values ​​in the sequence to obtain the frequency of occurrence of each encoding value. The frequency of occurrence is between 0 and 1, and the sum of the frequencies of occurrence of all encoding values ​​is 1.

[0071] In this embodiment, a frequency distribution vector is constructed using the occurrence frequencies of all semantic event type coding values. This frequency distribution vector serves as the baseline probability value for each semantic event type coding value. Specifically, the six frequency values—traffic accident coding value, road construction coding value, temporary traffic control coding value, severe weather coding value, special vehicle passage coding value, and coding value 0—are arranged in a fixed order to form a six-dimensional vector, which is the frequency distribution vector. The frequency value of each dimension in this six-dimensional vector is then used as the baseline probability value for the corresponding semantic event type coding value.

[0072] In this embodiment, the semantic symbol vector of each grid cell refers to the one-hot encoding of the latest semantic event type occurring in the current time window of each grid cell according to six semantic states. The six semantic states include no event, traffic accident, road construction, temporary traffic control, severe weather, and special vehicle passage. If the semantic event type of the grid cell is a traffic accident, the traffic accident corresponding dimension in its semantic symbol vector is 1, and the other dimensions are 0. If no semantic event occurs in the grid cell, the no event corresponding dimension in its semantic symbol vector is 1, and the other dimensions are 0. The semantic symbol vector is a six-dimensional vector, with each dimension being 0 or 1, and the sum of the values ​​of each dimension is 1.

[0073] In this embodiment, before calculating the cross-entropy, Laplace smoothing is applied to each dimension of the frequency distribution vector, and a very small positive number is added to the frequency value of each dimension so that the frequency values ​​of all dimensions are greater than zero, thus avoiding the occurrence of infinite values ​​in the cross-entropy calculation due to zero probability.

[0074] In this embodiment, the cross-entropy between the semantic symbol vector of each grid cell and the baseline probability value is calculated. The cross-entropy calculation result is used as the semantic situation convergence of each grid cell. Specifically, for the semantic symbol vector of each grid cell, the position of each dimension with a value of 1 is found, the natural logarithm of the baseline probability value corresponding to that position is taken, and then the negative value of the natural logarithm is taken to obtain the cross-entropy value of that grid cell. The smaller the cross-entropy value, the higher the probability of the semantic event type of that grid cell in the frequency distribution of the group semantic event types, that is, the more similar the semantic situation of that grid cell is to the semantic situation of other grid cells in the group. The value obtained after normalization and inversion is used as the semantic situation convergence of that grid cell.

[0075] The beneficial effects of the above technical solution are as follows: by obtaining the semantic event type encoding sequence of all grid cells in the group under the current time window, the frequency of each encoding value is statistically analyzed to form a frequency distribution vector and used as a baseline probability value. Then, the cross-entropy between the semantic symbol vector of each grid cell and the baseline probability value is calculated as the semantic situation convergence degree. The convergence degree is calculated based on the actual statistical distribution of semantic event types in the group. The smaller the cross-entropy value, the higher the probability of the semantic event type of the grid cell in the group semantic frequency distribution. That is, the semantic situation of the grid cell is more similar to the semantic situation of other grid cells in the group, thereby realizing the comparable quantification of discrete symbolic semantic information under group reference.

[0076] In another embodiment of the present invention, after constructing the physical state field, the method further includes the step of extracting the local field gradient vector of the physical state field: The first-order partial derivatives of the physical state parameters of each grid cell are calculated in the row and column directions of the road network spatial quantization grid to obtain the row and column gradient components of each grid cell in each physical state parameter dimension. The row and column gradient components of each grid cell in each physical state parameter dimension are vectorized to obtain the local field gradient vector of each grid cell. Using the field situational coordination deviation field as the current situational awareness data of the road network space, and combining it with the local field gradient vector, the data are input into the situational warning discriminator to obtain better warning results.

[0077] In this embodiment, the first-order partial derivatives of the physical state parameters of each grid cell are calculated in the row and column directions of the road network spatial quantization grid to obtain the row and column gradient components of each grid cell in each physical state parameter dimension. Specifically, for each grid cell in the road network spatial quantization grid, the average velocity value of the grid cell in the row direction is subtracted from the average velocity value of its adjacent grid cells in the row direction, and then divided by the spatial step size in the row direction to obtain the first-order partial derivative of the grid cell in the row direction in the average velocity dimension. Similarly, the time occupancy rate value of the grid cell in the row direction is calculated. Subtracting the values ​​and dividing by the spatial step size in the row direction yields the first-order partial derivative in the row direction for the time occupancy dimension. The same operation is performed on the spatial density and velocity variance to obtain the first-order partial derivatives in the row direction for the spatial density and velocity variance dimensions, respectively. The above four arrays of first-order partial derivatives in the row direction are combined to form the row direction gradient component of the grid cell. Then, the four physical state parameters of the grid cell in the column direction are subtracted from the four physical state parameter dimensions and divided by the spatial step size in the column direction to obtain the first-order partial derivatives in the column direction for the grid cell in the four physical state parameter dimensions. These are then combined to form the column direction gradient component of the grid cell.

[0078] In this embodiment, the physical state parameter dimension refers to the four independent measurement directions that constitute the physical state parameter, including the average speed dimension, time occupancy dimension, spatial density dimension, and speed variance dimension. Each dimension represents one aspect of the traffic physical state. The average speed dimension reflects the speed of traffic flow, the time occupancy dimension reflects the proportion of time the road is occupied, the spatial density dimension reflects the density of vehicles per unit length of road, and the speed variance dimension reflects the dispersion of traffic flow speed.

[0079] In this embodiment, the row and column gradient components of each grid cell in each physical state parameter dimension are vector-synthesized to obtain the local field gradient vector of each grid cell. Specifically, for each grid cell, the first-order partial derivative of the average velocity in the row direction gradient component and the first-order partial derivative of the average velocity in the column direction gradient component of the grid cell are vector-synthesized to obtain the synthesized gradient value of the grid cell in the average velocity dimension. The same method is applied to the time occupancy rate, spatial density, and velocity variance to obtain the synthesized gradient values ​​of each dimension. Then, the synthesized gradient values ​​of the four dimensions are combined into a four-dimensional vector, which is the local field gradient vector of the grid cell.

[0080] In this embodiment, the field-situational coordination deviation field is used as the current situational awareness data of the road network space, and combined with the local field gradient vector, and input together into the situational warning discriminator to obtain better warning results. Specifically, the field vectors of the field-situational coordination deviation field of all grid cells in the road network space quantization grid are arranged in space to form a multidimensional tensor, and the local field gradient vectors of all grid cells in the road network space quantization grid are arranged in the same space to form another multidimensional tensor. The two multidimensional tensors are spliced ​​in the channel dimension to form a fusion feature tensor. The fusion feature tensor contains both physical-semantic coordination deviation information and physical situational space change rate information. The fusion feature tensor is used as the input data of the situational warning discriminator. The situational warning discriminator calculates and outputs the warning level and warning type based on the richer input features.

[0081] The beneficial effects of the above technical solution are as follows: by extracting the local field gradient vector of each grid cell in the physical situation field, the gradient vector represents the spatial change rate of average velocity, time occupancy rate, spatial density and velocity variance in the row and column directions of the road network spatial quantization grid. The local field gradient vector and the field situation co-deviation field are jointly input into the situation warning discriminator, so that while the warning discriminator makes judgments based on the physical-semantic co-deviation, it can simultaneously obtain the degree of spatial change of the physical situation itself as an auxiliary criterion, thereby enriching the input feature dimensions of the warning decision and helping to improve the response sensitivity and accuracy of the warning results in areas of sudden changes in the physical situation.

[0082] In another embodiment of the present invention, the field-state cooperative deviation field is generated in the following manner: The field situation coordination deviation of each grid cell is used as a scalar value; The gradient components of the scalar values ​​of each grid cell in the row and column directions of the quantized grid in the road network space are used as the field vectors of the field-situational coordination deviation field in the corresponding grid cell to obtain the field-situational coordination deviation field.

[0083] In this embodiment, the field-state coordination deviation of each grid cell is used as a scalar value. For each grid cell in the road network spatial quantization grid, the difference between the physical state deviation of the grid cell and the semantic state convergence of the grid cell is the field-state coordination deviation value of the grid cell. This value has no directional attribute, only magnitude, and is a scalar value.

[0084] In this embodiment, the gradient components of the scalar values ​​of each grid cell in the row and column directions of the road network spatial quantization grid are used as the field vectors of the field-situation coordination deviation field in the corresponding grid cell. This process involves subtracting the scalar values ​​of the field-situation coordination deviations of the adjacent grid cells in the row direction and dividing by the spatial step size in the row direction to obtain the row direction gradient component of the grid cell; subtracting the scalar values ​​of the field-situation coordination deviations of the adjacent grid cells in the column direction and dividing by the spatial step size in the column direction to obtain the column direction gradient component of the grid cell; and combining the row and column direction gradient components into a vector to form a two-dimensional vector, which is the field vector of the grid cell. This process is repeated for all grid cells in the road network spatial quantization grid, and the two-dimensional vectors at each grid cell form a vector field, which is the field-situation coordination deviation field.

[0085] The beneficial effects of the above technical solution are as follows: by taking the difference between the physical state deviation and semantic state convergence of each grid cell as a scalar value, and then using the gradient components of the scalar value in the row and column directions of the grid in the road network space to synthesize a two-dimensional vector as the field vector of each grid cell, the scalar value with only magnitude attribute is converted into a vector field with both magnitude and direction attributes. This organizes the originally scattered cooperative deviation values ​​in each grid cell into a field structure with spatial continuity. The direction of the field vector represents the direction in which the cooperative deviation changes most drastically in space, and the magnitude of the field vector represents the rate of change in that direction. This provides input data with spatial structure for subsequent field line tracking and gradient analysis, enabling early warning judgment to improve accuracy by utilizing the spatial propagation and evolution law of cooperative deviation.

[0086] In another embodiment of the present invention, the method further includes field line tracing and abnormal region geometric calibration steps: The field vector of the field situation coordination deviation field is traced along the field vector direction at each grid cell to obtain multiple field lines of the field situation coordination deviation field. Based on the curvature of each field line and the convergence between field lines, the abnormal convergence region and abnormal divergence region of the field situation coordination deviation field are identified. The calibration results of the anomaly convergence region and the anomaly divergence region are superimposed with the magnitude of the field vector at each grid cell of the field-situation coordination deviation field to generate an enhanced field-situation coordination deviation field containing the region boundary markers. The enhanced field situational coordination deviation field input situational early warning discriminator is used to obtain enhanced early warning results.

[0087] In this embodiment, the field vector of the field situation cooperative deviation field refers to the two-dimensional vector at the position of each grid cell in the field situation cooperative deviation field. The two-dimensional vector is composed of the gradient component of the field situation cooperative deviation scalar value at the grid cell in the row direction and the gradient component in the column direction. The two-dimensional vector has magnitude and direction. Its direction represents the direction in which the field situation cooperative deviation value changes most drastically in space, and its magnitude represents the rate of change in that direction.

[0088] In this embodiment, the field vector of the field-state coordination deviation field is tracked along the field vector direction at each grid cell to obtain multiple field lines of the field-state coordination deviation field. Starting from the boundary grid cell of the road network spatial quantization grid, the field vector direction of the boundary grid cell is used as the initial tracking direction, and the position coordinates of the next tracking point are calculated with a preset step size. After reaching the next tracking point, the field vector direction at that point is obtained through bilinear interpolation, and this direction is used as the tracking direction for the next tracking step. The field vector divergence value at each tracking point is calculated synchronously, and the current tracking is terminated when the sign of the field vector divergence value changes among consecutive tracking points. All tracking points on the current tracking path are connected in the tracking order to form a curve, which is a field line. The above tracking process is performed on all boundary grid cells to obtain multiple field lines of the field-state coordination deviation field.

[0089] In this embodiment, the curvature of each field line refers to the rate of change of the curve direction at each tracking point on each field line, calculated based on the position coordinates of the tracking point and its adjacent tracking points. The larger the curvature value, the more severe the bending of the field line at the tracking point; a curvature value close to zero indicates that the field line at the tracking point is approximately a straight line. The curvature values ​​of all tracking points on each field line are arranged in the tracking order to form the curvature sequence of the field line.

[0090] In this embodiment, the convergence degree between field lines refers to the rate of change of the distance between any two adjacent field lines measured at the same spatial location, along the direction of field line extension. When the distance gradually decreases along the direction of field line extension, it indicates that the two field lines are approaching each other, and the location is in a convergent state. When the distance gradually increases along the direction of field line extension, it indicates that the two field lines are moving away from each other, and the location is in a divergent state. The absolute value and positive or negative sign of the rate of change of the distance are used as quantitative indicators of the convergence degree between field lines.

[0091] In this embodiment, based on the curvature of each field line and the convergence between field lines, the abnormal convergence region and abnormal divergence region in the field situation coordination deviation field are calibrated. Specifically, tracking points with curvature values ​​exceeding a preset curvature threshold are calibrated as high curvature feature points, field line segments with a negative rate of change in spacing between field lines and an absolute value exceeding a preset convergence threshold are calibrated as convergence segments, and field line segments with a positive rate of change in spacing between field lines and an absolute value exceeding a preset divergence threshold are calibrated as divergence segments. Then, connected component clustering is performed on spatially adjacent high curvature feature points, and the boundary of the clustered connected component is extended outward by a preset distance to form an abnormal convergence region. At the same time, connected component clustering is performed on the spatial area covered by the divergence segment, and the connected component obtained by clustering is the abnormal divergence region.

[0092] In this embodiment, the calibration results of the anomaly convergence region and the anomaly divergence region are superimposed with the magnitude of the field vector at each grid cell of the field-situation coordination deviation field to generate an enhanced field-situation coordination deviation field containing regional boundary markers. This involves obtaining the field-situation coordination deviation scalar value of each grid cell in the road network spatial quantization grid, while simultaneously obtaining the boundary location information of the anomaly convergence region and the boundary location information of the anomaly divergence region. For grid cells located inside the boundary of the anomaly convergence region, a preset positive gain value is added to their field-situation coordination deviation scalar value, and a regional attribution marker is added to that grid cell. For grid cells located inside the boundary of the anomaly divergence region, a preset negative gain value is added to their field-situation coordination deviation scalar value, and a regional attribution marker is added to that grid cell. For grid cells located outside the boundary of the anomaly region, their field-situation coordination deviation scalar value remains unchanged, and no regional attribution marker is added. After the above operations are completed, the field-situation coordination deviation scalar value and regional attribution marker at each grid cell together form the enhanced field-situation coordination deviation field.

[0093] In this embodiment, the enhanced situational coordination deviation field is input into the situational warning discriminator to obtain the enhanced warning result. This means that the enhanced scalar values ​​and regional attribution labels of all grid cells in the enhanced situational coordination deviation field are organized into a multidimensional tensor according to the spatial arrangement of the road network spatial quantization grid. Each position element in the multidimensional tensor contains both the enhanced scalar value and the regional attribution label encoding value of the grid cell. The multidimensional tensor is used as the input data of the situational warning discriminator. After calculating the input data, the situational warning discriminator outputs the warning level and warning type.

[0094] The beneficial effects of the above technical solution are as follows: By tracing the field vector of the field situation coordination deviation field along the field vector direction at each grid cell to obtain multiple field lines, and then calibrating the situational anomaly convergence region and situational anomaly divergence region in the field situation coordination deviation field based on the curvature of each field line and the convergence degree between field lines, the above region calibration results are superimposed with the field situation coordination deviation of each grid cell to generate an enhanced field situation coordination deviation field containing region boundary markers. Finally, the enhanced field situation coordination deviation field is input into the situational early warning discriminator, so that the field situation coordination deviation field, which originally only contained numerical information, is superimposed with anomaly region markers with clear spatial boundaries and geometric shapes. The anomaly patterns contained in the high-dimensional field data are extracted through the geometric features of field line curvature and convergence degree and explicitly marked in the form of region boundaries, thereby providing the early warning discriminator with more intuitive and information-rich input data, which helps to improve the accuracy and interpretability of the early warning results in anomaly region identification.

[0095] In another embodiment of the present invention, the field vector of the field-state cooperative deviation field is traced along the field vector direction at each grid cell to obtain multiple field lines of the field-state cooperative deviation field, including: Starting from the boundary grid cells of the road network spatial quantization grid, the field vector direction of each boundary grid cell is used as the initial tracking direction, and the position coordinates of the next tracking point are calculated sequentially with a preset step size; The field vector direction at each tracking point is obtained by bilinear interpolation, and the corresponding direction is used as the tracking direction for the next tracking step. At each tracking point location, the field vector divergence value at the corresponding point is calculated synchronously. When the sign of the field vector divergence value of consecutive tracking points changes or the tracking point exceeds the boundary of the road network spatial quantization grid, the current tracking is terminated, and all tracking points on the current tracking path are connected in sequence to form a single field line. The tracking process is repeated for all boundary grid cells to obtain multiple field lines of the field state cooperative deviation field.

[0096] In this embodiment, the boundary grid cell of the road network spatial quantization grid refers to all grid cells located on the outermost ring of the road network spatial quantization grid, including all grid cells in the first row, last row, first column and last column of the grid; these grid cells are located at the edge of the entire quantization grid and serve as the starting point for field line tracing, used to generate a complete set of field lines covering the entire field.

[0097] In this embodiment, the field vector of the boundary grid cell refers to the two-dimensional vector of the field-state coordination deviation field at the location of the boundary grid cell of the road network spatial quantization grid. The two-dimensional vector is composed of the gradient components of the field-state coordination deviation scalar value at the boundary grid cell in the row and column directions. The field vector serves as the starting direction for field line tracing starting from the boundary grid cell.

[0098] In this embodiment, the preset step size refers to the spatial distance that is advanced in each step during the field line tracking process. This distance is 0.5 times the center-to-center distance between adjacent grid cells in the road network spatial quantization grid, and is used to control the spatial resolution and tracking accuracy of the field line tracking. The smaller the step size, the smoother the tracked field lines are, and the larger the step size, the smaller the tracking computation.

[0099] In this embodiment, starting from the boundary grid cells of the road network spatial quantization grid, the field vector direction of each boundary grid cell is used as the initial tracking direction. The position coordinates of the next tracking point are calculated sequentially with a preset step size. Specifically, for each boundary grid cell, the center position of the boundary grid cell is used as the current tracking point, and the field vector direction of the boundary grid cell is used as the current tracking direction. The tracking point moves along the current tracking direction by a preset step size, and the coordinates of the new position reached after the movement are calculated. This new position is the next tracking point. Then, the above movement operation is repeated with this new position as the current tracking point, and the position coordinates of each subsequent tracking point are calculated sequentially.

[0100] In this embodiment, the field vector direction at each tracking point is obtained using bilinear interpolation, and the corresponding direction is used as the tracking direction for the next tracking step. Specifically, for each calculated tracking point, since the location may not be exactly at the center of a grid cell, the field vector direction at that location needs to be obtained for the next tracking step. The method for obtaining this direction is as follows: find the grid cell containing the location and its four adjacent grid cells, obtain the field vector directions of each of these four grid cells, and then perform a weighted average of the four field vector directions according to the distance between the location and the center of the four grid cells to obtain the field vector direction at that location. The field vector direction obtained after weighted averaging is used as the tracking direction for the next tracking step.

[0101] In this embodiment, the field vector divergence value at each tracking point is calculated synchronously. Specifically, at each tracking point location of the field line tracking, the row and column components of the field vector at that location are obtained using bilinear interpolation. Then, the first-order partial derivative of the row component of the field vector at that location is calculated in the row direction, and the first-order partial derivative of the column component of the field vector at that location is calculated in the column direction. The two first-order partial derivatives are added together to obtain the field vector divergence value at that location. The divergence value is used to characterize whether the field vector at that location is diverging outward or converging inward. A positive divergence value indicates outward divergence, and a negative divergence value indicates inward convergence.

[0102] In this embodiment, the sign of the field vector divergence value changes at continuous tracking points. This means that during the field line tracking process, the divergence value calculated at the current tracking point position is inconsistent with the divergence value calculated at the previous tracking point position in terms of sign. That is, one is positive and the other is negative, or one is positive and the other is zero, or one is negative and the other is zero. When the sign of the divergence value changes, it indicates that the field line segment has entered the convergence region from the divergence region, or entered the divergence region from the convergence region.

[0103] In this embodiment, when the sign of the field vector divergence value of a continuous tracking point changes, the current tracking is terminated, and all tracking points on the current tracking path are connected sequentially to form a single field line. This means that after calculating the divergence value of a new tracking point position, the divergence value is compared with the divergence value of the previous tracking point. If the signs are the same, the next tracking step continues; if the signs are different, the tracking process of the current tracking path is stopped, and all tracking points from the starting boundary grid cell to the current tracking point are connected sequentially in the tracking order to form a continuous curve, which is a complete field line. Each field line starts from the boundary and terminates at the boundary position where the positive and negative signs alternate in the divergence field.

[0104] In this embodiment, the tracking process is repeated for all boundary grid cells to obtain multiple field lines of the field-state coordination deviation field. This means that for each of the boundary grid cells of the road network spatial quantization grid, a complete field line tracking process starting from that boundary grid cell is executed independently. This includes advancing along the field vector direction with a preset step size, obtaining the field vector direction at each tracking point through bilinear interpolation during the advancement process, synchronously calculating the divergence value at each tracking point, terminating the tracking when the divergence value sign changes, and connecting the tracking points to form a field line. Each boundary grid cell generates an independent field line. After all boundary grid cells have completed the tracking, multiple field lines of the field-state coordination deviation field are obtained.

[0105] The beneficial effects of the above technical solution are as follows: Starting from the boundary grid cells of the road network spatial quantization grid, the field vector direction of each boundary grid cell is used as the initial tracking direction and the tracking is gradually advanced with a preset step size. At each tracking point, the field vector direction at that position is obtained by bilinear interpolation as the next tracking direction. At the same time, the field vector divergence value is calculated synchronously at each tracking point, and the current tracking is terminated when the divergence value sign changes at consecutive tracking points. All tracking points on the tracking path are connected to form a single field line. By repeating the above process for all boundary grid cells, multiple field lines are obtained, realizing the complete tracking of field lines extending from the boundary to the interior in the field situation cooperative deviation field. The change in the divergence value sign reflects the turning point of the field vector from the divergent state to the convergent state. Using this turning point as the field line termination position ensures that each field line terminates at the structural change boundary of the physical-semantic cooperative deviation field. Thus, the obtained field line set can effectively cover the entire road network spatial quantization grid and accurately reflect the spatial structural characteristics of the cooperative deviation field.

[0106] like Figure 3 As shown, in another embodiment of the present invention, a time-domain adaptive calibration step for the group statistical reference baseline within the current time window is further included: Obtain the physical and semantic situation fields of each grid cell in the road network spatial quantization grid within the same historical time window; Based on the physical and semantic situation fields of each grid cell within the historical time window, the mean physical situation parameters and the probability of occurrence of semantic event types of each grid cell in the historical period are calculated. The historical baseline physical and semantic situation fields are constructed using the mean physical situation parameters and the probability of occurrence of semantic event types of each grid cell. The physical situation field of each grid cell in the current time window is compared with the historical benchmark physical situation field in the same period by performing a grid-by-grid difference calculation to obtain the physical situation time domain offset of each grid cell. The semantic situation field of each grid cell in the current time window is compared with the historical benchmark semantic situation field by performing grid-by-grid relative entropy calculation to obtain the temporal offset of the semantic situation of each grid cell. The physical state time-domain offset and semantic state time-domain offset of each grid cell are concatenated to obtain the time-domain offset vector of each grid cell. The time-domain offset vector of each grid cell is divided by the historical standard deviation of the corresponding dimension to obtain the standardized time-domain offset vector of each grid cell. The magnitude of the standardized time-domain offset vector is used as the time-domain offset degree of each grid cell. Using the temporal offset of each grid cell as the update weight of the group statistical reference baseline of the grid cell group within the current time window, the calculation reference values ​​of the physical situation deviation and the semantic situation convergence within the current time window are weighted and updated to obtain the group statistical reference baseline after temporal calibration. Using the time-domain calibrated group statistical reference baseline, the physical state deviation and semantic state convergence of each grid cell within the current time window are corrected, and the corrected physical state deviation and semantic state convergence are used for the field state coordination deviation calculation in the current time window. Meanwhile, the time-domain calibrated group statistical reference baseline is used as the calculation reference value for the next time window, which is used to calculate the physical situation deviation and semantic situation convergence in the next time window. Meanwhile, the historical benchmark physical situation field, the historical benchmark semantic situation field, and the time domain offset are used for subsequent dynamic deviation index calculation.

[0107] In this embodiment, the historical time window refers to a set of historical data for multiple consecutive time periods prior to the current moment. Specifically, it refers to the time intervals within the past four weeks that are in the same weekday and the same time period as the current time window. For example, if the current time window is from 8:00 AM to 8:05 AM on Wednesday, then the historical time window is the four time intervals from 8:00 AM to 8:05 AM on Wednesday of the previous week, two weeks, three weeks, and four weeks ago, which are used to extract the long-term historical trend patterns of each grid unit in the road network spatial quantization grid under the same periodic time period.

[0108] In this embodiment, obtaining the physical and semantic situation fields of each grid cell in the road network spatial quantization grid within a historical time window involves extracting the physical state parameters and semantic event type codes of each grid cell from all time intervals within the historical time window for each grid cell in the road network spatial quantization grid. The extracted data are then arranged in chronological order to form the physical and semantic situation field data of the grid cell within the historical time window. The above operations are performed on all grid cells in the road network spatial quantization grid to obtain the physical and semantic situation field sets of all grid cells within the historical time window.

[0109] In this embodiment, based on the physical and semantic situation fields of each grid cell within the historical time window, the mean of physical situation parameters and the probability of occurrence of semantic event types for each grid cell within the historical time window are calculated. Specifically, for each grid cell, the average velocity values ​​of the grid cell in all time intervals within the historical time window are summed and then divided by the total number of time intervals to obtain the historical average velocity mean of the grid cell. The historical average values ​​of time occupancy rate, spatial density, and velocity variance are calculated in the same way, and these four average values ​​are used as the mean of physical situation parameters for the grid cell. At the same time, the number of occurrences of each semantic event type of the grid cell within the historical time window is divided by the total number of time intervals to obtain the historical probability of occurrence of each semantic event type of the grid cell.

[0110] In this embodiment, a historical baseline physical situation field and a historical baseline semantic situation field are constructed using the average physical situation parameters and the probability of occurrence of semantic event types of each grid cell. This means that the average physical situation parameters of each grid cell in the road network spatial quantization grid are organized into a multi-dimensional matrix according to the spatial arrangement of the grid. The element at each position in this matrix is ​​the historical average speed, average time occupancy rate, average spatial density, and average speed variance of the corresponding grid cell. This multi-dimensional matrix is ​​the historical baseline physical situation field. At the same time, the probability of occurrence of semantic event types of each grid cell is organized into another multi-dimensional matrix according to the same spatial arrangement. The element at each position in this matrix is ​​the historical probability of occurrence of the six semantic states of the corresponding grid cell. This multi-dimensional matrix is ​​the historical baseline semantic situation field.

[0111] In this embodiment, the physical situation field of each grid cell in the current time window is compared with the historical reference physical situation field of the same period by performing a grid-by-grid difference calculation to obtain the physical situation time-domain offset of each grid cell. Specifically, for each grid cell in the road network spatial quantization grid, the average velocity time-domain offset of the grid cell is obtained by subtracting the average velocity mean of the grid cell in the historical reference physical situation field from the average velocity value of the grid cell in the current time window. The time-domain offsets of time occupancy rate, spatial density and velocity variance are calculated in the same way. The above four time-domain offsets are combined into a four-dimensional vector, which is the physical situation time-domain offset of the grid cell.

[0112] In this embodiment, the semantic situation field of each grid cell in the current time window is compared with the historical reference semantic situation field to perform grid-by-grid relative entropy calculation to obtain the semantic situation time-domain offset of each grid cell. Specifically, for each grid cell in the road network spatial quantization grid, the semantic symbol vector of the grid cell in the current time window is obtained, and the probability of occurrence of the six semantic states in the historical reference semantic situation field of the grid cell in the same period ...

[0113] In this embodiment, the physical situation time-domain offset and semantic situation time-domain offset of each grid cell are concatenated to obtain the time-domain offset vector of each grid cell. Specifically, for each grid cell, the four-dimensional physical situation time-domain offset of the grid cell is used as the physical component, and the one-dimensional semantic situation time-domain offset of the grid cell is used as the semantic component. The physical component and the semantic component are concatenated end to end to form a five-dimensional vector, which is the time-domain offset vector of the grid cell.

[0114] In this embodiment, the time-domain offset vector of each grid cell is divided by the historical standard deviation of the corresponding dimension to obtain the standardized time-domain offset vector of each grid cell. The magnitude of the standardized time-domain offset vector is used as the time-domain offset of each grid cell. Specifically, for each grid cell, the standard deviation of each dimension of the grid cell within the historical time window is obtained. The value of each dimension in the time-domain offset vector is divided by the historical standard deviation of the corresponding dimension to obtain the standardized time-domain offset vector. The values ​​of the five dimensions in the standardized time-domain offset vector are squared and added together. The square root of the sum is then taken to obtain the value of the time-domain offset of the grid cell.

[0115] In this embodiment, the temporal offset of each grid cell is used as the update weight of the group statistical reference baseline of the grid cell group within the current time window. This means that for each grid cell, its temporal offset is used as the weight coefficient of the group to which the grid cell belongs when updating the group statistical reference baseline. When the temporal offset is large, it indicates that the current situation of the grid cell deviates significantly from the historical same period, and the grid cell occupies a larger proportion in the weighted update, so that the updated group statistical reference baseline reflects the current significant changes in the situation more. When the temporal offset is small, it indicates that the situation of the grid cell is stable, and the grid cell occupies a smaller proportion in the weighted update.

[0116] In this embodiment, the reference values ​​for calculating the physical situation deviation and semantic situation convergence within the current time window refer to the statistical quantities of physical state parameters within the group on which the calculation of physical situation deviation depends in "calculating the physical situation deviation and semantic situation convergence of each grid cell within each grid cell group respectively". Specifically, these are the mean vector and covariance matrix of the physical state feature vectors of all grid cells within the group. The statistical quantities of semantic event types within the group on which the calculation of semantic situation convergence depends are specifically the frequency distribution vector of the semantic event type encoding of all grid cells within the group. The mean vector, covariance matrix, and frequency distribution vector together constitute the reference values ​​for calculating the physical situation deviation and semantic situation convergence.

[0117] In this embodiment, the reference values ​​for calculating the physical situation deviation and semantic situation convergence within the current time window are updated with weights to obtain the time-domain calibrated group statistical reference baseline. Specifically, for each grid cell group, the time-domain offset of each grid cell in the group is used as a weight to calculate the weighted average of the values ​​of each dimension in the mean vector of the group, resulting in an updated mean vector; the elements in the covariance matrix of the group are weighted to calculate an updated covariance matrix; and the frequency values ​​in the frequency distribution vector of the group are weighted to calculate an updated frequency distribution vector. The updated mean vector, the updated covariance matrix, and the updated frequency distribution vector together constitute the time-domain calibrated group statistical reference baseline.

[0118] In this embodiment, the physical situation deviation and semantic situation convergence of each grid cell within the current time window are corrected using a time-domain calibrated group statistical reference baseline. The corrected physical situation deviation and semantic situation convergence are then used to calculate the field situation coordination deviation for the current time window. Specifically, for each grid cell, the Mahalanobis distance of that grid cell is recalculated using the updated mean vector and updated covariance matrix from the time-domain calibrated group statistical reference baseline. The ratio of the recalculated Mahalanobis distance to the maximum value within the group is used as the corrected physical situation deviation. Simultaneously, the cross-entropy of that grid cell is recalculated using the updated frequency distribution vector from the time-domain calibrated group statistical reference baseline. The recalculated cross-entropy value is used as the corrected semantic situation convergence. The corrected physical situation deviation and corrected semantic situation convergence replace the original deviation and convergence for calculating the field situation coordination deviation for the current time window.

[0119] In this embodiment, the time-domain calibrated group statistical reference baseline is used as the calculation reference value for the next time window. It is used to calculate the physical situation deviation and semantic situation convergence in the next time window. This means that the time-domain calibrated group statistical reference baseline obtained in the current time window is stored in a cache. After the current time window ends and the next time window begins, the time-domain calibrated group statistical reference baseline is used as the initial calculation reference value for each grid cell group in the next time window. It is used to calculate the physical situation deviation and semantic situation convergence in the next time window. After the time-domain calibrated group statistical reference baseline is used for the next time window, the real-time data of the next time window will be calculated based on the initial reference value of the baseline. After the next time window ends, time-domain adaptive calibration will be performed again to obtain a new time-domain calibrated reference value, thereby realizing the continuous iterative update of the group statistical reference baseline in the time dimension.

[0120] The beneficial effects of the above technical solution are as follows: By acquiring the physical and semantic situation fields of each grid cell within the historical time window, the mean of physical situation parameters and the probability of occurrence of semantic event types are calculated to construct the historical baseline physical and semantic situation fields. Then, the physical situation field of the current time window is interpolated with the historical baseline physical situation field to obtain the physical situation time-domain offset. Similarly, the semantic situation field of the current time window is interpolated with the historical baseline semantic situation field to obtain the semantic situation time-domain offset. Finally, the two offset vectors are synthesized, and their magnitude is used as the time-domain offset. This time-domain offset is then used as the weight for... Within the current time window, the group statistical reference baseline is weighted and updated to obtain a time-domain calibrated group statistical reference baseline. Then, this calibrated baseline is used to correct the physical situation deviation and semantic situation convergence of the current time window and is used to calculate the current field situation coordination deviation. At the same time, the calibrated baseline is used as the calculation reference value for the next time window, thereby realizing the adaptive iterative update of the group statistical reference baseline in the time dimension. This allows the calculation reference values ​​of physical situation deviation and semantic situation convergence to be dynamically adjusted according to the time-varying characteristics of traffic situation, avoiding the warning deviation caused by the failure of fixed reference values ​​in scenarios such as seasonal traffic flow changes and the impact of special events.

[0121] like Figure 3 As shown, in another embodiment of the present invention, the method further includes steps of dynamic deviation index calculation, adaptive dynamic threshold calibration, and field vector weighted modulation. The difference between the field situation coordination deviation of each grid cell and the historical reference field situation coordination deviation of the same grid cell is calculated to obtain the reference deviation of each grid cell. The historical reference field situation coordination deviation is calculated based on the average physical situation deviation and the average semantic situation convergence of each grid cell within the historical time window. The absolute value of the reference deviation of each grid cell is multiplied by the time-domain offset to obtain the field situation dynamic deviation index of each grid cell. Based on the spatial distribution of the dynamic deviation index of the field situation in each grid cell on the road network spatial quantization grid, a dynamic deviation index field is constructed. Calculate the mean and standard deviation of the field state dynamic deviation index for all grid cells, and use the product of the mean plus the standard deviation and the preset multiplication factor as the dynamic threshold. The ratio of the field situation dynamic deviation index to the dynamic threshold of each grid cell is used as the actual weight coefficient of each grid cell. The field vector of the field situation cooperative deviation field is weighted and modulated using the actual weight coefficient to obtain the weighted and modulated field situation cooperative deviation field. The weighted modulated field situation collaborative deviation field is input into the situation warning discriminator to obtain the weighted warning result.

[0122] In this embodiment, the field situation coordination deviation of each grid cell is compared with the historical benchmark field situation coordination deviation of the same grid cell to obtain the benchmark deviation of each grid cell. Specifically, for each grid cell in the road network spatial quantization grid, the field situation coordination deviation value of the grid cell in the current time window is obtained, and the historical benchmark field situation coordination deviation value of the grid cell in the historical time window is obtained. The current field situation coordination deviation value is subtracted from the historical benchmark field situation coordination deviation value, and the difference is the benchmark deviation of the grid cell. A positive benchmark deviation value indicates that the current coordination deviation is higher than the historical level, and a negative value indicates that the current coordination deviation is lower than the historical level.

[0123] In this embodiment, the historical synchronous reference field situation coordination deviation is calculated based on the average physical situation deviation and the average semantic situation convergence of each grid cell within the historical synchronous time window. Specifically, for each grid cell, the average physical situation deviation of the grid cell in all time intervals within the historical synchronous time window is obtained by summing the physical situation deviations of the grid cell in all time intervals within the historical synchronous time window and dividing by the total number of time intervals; the average semantic situation convergence of the grid cell in all time intervals within the historical synchronous time window is obtained by summing the semantic situation convergence of the grid cell in all time intervals within the historical synchronous time window and dividing by the total number of time intervals; the difference between the average physical situation deviation and the average semantic situation convergence is the historical synchronous reference field situation coordination deviation of the grid cell.

[0124] In this embodiment, the absolute value of the reference deviation of each grid cell is multiplied by the time-domain offset to obtain the field situation dynamic deviation index of each grid cell. Specifically, for each grid cell, the absolute value of the reference deviation of the grid cell is multiplied by the time-domain offset of the grid cell, and the resulting product is the field situation dynamic deviation index of the grid cell. This index comprehensively reflects the deviation of the current cooperative deviation from the historical period and the degree of deviation of the current situation from the historical period. The larger the index, the more significant the deviation of the current situation of the grid cell from the historical pattern.

[0125] In this embodiment, a dynamic deviation index field is constructed based on the spatial distribution of the field status dynamic deviation index of each grid cell on the road network spatial quantization grid. This means that the field status dynamic deviation index values ​​of all grid cells in the road network spatial quantization grid are organized into a two-dimensional matrix according to the spatial arrangement of the grid. Each element in the matrix corresponds to the field status dynamic deviation index value at the location of a grid cell. This two-dimensional matrix is ​​the dynamic deviation index field, which is used to characterize the distribution of the field status dynamic deviation index on the entire road network spatial quantization grid.

[0126] In this embodiment, the mean and standard deviation of the field situation dynamic deviation index of all grid cells are calculated. The sum of the products of the mean and standard deviation and the preset multiplier coefficient is used as the dynamic threshold. The mean of the field situation dynamic deviation index is obtained by adding the field situation dynamic deviation index values ​​of all grid cells in the road network spatial quantization grid and dividing by the total number of grid cells. The square of the difference between the field situation dynamic deviation index of each grid cell and the mean is calculated. The square values ​​of all the squares are added together and divided by the total number of grid cells. The square root of the calculation result is then taken to obtain the standard deviation of the field situation dynamic deviation index. The mean is added to the product of the standard deviation and the preset multiplier coefficient, and the resulting value is the dynamic threshold.

[0127] In this embodiment, the preset multiplier coefficient is a pre-set constant with a value of 2, which is used to control the strictness of the dynamic threshold. The larger the multiplier coefficient, the higher the dynamic threshold and the fewer grid cells are judged as having high dynamic deviation. The smaller the multiplier coefficient, the lower the dynamic threshold and the more grid cells are judged as having high dynamic deviation.

[0128] In this embodiment, the ratio of the field situation dynamic deviation index of each grid cell to the dynamic threshold is used as the actual weight coefficient of each grid cell. Specifically, for each grid cell, the field situation dynamic deviation index of that grid cell is divided by the dynamic threshold, and the resulting ratio is the actual weight coefficient of that grid cell. When the dynamic deviation index is greater than the dynamic threshold, the actual weight coefficient is greater than 1, and the grid cell obtains enhanced weight in subsequent weighted modulation. When the dynamic deviation index is equal to the dynamic threshold, the actual weight coefficient is equal to 1, and the grid cell obtains a baseline weight. When the dynamic deviation index is less than the dynamic threshold, the actual weight coefficient is less than 1, and the grid cell obtains attenuation weight.

[0129] In this embodiment, the field vector of the field situation coordination deviation field is weighted and modulated using actual weight coefficients to obtain the weighted and modulated field situation coordination deviation field. Specifically, for each grid cell in the road network spatial quantization grid, the actual weight coefficient of the grid cell is multiplied by the row direction component of the field vector of the field situation coordination deviation field at that grid cell to obtain the weighted and modulated row direction component of the grid cell; the actual weight coefficient is multiplied by the column direction component of the field vector of the field situation coordination deviation field at that grid cell to obtain the weighted and modulated column direction component of the grid cell; the weighted and modulated row direction component and column direction component are combined to form the weighted and modulated field vector of the grid cell; after performing the above operations on all grid cells in the road network spatial quantization grid, the weighted and modulated field situation coordination deviation field is formed by the weighted and modulated field vectors of all grid cells.

[0130] In this embodiment, the weighted modulated field situation coordination deviation field is input into the situation warning discriminator to obtain a weighted warning result. This means that the weighted field vectors of all grid cells in the weighted modulated field situation coordination deviation field are organized into a multidimensional tensor according to the spatial arrangement of the road network spatial quantization grid. Each position element in the multidimensional tensor contains the weighted modulated row direction component and column direction component of the grid cell. The multidimensional tensor is used as the input data of the situation warning discriminator. After calculating the input data, the situation warning discriminator outputs the warning level and warning type.

[0131] The beneficial effects of the above technical solution are as follows: By calculating the difference between the field situation coordination deviation of each grid unit and the historical reference field situation coordination deviation, the reference deviation is obtained. Then, the reference deviation is multiplied by the time domain offset to obtain the field situation dynamic deviation index. Based on the distribution of this index in the road network space, a dynamic deviation index field is constructed. The mean and standard deviation of the dynamic deviation index of all grid units are calculated. The product of the mean and standard deviation and the preset multiplier coefficient is used as the dynamic threshold. Then, the ratio of the dynamic deviation index of each grid unit to the dynamic threshold is used as the actual weight coefficient to perform weighted modulation on the field vector of the field situation coordination deviation field and input into the early warning discriminator. This allows the grid unit with the higher dynamic deviation index to obtain a higher enhancement weight in the weighted modulation, so that the attention of the early warning discriminator is adaptively focused on the spatial region where the current situation deviates most significantly from the historical period. This realizes the dynamic response of early warning discrimination to the differentiated importance of different regions.

[0132] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for intelligent early warning of urban commuter traffic situation based on multi-source data fusion, characterized in that, include: Acquire multi-source sensing data from the urban commuter road network. The multi-source sensing data includes physical sensing data and semantic description data. Construct a spatial quantization grid for the road network, map physical sensing data to each grid cell and extract physical state parameters, and construct a physical situation field based on the physical state parameters of each grid cell. The physical situation field is a continuous numerical vector field. Semantic event types are extracted from semantic description data, and a semantic situation field is constructed based on the semantic event types. The semantic situation field is a discrete symbol field. All grid units are divided according to the preset grid unit division attributes and division levels to obtain grid unit groups under each division attribute and division level; the preset grid unit division attributes include road level attributes, functional area type attributes, and historical congestion level attributes. Based on the physical situation field and the semantic situation field, within each grid cell group, with other grid cells in the group as references, the physical situation deviation and semantic situation convergence of each grid cell are calculated respectively. The difference between the physical state deviation and the semantic state convergence of each grid cell is used as the field state coordination deviation of each grid cell. The field state coordination deviation field is generated based on the gradient of the field state coordination deviation on the road network spatial quantization grid. The field situational coordination deviation field is used as the current situational awareness data of the road network space and input into the situational warning discriminator to obtain the warning result.

2. The intelligent early warning method for urban commuter traffic situation based on multi-source data fusion according to claim 1, characterized in that, Before dividing all grid cells according to the preset grid cell division attributes and division hierarchy, there is also an adaptive construction step of the group reference system: Obtain the time series sequence of physical state parameters and semantic event type of each grid cell in the road network spatial quantization grid within the historical time window; The physical state parameters of each grid cell are used to extract the physical situation time-series feature vector of each grid cell. The physical situation time-series feature vector contains the time sequence of the average velocity, time occupancy rate, spatial density and velocity variance of each grid cell at each historical moment. Based on the semantic event type time series of each grid cell, the semantic situation frequency feature vector of each grid cell is statistically analyzed. The semantic situation frequency feature vector contains the normalized value of the occurrence frequency of each semantic event type in each grid cell within the historical time window. Density-based clustering algorithm is used to cluster the time series feature vectors of each physical situation to obtain multiple initial physical groups. Each initial physical group is identified by the group center vector of the time series feature vector of the physical situation. A density-based clustering algorithm is used to cluster the semantic situation frequency feature vectors to obtain multiple initial semantic groups. Each initial semantic group is identified by the group center vector of the semantic situation frequency feature vector. The spatial intersection grid cells of the initial physical group and the initial semantic group are taken as hybrid group cells. The group center vector of the physical situation temporal feature vector and the group center vector of the semantic situation frequency feature vector in the hybrid group cell are concatenated to obtain the hybrid group center vector. The hybrid group center vector is used as the initial reference benchmark for determining the group affiliation of grid cells under the preset partitioning attributes and partitioning levels. Based on the similarity between the physical situation temporal feature vector and the semantic situation frequency feature vector of each grid cell and the hybrid group center vector, the initial group affiliation of each grid cell is completed.

3. The intelligent early warning method for urban commuter traffic situation based on multi-source data fusion according to claim 1, characterized in that, Calculate the physical state deviation of each grid cell within each grid cell group, including: Obtain the physical state feature vector sample set of all grid cells in the grid cell group under the current time window. Each physical state feature vector contains four dimensions: average velocity, time occupancy rate, spatial density, and velocity variance. The covariance matrix and mean vector of the physical state feature vector sample set are calculated to obtain the covariance matrix and mean vector of the physical state feature vectors of all grid cells in the grid cell group. Calculate the Mahalanobis distance between the physical state feature vector and the mean vector of each grid cell, and take the ratio of the Mahalanobis distance to the maximum value of the Mahalanobis distance of all grid cells in the grid cell group as the physical state deviation of each grid cell.

4. The intelligent early warning method for urban commuter traffic situation based on multi-source data fusion according to claim 1, characterized in that, Calculate the semantic situation convergence of each grid cell within each grid cell group, including: Obtain the semantic event type encoding sequence of all grid cells within the corresponding grid cell group under the current time window; The frequency of occurrence of each semantic event type encoding value within the grid cell group is statistically analyzed. A frequency distribution vector is constructed using the frequency of occurrence of all semantic event type encoding values. The frequency distribution vector is used as the baseline probability value of each semantic event type encoding value. Calculate the cross-entropy between the semantic symbol vector and the baseline probability value of each grid cell, normalize the cross-entropy by the maximum and minimum values ​​within the grid cell group, and then invert it to obtain the semantic situation convergence degree of each grid cell.

5. The intelligent early warning method for urban commuter traffic situation based on multi-source data fusion according to claim 1, characterized in that, After constructing the physical state field, the process also includes extracting the local field gradient vector from the physical state field: The first-order partial derivatives of the physical state parameters of each grid cell are calculated in the row and column directions of the road network spatial quantization grid to obtain the row and column gradient components of each grid cell in each physical state parameter dimension. The row and column gradient components of each grid cell in each physical state parameter dimension are vectorized to obtain the local field gradient vector of each grid cell. Using the field situational coordination deviation field as the current situational awareness data of the road network space, and combining it with the local field gradient vector, the data are input into the situational warning discriminator to obtain better warning results.

6. The intelligent early warning method for urban commuter traffic situation based on multi-source data fusion according to claim 1, characterized in that, The field situation coordination deviation field is generated in the following way: The field situation coordination deviation of each grid cell is used as a scalar value; The gradient components of the scalar values ​​of each grid cell in the row and column directions of the quantized grid in the road network space are used as the field vectors of the field-situational coordination deviation field in the corresponding grid cell to obtain the field-situational coordination deviation field.

7. The intelligent early warning method for urban commuter traffic situation based on multi-source data fusion according to claim 6, characterized in that, It also includes field line tracking and anomalous region geometry calibration steps: The field vector of the field situation coordination deviation field is traced along the field vector direction at each grid cell to obtain multiple field lines of the field situation coordination deviation field. Based on the curvature of each field line and the convergence between field lines, the abnormal convergence region and abnormal divergence region of the field situation coordination deviation field are identified. The calibration results of the anomaly convergence region and the anomaly divergence region are superimposed with the magnitude of the field vector at each grid cell of the field-situation coordination deviation field to generate an enhanced field-situation coordination deviation field containing the region boundary markers. The enhanced field situational coordination deviation field input situational early warning discriminator is used to obtain enhanced early warning results.

8. The intelligent early warning method for urban commuter traffic situation based on multi-source data fusion according to claim 7, characterized in that, By tracing the field vector of the field-state coordination deviation field along the field vector direction at each grid cell, multiple field lines of the field-state coordination deviation field are obtained, including: Starting from the boundary grid cells of the road network spatial quantization grid, the field vector direction of each boundary grid cell is used as the initial tracking direction, and the position coordinates of the next tracking point are calculated sequentially with a preset step size; The field vector direction at each tracking point is obtained by bilinear interpolation, and the corresponding direction is used as the tracking direction for the next tracking step. At each tracking point location, the field vector divergence value at the corresponding point is calculated synchronously. When the sign of the field vector divergence value of consecutive tracking points changes or the tracking point exceeds the boundary of the road network spatial quantization grid, the current tracking is terminated, and all tracking points on the current tracking path are connected in sequence to form a single field line. The tracking process is repeated for all boundary grid cells to obtain multiple field lines of the field state cooperative deviation field.

9. The intelligent early warning method for urban commuter traffic situation based on multi-source data fusion according to claim 1, characterized in that, It also includes a time-domain adaptive calibration step for the group statistical reference baseline within the current time window: Obtain the physical and semantic situation fields of each grid cell in the road network spatial quantization grid within the same historical time window; Based on the physical and semantic situation fields of each grid cell within the historical time window, the mean physical situation parameters and the probability of occurrence of semantic event types of each grid cell in the historical period are calculated. The historical baseline physical and semantic situation fields are constructed using the mean physical situation parameters and the probability of occurrence of semantic event types of each grid cell. The physical situation field of each grid cell in the current time window is compared with the historical benchmark physical situation field in the same period by performing a grid-by-grid difference calculation to obtain the physical situation time domain offset of each grid cell. The semantic situation field of each grid cell in the current time window is compared with the historical benchmark semantic situation field by performing grid-by-grid relative entropy calculation to obtain the temporal offset of the semantic situation of each grid cell. The physical state time-domain offset and semantic state time-domain offset of each grid cell are concatenated to obtain the time-domain offset vector of each grid cell. The time-domain offset vector of each grid cell is divided by the historical standard deviation of the corresponding dimension to obtain the standardized time-domain offset vector of each grid cell. The magnitude of the standardized time-domain offset vector is used as the time-domain offset degree of each grid cell. Using the temporal offset of each grid cell as the update weight of the group statistical reference baseline of the grid cell group within the current time window, the calculation reference values ​​of the physical situation deviation and the semantic situation convergence within the current time window are weighted and updated to obtain the group statistical reference baseline after temporal calibration. Using the time-domain calibrated group statistical reference baseline, the physical state deviation and semantic state convergence of each grid cell within the current time window are corrected, and the corrected physical state deviation and semantic state convergence are used for the field state coordination deviation calculation in the current time window. Meanwhile, the time-domain calibrated group statistical reference baseline is used as the calculation reference value for the next time window, which is used to calculate the physical situation deviation and semantic situation convergence in the next time window.

10. The intelligent early warning method for urban commuter traffic situation based on multi-source data fusion according to claim 9, characterized in that, It also includes the steps of dynamic deviation index calculation, adaptive dynamic threshold calibration, and field vector weighted modulation: The difference between the field situation coordination deviation of each grid cell and the historical reference field situation coordination deviation of the same grid cell is calculated to obtain the reference deviation of each grid cell. The historical reference field situation coordination deviation is calculated based on the average physical situation deviation and the average semantic situation convergence of each grid cell within the historical time window. The absolute value of the reference deviation of each grid cell is multiplied by the time-domain offset to obtain the field situation dynamic deviation index of each grid cell. Based on the spatial distribution of the dynamic deviation index of the field situation in each grid cell on the road network spatial quantization grid, a dynamic deviation index field is constructed. Calculate the mean and standard deviation of the field state dynamic deviation index for all grid cells, and use the product of the mean plus the standard deviation and the preset multiplication factor as the dynamic threshold. The ratio of the field situation dynamic deviation index to the dynamic threshold of each grid cell is used as the actual weight coefficient of each grid cell. The field vector of the field situation cooperative deviation field is weighted and modulated using the actual weight coefficient to obtain the weighted and modulated field situation cooperative deviation field. The weighted modulated field situation collaborative deviation field is input into the situation warning discriminator to obtain the weighted warning result.

Images