A Highway Traffic Data Analysis Method Based on ETC System

Abstract

The present application belongs to the field of expressway traffic data analysis, and relates to an expressway traffic data analysis method based on an ETC system, which acquires all ETC gantry collected original vehicle passing data streams including vehicle identification, passing time stamp, gantry position code and vehicle type code in a preset time period of a monitored road section, analyzes the original vehicle passing data streams to generate individual dynamic characteristics of vehicle speed and static characteristics of vehicle type groups, constructs a multi-dimensional risk data set indexed by road section identification and time window, inputs the multi-dimensional risk data set into a pre-trained deep learning model, outputs an accident hidden danger probability value result, performs early warning state determination based on a probability gradient determination rule, generates an accident risk heat map of the monitored road section according to the early warning state and feeds back, effectively solves the problems of insufficient forward-looking prediction and low regional accident hidden danger evaluation accuracy in the prior art, and significantly improves the timeliness and accuracy of traffic risk evaluation.

Description

Technical Field

[0001] This invention belongs to the field of highway traffic data analysis and relates to a highway traffic data analysis method based on the ETC system. Background Technology

[0002] With the continuous improvement of the expressway network and the normalized increase in traffic volume, traffic operation management faces increasingly higher requirements for accuracy and efficiency. The ETC system has achieved nationwide network coverage, and it can collect massive amounts of multi-dimensional dynamic traffic data in real time and automatically during vehicle passage. This massive amount of dynamic data provides support for understanding the operation patterns of the road network. Furthermore, through in-depth analysis of ETC data, the accurate prediction or early detection of potential expressway accident hazards has become a key technical requirement for improving traffic safety and ensuring the efficient operation of the road network.

[0003] Currently, some technical solutions are dedicated to using traffic data for highway anomaly identification or accident risk prediction. For example, Chinese patent application CN119720023A discloses a highway anomaly detection method and system based on support vector machine. It processes vehicle trajectory data and ETC gantry data through Kalman filtering and adaptive weighted averaging, introduces Gaussian kernel function to construct SVM model, maps the original feature space to a high-dimensional space to handle nonlinear problems, trains multi-class SVM model using labeled samples, and achieves accurate detection of various anomalies such as traffic congestion, accidents and vehicle malfunctions through a voting mechanism.

[0004] Chinese patent application CN120011861A discloses an accident prediction method and system based on ETC data from the Internet of Things. It analyzes historical traffic information of vehicles to identify risky behaviors, combines historical ETC data of the first section of the highway with vehicle driving information to assess accident risks, and predicts the accident risks of specific road sections through driving history, environmental and road condition data. It can discover potential dangers and issue warnings, help drivers to respond in advance, reduce traffic accidents, and improve highway safety.

[0005] Although the existing solutions mentioned above have made useful explorations in highway anomaly detection or accident risk analysis, in-depth analysis reveals the following significant technical limitations that restrict the accuracy, timeliness, and universality of accident hazard prediction. Specifically, these limitations are: 1. Existing technologies for accident hazard prediction have relatively limited data dimensions, relying heavily on basic vehicle traffic information or local static statistical indicators. They lack in-depth integration of multi-source data such as instantaneous speed sequences and group static characteristics, and are unable to construct a complete risk assessment element system.

[0006] 2. Existing technologies are fragmented in handling the dynamic interaction characteristics of multiple vehicles. Individual behavior analysis and group risk assessment are performed separately. Not only is there a lack of spatiotemporal correlation analysis of speed fluctuations and traffic composition, but the dynamic impact of different vehicle types on regional accident hazards is also not fully quantified, which limits the timeliness and accuracy of prediction results. Summary of the Invention

[0007] In view of this, in order to solve the problems mentioned in the background technology, a highway traffic data analysis method based on the ETC system is proposed.

[0008] The objective of this invention can be achieved through the following technical solution: This invention provides a highway traffic data analysis method based on the ETC system, comprising: acquiring the original vehicle passage data stream collected by all ETC gantries within a preset time period of the monitored road segment, wherein the original vehicle passage data stream includes vehicle identification, passage timestamp, gantry location code and vehicle type code.

[0009] The original vehicle data stream is analyzed to generate individual dynamic features of vehicle speed and static features of vehicle type groups, and a multidimensional risk dataset is constructed with road segment identification and time window as indexes.

[0010] The multidimensional risk dataset is input into a pre-trained deep learning model, which outputs the probability value of accident hazards and performs early warning status determination based on the probability gradient determination rule.

[0011] Based on the early warning status, an accident risk heat map of the monitored road section is generated and fed back.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention integrates the dynamic characteristics of individual vehicle speed and the static characteristics of vehicle group, introduces a deep learning model, deeply explores the implicit mapping law between the two types of features and accident risks, and combines the feature aggregation mechanism of dynamic sliding time window and probability gradient judgment rule to achieve accurate identification of the cumulative growth and sudden change of traffic risks, thereby improving the accuracy of accident hazard prediction.

[0013] (2) This invention provides traffic management departments with an intuitive decision support tool by visualizing regional accident hazard prediction maps, which helps to take measures in advance to reduce the accident rate.

[0014] (3) The present invention allows the model to be incrementally trained based on the actual accident situation through the feedback correction step, ensuring that the system maintains high accuracy in the long term, which is in line with the development direction of intelligent transportation. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the implementation steps of the method of the present invention.

[0017] Figure 2 This is a schematic diagram illustrating the construction logic of the multidimensional risk dataset of this invention.

[0018] Figure 3 This is a schematic diagram of the probability gradient determination rule logic of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 As shown, the present invention provides a highway traffic data analysis method based on the ETC system, including: S11. Obtaining the original vehicle passage data stream collected by all ETC gantries within a preset time period of the monitored road segment, wherein the original vehicle passage data stream includes vehicle identification, passage timestamp, gantry location code and vehicle type code.

[0021] In a preferred embodiment of the present invention, the process of acquiring the original vehicle passage data stream includes: establishing a data communication link with the highway data center to receive vehicle passage data packets uploaded by all ETC gantries in the monitored road section in real time.

[0022] The target fields of the vehicle data message are located, extracted, and transformed accordingly. The target fields include vehicle identification field, timestamp field, gantry physical location field, and vehicle type classification field.

[0023] It should be added that the above-mentioned corresponding data conversion process specifically includes the following: extracting the original electronic tag or license plate recognition string from the vehicle identification field as the vehicle identification.

[0024] Locate the timestamp field and convert it to a standard time format value to serve as the pass timestamp.

[0025] Extract the three-dimensional coordinate data of the gantry physical location field and match the corresponding location code by querying the preset gantry code library.

[0026] Read the vehicle type classification field and decode it into a standardized vehicle type code based on a predefined vehicle type classification dictionary.

[0027] The converted field data is input into the data buffer for serialization processing. The processing steps are as follows: (a) The data records are reordered according to the timestamp order.

[0028] (b) Perform field integrity checks and mark missing data items.

[0029] (c) For the marked missing data items, perform interpolation repair based on the passage records of the same vehicle in adjacent gantries.

[0030] The output is an ordered data stream after integrity restoration, which serves as the original vehicle passage data stream.

[0031] It should be noted that the above field integrity verification process is as follows: the vehicle identification, passage timestamp, gantry location code and vehicle type code are regarded as the set of required fields. Data records are scanned one by one. When a required field is found to be empty or in an illegal format, it is confirmed that there is a field missing phenomenon. At the same time, a tag tuple containing the missing field type and record location is generated.

[0032] The interpolation repair process for the missing data items is as follows: locate the vehicle and gantry position to which the missing field belongs based on the marked tuple, retrieve the complete passage records of the same vehicle at adjacent gantries in time and space, and perform differentiated interpolation repair based on the missing field type: i. If the vehicle identifier or vehicle type code is missing, copy the field value corresponding to the nearest valid gantry of the same vehicle.

[0033] ii. If a passage timestamp is missing, extract the passage timestamps of the two nearest valid gantries before and after the gantry to which the missing field belongs, as well as the actual road distance between the gantries. Calculate the ratio between the sum of the distance from the preceding adjacent gantry to the gantry to which the missing field belongs and the distance from the gantry to the following adjacent gantry, and the difference between the passage timestamps of the preceding and following gantries to obtain the average driving speed of the vehicle in the two segments. Quantify the theoretical passage time of the vehicle from the preceding adjacent gantry to the gantry to which the missing field belongs based on the average speed. Add the duration to the passage timestamp of the preceding gantry to obtain the repaired value corresponding to the missing passage timestamp. The theoretical passage time can be checked for reasonableness. If the deviation from the average passage time of the vehicle on the same road segment or the same distance segment in history exceeds a preset permissible deviation threshold, the historical average speed can be retrieved and recalculated.

[0034] iii. If a gantry location code is missing, retrieve the three-dimensional coordinate data of the preceding and following adjacent gantry frames of the missing gantry. Obtain the gantry design and layout guidelines for the monitored road segment, and acquire the standard spacing and curve adjustment parameters for gantry layout in that segment. Calculate the theoretical layout spacing based on the three-dimensional coordinate data of the preceding and following adjacent gantry frames, and compare it with the design standard spacing. If the deviation rate between the theoretical layout spacing and the design standard spacing is less than or equal to the preset allowable rate, use linear interpolation to calculate the three-dimensional coordinates of the missing gantry. Otherwise, combine the curve adjustment parameters for curve interpolation correction. Spatially match the calculated three-dimensional coordinates of the missing gantry with the preset gantry code library, and retrieve the gantry location code with the smallest spatial error distance as the missing gantry location repair code.

[0035] The deviation rate between the theoretical spacing of the gantry and the design standard spacing is specifically the ratio of the absolute difference between the theoretical spacing of the gantry and the design standard spacing to the design standard spacing.

[0036] The preset permission deviation threshold and preset permission rate value are calibrated through statistical analysis experiments based on massive historical normal traffic data. That is, big data mining is performed on the traffic characteristics of different road sections, vehicle types, and time periods to determine the deviation range covering normal traffic scenarios as the threshold benchmark.

[0037] S12. Parse the original vehicle data stream to generate individual dynamic features of vehicle speed and static features of vehicle type groups, and construct a multidimensional risk dataset indexed by road segment identification and time window.

[0038] In a preferred embodiment of the present invention, the process of analyzing individual vehicle speed dynamic features includes:

[0039] Filter the records of at least two adjacent gantries that the target vehicle passes through in succession on the monitored road segment according to the passage timestamp order.

[0040] Based on the gantry location code, the spatial distance between adjacent gantries is queried. Combined with the difference in the timestamps of passage between adjacent gantries, the average driving speed of the target vehicle in the interval between two adjacent gantries is determined, and the average driving speed is used as the instantaneous speed estimate at the end gantry of the interval.

[0041] Iterate through all continuous gantry sections that the target vehicle passes through in the monitored road segment, and generate its instantaneous speed set in chronological order.

[0042] In a preferred embodiment of the present invention, the process of analyzing individual dynamic features of vehicle speed further includes: setting a sliding time window based on the dynamic configuration of real-time traffic density.

[0043] It should be noted that the dynamic configuration of the aforementioned sliding time window is based on real-time traffic density as the core adjustment criterion. The specific mechanism is as follows: Real-time traffic density of the monitored road segment is statistically analyzed using gantry monitoring data. A negative correlation mapping rule between traffic density and window duration is established, dividing traffic density into three levels: low, medium, and high. Each level corresponds to a specific density value range and a preset window duration. The higher the level, the denser the traffic flow, and the shorter the corresponding window duration. In practical applications, the system matches the real-time traffic density to the appropriate level and automatically calls the preset window duration parameter for that level. This hierarchical dynamic adjustment mechanism can capture vehicle speed characteristics under different traffic conditions in a timely manner through duration adaptation, and can also avoid frequent jumps in window duration due to small fluctuations in traffic flow through level range buffering. This ensures the timeliness of feature extraction while maintaining the continuity of the feature sequence, ultimately improving the scenario adaptability and analysis accuracy of individual vehicle speed dynamic features.

[0044] Calculate the statistical distribution index of the instantaneous velocity set within the window. The statistical distribution index includes standard deviation, variance, and coefficient of variation.

[0045] The statistical distribution indicators are integrated into a feature vector to characterize the individual dynamic characteristics of vehicle speed.

[0046] It should be added that the above-mentioned standard deviation, variance, and coefficient of variation are the core indicators for characterizing the dynamic characteristics of individual vehicle speeds. Each of them quantifies the fluctuation characteristics of vehicle speed from different dimensions. The selection criteria are as follows: Variance reflects the overall amplitude of vehicle speed fluctuations in absolute terms. The larger the value, the more significant the deviation of the speed from the average level. The calculation method of its squared term will amplify the impact of extreme speed values ​​and more sensitively capture the abrupt changes in driving behavior to measure whether there are frequent sudden changes in speed during vehicle operation.

[0047] Standard deviation and variance are both absolute discrete indicators, but relative variance more intuitively reflects the average range of speed fluctuations, thus characterizing the regular fluctuation of vehicle speed.

[0048] The coefficient of variation is the ratio of the standard deviation to the mean. It eliminates the influence of the speed mean order of magnitude and focuses on the relative intensity of the fluctuation to compare driving stability at different speed levels.

[0049] Variance and standard deviation describe the magnitude of vehicle speed fluctuations in absolute terms, while the coefficient of variation describes the severity of fluctuations in relative terms. Together, they construct multi-dimensional individual dynamic characteristics of vehicle speed, which can reflect both the absolute stability of driving behavior and the relative stability under different scenarios, providing a quantitative basis for distinguishing individual driving styles.

[0050] In a preferred embodiment of the present invention, the static feature analysis process of the vehicle group includes:

[0051] The total number of vehicles entering the monitored road section within the preset time period is counted.

[0052] Based on the vehicle model code classification, the number of each type of vehicle passing through is statistically analyzed, and the proportion of each type of vehicle's traffic to the total number of vehicles is calculated to form a vehicle model distribution vector, which is used as a static feature of the vehicle model group.

[0053] See Figure 2 As shown, in a preferred embodiment of the present invention, the process of constructing the multidimensional risk dataset includes:

[0054] The monitored road segment is divided into multiple continuous sub-segments according to a fixed spatial distance, and each sub-segment is uniquely identified by its center mileage marker.

[0055] The preset time period is divided into multiple consecutive time slices according to the sliding time window configuration, and each time slice is uniquely identified by its starting timestamp.

[0056] For each spatiotemporal cell defined by sub-segments and time slices, aggregate the individual dynamic characteristics of vehicle speed and the group static characteristics of vehicle type for all vehicles within that cell.

[0057] It should be added that the static characteristics of vehicle groups in the spatiotemporal cell are not calculated based on the overall preset time period, but are counted independently for each sliding time window. That is, the number of each type of vehicle passing through the window is counted and the proportion of each type to the total number of vehicles in the window is calculated. This serves as the basis for the composition of the static characteristics of vehicle groups in the corresponding spatiotemporal cell, which helps to more accurately reflect the differences in vehicle distribution in different time slices. This keeps the characteristics of the spatiotemporal cell synchronized with the dynamically changing traffic flow, and avoids masking short-term fluctuations in vehicle structure due to the averaging of the overall time period.

[0058] The aggregated features are normalized to form feature vectors, and the feature vectors of all spatiotemporal cells are integrated to form a multidimensional risk dataset.

[0059] S13. Input the multidimensional risk dataset into the pre-trained deep learning model, output the probability value of accident hazards, and perform early warning status determination based on the probability gradient determination rule.

[0060] In a preferred embodiment of the present invention, the pre-trained deep learning model construction process includes: acquiring historical vehicle data streams and corresponding accident label datasets, wherein the label is 1 when an accident occurs and 0 when no accident occurs.

[0061] The historical vehicle data stream is converted into a historical multidimensional risk dataset, and the historical multidimensional risk dataset is divided into an accident sample set and a non-accident sample set based on accident labels.

[0062] Calculate the distribution statistics of each feature dimension in the accident sample set and the non-accident sample set respectively, quantify the distribution difference of each feature dimension under accident and non-accident conditions, and generate a feature correlation strength vector.

[0063] It should be noted that the quantification process for the distribution difference of each feature dimension under accident and non-accident conditions is as follows: the distribution difference on the static feature dimension of the vehicle group is quantified by calculating the JS divergence between the vehicle distribution vectors of the accident sample set and the non-accident sample set. The specific calculation process of the JS divergence is as follows: the vehicle distribution vector is a probability distribution function, where... Represent the accident sample set and the non-accident sample set, respectively. The average percentage of different car models For the numbering of various vehicle models, ,Pick The average distribution is used as the intermediate probability distribution function That is, the median percentage for each vehicle model is the average of the percentages of accident samples and non-accident samples, obtained through... Formula quantification relatively The KL divergence, similarly, can be calculated using the formula... Replace with Quantification relatively The KL divergence is calculated by averaging the two KL divergences to obtain the JS divergence between the vehicle model distribution vectors of the accident sample set and the non-accident sample set. The JS divergence ranges from [0, 1]. The closer the value is to 1, the greater the difference in the proportion distribution of vehicle models in the two sample sets.

[0064] The Wasserstein distance is calculated between the probability distributions of the aggregated vehicle speed statistical feature vectors of accident sample sets and non-accident sample sets to measure the distributional differences in the individual dynamic feature dimension of vehicle speed. The specific calculation process of the Wasserstein distance is as follows: the vehicle speed statistical feature vectors are aggregated into a probability distribution function in three-dimensional Euclidean space, and the Euclidean distance is used as the transportation cost metric. The minimum total transportation cost that transforms the accident sample distribution into the non-accident sample distribution is retrieved as the Wasserstein distance.

[0065] The above-mentioned feature association strength vector generation process is as follows: weights are assigned to the distribution differences of each feature dimension under accident and non-accident conditions, and numerical corrections are performed. After correction, a two-dimensional vector is formed as the feature association strength vector. The weight allocation of each feature dimension can be set according to industry experience, or it can be obtained through a limited number of historical test data. For example, historical data of individual dynamic characteristics of vehicle speed and static characteristics of vehicle type group at each occurrence of an accident on the monitored road section are collected first. Regression analysis or logistic regression analysis is used to determine the contribution of vehicle speed and vehicle type to the occurrence of the accident. Finally, after normalization, the contribution is converted into weights and their sum is 1.

[0066] A time series deep learning model architecture is constructed, and the feature association strength vector is embedded in the hidden layer as a prior constraint.

[0067] Using historical multidimensional risk datasets as model input and binary accident labels as training targets, the model outputs probability predictions through forward propagation. The model network weight parameters are iteratively updated by minimizing the cross-entropy loss function between the probability predictions and the binary labels until the consistency index between the probability predictions output on the validation set and the actual accident status reaches a preset threshold, at which point training terminates.

[0068] It should be noted that the process of calculating the probability prediction value through forward propagation is as follows: the input layer receives the historical multidimensional risk dataset of spatiotemporal cells and converts it into a tensor format that the model can process.

[0069] The embedding layer combines the feature association strength vector to perform weighted encoding of the input features.

[0070] The weighted feature input hidden layer captures the dynamic dependencies in the time series through a gating mechanism, namely the coordinated changes in vehicle speed fluctuations and vehicle type distribution within continuous time slices.

[0071] The output layer uses the sigmoid activation function to map the output of the hidden layer to the interval [0, 1] to obtain the predicted value of the accident hazard probability.

[0072] See Figure 3 As shown, in a preferred embodiment of the present invention, the probability gradient determination rule includes the following:

[0073] Set an observation window that includes a preset number of consecutive time slices to output the probability value of potential accidents.

[0074] The rate of change of the relative probability values ​​of potential accidents between adjacent time slices within the observation window is calculated using the finite difference method to generate a gradient sequence.

[0075] An early warning state is triggered when multiple consecutive gradient values ​​in the gradient sequence exceed a preset positive gradient threshold or a single gradient value exceeds a sudden change standard gradient threshold.

[0076] It should be added that the core of setting the warning state triggering conditions is based on two typical laws of traffic risk evolution: the cumulative growth of risk and sudden jump. The continuous positive gradient condition focuses on chronic risks, solving the problem of risks slowly deteriorating but easily overlooked, while the sudden change gradient condition focuses on acute risks, solving the problem of risks suddenly appearing but deteriorating rapidly, so as to achieve full-scenario coverage of traffic risks. Among them, the preset positive gradient threshold is relatively small compared with the sudden change standard gradient threshold.

[0077] This invention integrates the dynamic characteristics of individual vehicle speeds and the static characteristics of vehicle types, introduces a deep learning model, and deeply explores the implicit mapping patterns between these two types of features and accident risks. By combining a feature aggregation mechanism with a dynamic sliding time window and a probability gradient judgment rule, it achieves accurate identification of the cumulative growth and sudden changes in traffic risks, thereby improving the accuracy of accident hazard prediction.

[0078] S14. Generate and feed back the accident risk heat map of the monitored road section based on the early warning status.

[0079] In a preferred embodiment of the present invention, the process of generating the accident risk heat map of the monitored road section includes: defining the probability value of accident hazards and discrete risk level, as well as the mapping rules between discrete risk level and visual display color, wherein the discrete risk level includes low risk, medium risk, high risk and extremely high risk.

[0080] For the latest accident hazard probability value of each sub-segment in the monitored road segment, the discrete risk level is queried from the mapping rules and the visual display color is determined.

[0081] The corresponding sub-segment grids are rendered using the determined colors on the geographic layer of the monitored road segment, and the probability values ​​and risk levels of accident hazards are marked. In the early warning state, a flashing effect is additionally displayed, and an accident risk heat map of the monitored road segment is generated in real time.

[0082] This invention provides traffic management departments with an intuitive decision support tool by creating a visualized regional accident hazard prediction map, which helps to take measures in advance to reduce the accident rate.

[0083] In a preferred embodiment of the present invention, the method further includes: upon receiving a real accident alarm signal, performing spatiotemporal matching and comparative analysis between the actual traffic event information and the accident risk heat map prediction results of the monitored road segment.

[0084] Quantitatively evaluate the prediction bias of the deep learning model in real-world scenarios.

[0085] It should be added that the above prediction deviation quantification process includes: obtaining whether the high-risk and extremely high-risk sub-segments in the accident risk heatmap corresponding to the actual occurrence timestamp of the traffic incident cover the actual occurrence location of the traffic incident. If not, the distance between the actual occurrence location of the traffic accident and the nearest predicted high-risk or extremely high-risk area is obtained. The ratio of this distance to a fixed spatial distance is then superimposed with the difference between the probability prediction value of the sub-segment where the actual occurrence location of the traffic accident is located and 1. The superimposed calculation result is used as the prediction deviation. Conversely, if the sub-segment is covered, the difference between the probability prediction value and 1 is used as the prediction deviation.

[0086] If the prediction deviation exceeds a preset tolerance threshold, new, verified accident sample data will be added to the training set to trigger incremental training of the deep learning model.

[0087] The embodiments of the present invention allow the model to be incrementally trained based on actual accident situations through a feedback correction step, ensuring that the system maintains high accuracy over a long period of time, which is in line with the development direction of intelligent transportation.

[0088] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A method for analyzing highway traffic data based on an ETC system, characterized by, The method comprises the following steps: acquiring all original vehicle passing data streams collected by ETC gantries in a preset period of time on a monitored road section, wherein the original vehicle passing data streams comprise vehicle identification, passing time stamp, gantry position code and vehicle type code; analyzing the original vehicle passing data streams to generate individual dynamic characteristics of vehicle speed and static characteristics of vehicle type group, and constructing a multi-dimensional risk data set indexed by road section identification and time window; inputting the multi-dimensional risk data set into a pre-trained deep learning model to output an accident risk probability value result and execute early warning state judgment based on a probability gradient judgment rule; the pre-trained deep learning model construction process comprises: acquiring historical vehicle passing data streams and corresponding accident label data set, wherein the label is 1 when an accident occurs and 0 when no accident occurs; converting the historical vehicle passing data streams into a historical multi-dimensional risk data set, and dividing the historical multi-dimensional risk data set into an accident sample set and a non-accident sample set based on the accident label; calculating the distribution statistical indicators of each feature dimension in the accident sample set and the non-accident sample set respectively, quantifying the distribution difference degree of each feature dimension in the accident and non-accident states, and generating a feature correlation strength vector; constructing a time series deep learning model architecture and embedding the feature correlation strength vector in the hidden layer as a prior constraint; taking the historical multi-dimensional risk data set as the model input and the binary accident label as the training target, calculating the probability prediction value of the model output through forward propagation, and iteratively updating the model network weight parameters by means of the cross-entropy loss function of minimizing the probability prediction value and the binary label until the correlation consistency index of the probability prediction value output by the verification set and the actual accident state reaches a preset threshold value; the feature correlation strength vector generation process is: assigning weights to the distribution difference degrees of each feature dimension in the accident and non-accident states to perform numerical correction, and the corrected two-dimensional vector is used as the feature correlation strength vector, wherein the weight distribution of each feature dimension can be set according to industry experience or obtained through limited historical test data; the distribution difference degree in the vehicle type group static feature dimension is quantified by calculating the JS divergence between the vehicle type distribution vectors of the accident sample set and the non-accident sample set, and the distribution difference degree in the vehicle speed individual dynamic feature dimension is measured by calculating the Wasserstein distance between the probability distributions aggregated by the vehicle speed statistical feature vectors of the accident sample set and the non-accident sample set; generating an accident risk heat map of the monitored road section according to the early warning state and feeding back. 2.The ETC system-based expressway traffic data analysis method of claim 1, wherein, The original vehicle passing data stream acquisition process comprises: establishing a data communication link with a highway data center to receive vehicle passing data messages uploaded by all ETC gantries in the monitored road section in real time; performing target field positioning extraction and corresponding data conversion processing on the vehicle passing data messages, wherein the target fields include vehicle identification field, time stamp field, gantry physical position field and vehicle type classification field; The converted field data is input into a data buffer area to perform serialization processing, and the processing steps are as follows: (a) reordering the data records in timestamp order; (b) performing field integrity checking and marking missing data items; (c) for the marked missing data items, performing interpolation repair based on the passing records of the same vehicle at adjacent gantries; outputting the integrity-repaired ordered data stream as the original passing vehicle data stream. 3.The ETC system-based expressway traffic data analysis method of claim 1, wherein, The vehicle speed individual dynamic feature analysis process includes: filtering at least two adjacent gantry records of the target vehicle continuously passing through the monitored road section in time stamp order; based on the gantry position coding, querying the spatial distance between adjacent gantries, combining the adjacent gantry passing time stamp difference, determining the average driving speed of the target vehicle in the interval between the two adjacent gantries, and taking the average driving speed as the instantaneous speed estimation value at the interval end gantry; traversing all continuous gantry intervals of the target vehicle passing through the monitored road section, and generating its instantaneous speed set in time order. 4.The ETC system-based expressway traffic data analysis method of claim 3, wherein, The vehicle speed individual dynamic feature analysis process further includes: setting a sliding time window dynamically configured based on real-time traffic density; calculating statistical distribution indicators of the instantaneous speed set within the window, the statistical distribution indicators including standard deviation, variance and coefficient of variation; integrating the statistical distribution indicators into a feature vector to represent the vehicle speed individual dynamic feature. 5.The ETC system-based expressway traffic data analysis method of claim 1, wherein, The vehicle type group static feature analysis process includes: counting the total number of vehicles entering the monitored road section within a preset period; counting the number of each type of vehicle passing according to the vehicle type coding, calculating the proportion of the number of each type of vehicle passing to the total number of vehicles to form a vehicle type distribution vector, and taking it as the vehicle type group static feature. 6.The ETC system-based expressway traffic data analysis method of claim 4, wherein, The construction process of the multi-dimensional risk data set includes: dividing the monitored road section into multiple continuous sub-road sections according to fixed spatial distance, and taking the central milestone number of each sub-road section as the unique identifier; dividing the preset period into multiple continuous time slices according to the sliding time window configuration, and taking the start time stamp of each time slice as the unique identifier; for each space-time cell defined by a sub-road section and a time slice, aggregating the vehicle speed individual dynamic feature and the vehicle type group static feature of all vehicles in the cell; normalizing the aggregated features to form a feature vector, and integrating the feature vectors of all space-time cells to form a multi-dimensional risk data set. 7.The method of claim 6, wherein, The probability gradient judgment rule includes the following contents: setting an observation window containing a continuous preset number of time slice output accident risk probability values; calculating the change rate of the relative accident risk probability values of adjacent time slices in the observation window by difference method to generate a gradient sequence; when there are continuous multiple gradient values exceeding the preset positive gradient threshold value or a single gradient value exceeding the sudden change standard gradient threshold value in the gradient sequence, triggering the warning state. 8.The ETC system-based expressway traffic data analysis method of claim 6, wherein, The generation process of the monitored road section accident risk heat map includes: defining the mapping rules of accident risk probability value and discrete risk level, and discrete risk level and visual display color, the discrete risk level including low risk, medium risk, high risk and extremely high risk; for each sub-road section in the monitored road section, querying its discrete risk level and determining the visual display color from the mapping rules according to the latest accident risk probability value. The corresponding sub-road segment grid is rendered using the determined color on the monitoring road segment geographic layer, the accident risk probability value and the risk level text are marked, and the flashing effect is additionally superimposed on the early warning state, and the accident risk heat map of the monitoring road segment is generated in real time. 9.The ETC system-based expressway traffic data analysis method of claim 1, wherein, Also include: When receiving a real accident alarm signal, the actual traffic event information is spatiotemporally matched and compared with the prediction result of the accident risk heat map of the monitoring road segment; Quantitative evaluation of the prediction deviation of the deep learning model in the real scene; If the prediction deviation exceeds the preset tolerance threshold, new verified accident sample data is added to the training set to trigger incremental training of the deep learning model.

Images