Abnormal traffic event identification method and system based on traffic large model

Abstract

The invention relates to the technical field of traffic event identification, and discloses an abnormal traffic event identification method and system based on a traffic large model, and the method comprises the steps: obtaining a traffic data flow, extracting an abnormal feature vector, and obtaining an abnormal signal candidate set; grouping the candidate sets and calculating a deviation degree, and if the deviation degree exceeds a threshold value, taking the deviation degree as a risk signal to form an input subset; environment variables are extracted from the subsets, a mapping relation is established, and anomaly recognition embedding representation is obtained; classifying the embedded representation, judging a congestion precursor and generating an early warning signal to obtain an early warning signal sequence; matching the sequence to obtain an abnormal event chain; if the integrity is higher than a threshold value, analyzing the type to obtain an abnormal event type; extracting a correlation feature vector from the type, pushing the correlation feature vector to a traffic management platform to obtain an instruction, and obtaining an emergency response trigger instruction sequence; and executing the instruction sequence to extract a feedback data stream, inputting the traffic large model to judge the accuracy rate, and if the judgment accuracy rate is met, determining an optimized anomaly recognition framework. The method can solve the problem of insufficient early warning capability.

Description

Technical Field

[0001] This invention relates to the field of traffic incident recognition technology, and in particular to a method and system for identifying abnormal traffic incidents based on a large traffic model. Background Technology

[0002] With the acceleration of urbanization, the complexity of transportation systems is increasing day by day. How to effectively deal with emergencies in traffic has become an issue that cannot be ignored. The application of satellite positioning technology can accurately track vehicle dynamics and grasp the traffic flow distribution of each section of the road network in real time, providing strong support for traffic management departments to efficiently deal with congestion, accidents and other situations.

[0003] In one existing technology, a passive architecture of sensor acquisition-static threshold judgment-human-assisted response is adopted. With fixed detection equipment as the core, induction coil detectors are buried at specific locations on the road to collect statistics on traffic flow and average speed, or simulated cameras are installed for human visual observation. Limited-dimensional data is collected, and fixed thresholds are set for the collected data. When the threshold is exceeded, an alarm signal is generated and uniformly sent to the traffic control center for manual identification and handling by maintenance personnel.

[0004] However, existing technologies rely solely on coils and analog cameras, resulting in limited coverage, a single data dimension, and an inability to capture deep connections among traffic participants. The use of fixed-frequency data collection and manual response leads to delayed anomaly identification, and static thresholds cannot distinguish between normal traffic conditions and abnormal events, resulting in false alarms. In summary, existing technologies suffer from insufficient early warning capabilities. Summary of the Invention

[0005] This invention provides a method and system for identifying abnormal traffic events based on a large traffic model, in order to solve the problem of insufficient early warning capabilities in existing technologies.

[0006] Firstly, to address the aforementioned technical problems, this invention provides a method for identifying abnormal traffic events based on a large traffic model, comprising: Acquire real-time traffic data streams and extract abnormal feature vectors to obtain a candidate set of abnormal signals; The abnormal signal candidate set is grouped and the trajectory deviation of the vehicle trajectory data within the group is calculated. If the trajectory deviation exceeds a preset deviation threshold, it is judged as a potential risk signal and used as an input subset to obtain the correlation analysis input subset. The surrounding environmental variables are extracted from the input subset of the correlation analysis, and a mapping relationship is established based on the traffic density and intersection interaction data in the surrounding environmental variables to obtain the anomaly identification embedded representation. The anomaly identification embedding is used for classification, and the classification results are used to determine whether it is a precursor to congestion risk and generate a corresponding warning signal to obtain a warning signal sequence. The abnormal event chain is obtained by performing time-series matching on the warning signal sequence; Calculate the chain integrity of the abnormal event chain. If the chain integrity is higher than a preset integrity threshold, analyze the type of abnormal event to obtain the final abnormal event type. Extract the associated feature vector from the final abnormal event type and push it to the pre-established traffic management platform to obtain instructions, thus obtaining the emergency response trigger instruction sequence; The trigger command sequence is executed, the feedback data stream is extracted and input into the pre-trained traffic model, and the accuracy of the model output is judged. If the accuracy meets the conditions, the optimized anomaly recognition framework is determined.

[0007] Secondly, the present invention provides an abnormal traffic event identification system based on a large traffic model, comprising: The data acquisition module is used to acquire real-time traffic data streams and extract abnormal feature vectors to obtain a candidate set of abnormal signals; The data grouping module is used to group the candidate set of abnormal signals and calculate the trajectory deviation of the vehicle trajectory data within the group. If the trajectory deviation exceeds a preset deviation threshold, it is judged as a potential risk signal and used as an input subset to obtain the correlation analysis input subset. The data mapping module is used to extract surrounding environmental variables from the correlation analysis input subset, establish a mapping relationship based on traffic density and intersection interaction data in the surrounding environmental variables, and obtain anomaly identification embedding representation; The data generation module is used to classify the anomaly identification embedded representation, determine whether it is a precursor to congestion risk based on the classification result, and generate a corresponding early warning signal to obtain an early warning signal sequence. The data matching module is used to perform time-series matching on the warning signal sequence to obtain an abnormal event chain; The data calculation module is used to calculate the chain integrity of the abnormal event chain. If the chain integrity is higher than a preset integrity threshold, the type of abnormal event is analyzed to obtain the final abnormal event type. The data extraction module is used to extract the associated feature vectors from the final abnormal event type and push them to the pre-established traffic management platform to obtain instructions and get the emergency response trigger instruction sequence; The data judgment module is used to execute the trigger instruction sequence, extract the feedback data stream, input it into the pre-trained traffic model, judge the accuracy of the model output results, and determine the optimized anomaly recognition framework if the accuracy meets the conditions.

[0008] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the abnormal traffic event identification method based on a large traffic model as described above.

[0009] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the abnormal traffic event identification method based on the large traffic model described above.

[0010] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention establishes a dynamic baseline based on historical data and combines multiple features such as "speed-flow-spatial distribution" for collaborative judgment; it groups abnormal signals by K-means clustering, calculates trajectory deviation, accurately screens potential risk signals, avoids the oversensitivity problem of traditional solutions that alarm for a single vehicle abnormality, and reduces the false alarm rate.

[0011] (2) This invention automates the process of real-time analysis of traffic data streams, extraction of abnormal signal candidate sets, and generation of early warning sequences, eliminating the need for manual intervention and shortening response time. Through linkage with the traffic management platform, it automatically transforms abnormal feature vectors into triggering instructions that are "time-ordered and clearly defined in terms of responsibility" and tracks execution feedback, avoiding the inefficiency of the traditional solution of "manual dispatching - information gap", improving emergency response efficiency and early warning capabilities. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the abnormal traffic event identification method based on a large traffic model provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the abnormal traffic event identification system based on a large traffic model provided in the second embodiment of the present invention. Detailed Implementation

[0013] 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.

[0014] Reference Figure 1 The first embodiment of the present invention provides a method for identifying abnormal traffic events based on a large traffic model, comprising the following steps: S11: Obtain real-time traffic data stream and extract abnormal feature vectors to obtain a candidate set of abnormal signals; S12, group the candidate abnormal signals and calculate the trajectory deviation of the vehicle trajectory data within the group. If the trajectory deviation exceeds the preset deviation threshold, it is judged as a potential risk signal and used as an input subset to obtain the correlation analysis input subset. S13, extract surrounding environmental variables from the input subset of the correlation analysis, establish a mapping relationship based on the traffic density and intersection interaction data in the surrounding environmental variables, and obtain the anomaly identification embedded representation; S14, classify according to the anomaly identification embedding representation, and determine whether it is a precursor to congestion risk based on the classification result and generate a corresponding warning signal to obtain a warning signal sequence; S15, perform time-series matching on the warning signal sequence to obtain the abnormal event chain; S16, calculate the chain integrity of the abnormal event chain. If the chain integrity is higher than a preset integrity threshold, analyze the type of abnormal event and obtain the final abnormal event type. S17. Extract the associated feature vector from the final abnormal event type and push it to the pre-established traffic management platform to obtain instructions and get the emergency response trigger instruction sequence; S18, execute the trigger instruction sequence, extract the feedback data stream, and input it into the pre-trained traffic model. Determine the accuracy of the model output results. If the accuracy meets the conditions, determine the optimized anomaly recognition framework.

[0015] In step S11, real-time traffic data streams are acquired, and abnormal feature vectors are extracted to obtain a candidate set of abnormal signals, including: Vehicle trajectory data and traffic flow indicators are extracted from the real-time traffic data stream, and abnormal features are filtered out to obtain an abnormal feature set. Determine the individual behavior patterns of vehicle trajectory data in the abnormal feature set. When the individual behavior pattern deviates from the preset normal behavior pattern, mark it as a potential abnormal signal to obtain an initial abnormal signal set. The initial set of abnormal signals is compared and verified a second time to obtain a precise set of abnormal signals; From the precise set of abnormal signals, abnormal signals that meet the preset abnormal feature criteria are sorted out to obtain a candidate set of abnormal signals.

[0016] It should be noted that multiple sensor nodes deployed on urban main roads collect information such as vehicle speed, location, and timestamps in real time to obtain real-time traffic data streams. These real-time traffic data streams include vehicle trajectory data and traffic flow index data. Using a dynamic baseline approach, features that deviate from the normal range are filtered from the vehicle trajectory data and traffic flow index data to obtain an abnormal feature set. A dynamic baseline is established based on historical data from the same period over the past 7 days. For example, if the average headway of a vehicle during the morning rush hour on a certain road segment is 2 seconds, a fluctuation of ±0.5 seconds is allowed. When the real-time data deviates from the baseline by more than 20%, it is determined to be an abnormal feature.

[0017] K-means clustering algorithm is used to cluster vehicle trajectories of a certain road segment over the past month and to statistically analyze the frequency of individual behaviors under normal traffic conditions, thus obtaining individual behavior patterns. For example, for vehicles traveling east to west on main road A, the trajectory x-coordinate should be in the range of 90-210, and the y-coordinate should be in the range of 200-220. For each trajectory data in the abnormal feature set, its deviation from the normal behavior pattern is calculated. Euclidean distance is used to calculate the distance between the real-time trajectory and the cluster center of the normal path. If the distance is greater than 5 meters (a preset deviation threshold, set based on the 95% confidence interval of historical data over the past 90 days; the trajectory of vehicles during normal driving will form clusters around the optimal path, and the boundary of the 95% confidence interval can reflect the range of the vast majority of normal trajectories, i.e., set to 5 meters), the 5-meter threshold covers the range of normal trajectories. The system identifies lane changes and minor deviations within the normal range, as well as significant deviations caused by accidents or illegal lane changes, and classifies them as trajectory deviations. For example, if a vehicle's trajectory x-coordinate suddenly changes to 230 (exceeding the normal range of 90-210), and the Euclidean distance reaches 25 meters, it is considered a deviation. The system also counts the number of individual behaviors per unit time. If the number exceeds the baseline threshold (the target indicator data from the same period of the past 7 days is selected, and the average value is calculated as the dynamic baseline; the deviation ratio is set to 10%, and the threshold exceeding the baseline is equal to the dynamic baseline multiplied by (1 plus the deviation ratio)), it is considered an abnormal behavior. For example, if a vehicle changes lanes 5 times in 10 minutes (exceeding the normal threshold of 2 times / minute), it is marked as a potential abnormal signal. Based on the abnormal feature set, deviation type (trajectory / behavior), deviation degree value, and normal baseline value data are added to obtain the initial abnormal signal set.

[0018] The K-means clustering algorithm is implemented as follows: A preset K value is established based on the spatial and temporal granularity of the traffic scenario. For example, for abnormal signals on urban main roads, clustering units are divided according to "every 200-meter road segment + 15-minute time window". If the target area contains 6 such units, then K=6. K initial centers are randomly selected from the data to be clustered (such as the geographical coordinates and timestamps of the abnormal signals). For example, the coordinates (100, 200), (150, 205), and (200, 210) of three representative abnormal signals during the morning rush hour are selected as initial centers. The distance between the data and the cluster centers is calculated using Euclidean distance. The Euclidean distance calculation formula is: , , For the coordinates of the data points, , Using the coordinates of the cluster centers, assign each data point to the nearest cluster group. Calculate the mean coordinates (or mean timestamps) of all data points within each group and use this as the new cluster center. Repeat the steps to calculate the distance between the data points and the cluster centers and iteratively update the cluster centers until the cluster center positions no longer change, then output the final grouping results.

[0019] Based on the initial abnormal signal set, secondary verification is performed from three dimensions: duration, historical correlation, and external factors. The duration of the abnormal signal is monitored. If the abnormality lasts for less than 3 minutes (such as a brief deceleration caused by a vehicle temporarily avoiding a pedestrian), it is judged as a normal fluctuation and removed from the initial set. For example, if a vehicle's speed drops to 28 km / h and only lasts for 2 minutes before recovering, it is excluded. Historical data of the vehicle / road segment is retrieved. If the vehicle has not exhibited similar abnormal behavior in the past 3 months (such as never having changed lanes frequently), or if there is no such abnormality in the historical data of the road segment during the same period (such as the first concentrated low speed at an intersection), the signal is retained. Conversely, if a vehicle frequently slows down temporarily in the road segment (historical data shows 3 times per month), it is judged as an individual habit and is excluded. The construction / weather database of the traffic management platform is connected. If there is a temporary construction notice at the location of the abnormal signal (such as construction on main road A from 8:00 to 9:00), or the weather is heavy rain (visibility less than 200 meters leading to widespread deceleration), it is judged as a reasonable abnormality and removed from the initial set. Only retain signals that pass all verifications and add them to the precise set. The set is then updated with verification results and exclusion reason fields to obtain the precise set of abnormal signals.

[0020] From the precise set of abnormal signals, abnormal signals that meet the preset abnormal feature criteria are selected. The preset abnormal feature criteria are adapted to high-density scenarios and need to simultaneously meet the requirements of spatial clustering, temporal concentration, and impact correlation. Abnormal signals are spatially concentrated (e.g., more than 5 abnormal signals within 200 meters of an intersection); abnormal signals are temporally synchronized (e.g., concentrated during the morning rush hour from 8:00 to 8:15, accounting for more than 30% of the total signals during that period); and abnormal signals are related to the overall traffic situation (e.g., an individual low-speed signal is accompanied by a chain reaction of deceleration of surrounding vehicles, rather than an isolated event). A candidate set is selected using a combination of clustering and standard verification. The K-means clustering algorithm is used to group precise abnormal signals according to geographical location (latitude and longitude grid) and occurrence time (15-minute window). For example, the 6 low-speed signals around intersection A on the main road from 8:00 to 8:15 are grouped into Group 1. Each group of signals is verified to see if it meets the abnormal characteristic criteria. For example, if Group 1 meets the criteria of "spatial clustering (6 signals within 200 meters), temporal concentration (8:00-8:15), and impact correlation (3 surrounding vehicles decelerate simultaneously)," then all signals in this group are included in the candidate set. If a group of signals has only 1 isolated low-speed signal and does not meet the spatial clustering criteria, it is removed. The candidate set is stored in units of group ID and includes "a list of signals in the group, clustering range, occurrence time, and the criteria met," thus obtaining the abnormal signal candidate set.

[0021] In step S12, the candidate set of abnormal signals is grouped and the trajectory deviation of vehicle trajectory data within each group is calculated. If the trajectory deviation exceeds a preset deviation threshold, it is determined to be a potential risk signal and used as an input subset to obtain the correlation analysis input subset, including: The abnormal signal candidate set is grouped and processed to obtain the behavioral characteristics of vehicle trajectory data in each group, and the distribution of the behavioral characteristics is analyzed to obtain the distribution range of the behavioral characteristics. Calculate the trajectory deviation of the behavioral features within the distribution range of the behavioral features. If the trajectory deviation exceeds a preset deviation threshold, it is marked as a potential risk signal, thus obtaining a set of potential risk signals. Analyze vehicle behavior changes based on the set of potential risk signals, and extract key points that match the vehicle trajectory data to obtain a subset of the correlation analysis input.

[0022] It should be noted that the candidate set of abnormal signals is grouped, and the K-means clustering algorithm is used to group abnormal signals with similar behavioral characteristics together to avoid interference from isolated signals. For the abnormal signals in each group, quantitative behavioral features are extracted and their distribution patterns are analyzed. For the vehicle trajectory data within the group, three types of core features are extracted, including speed features, trajectory features, and interaction features. Among them, speed features include the average speed of vehicles in the group and the speed fluctuation range (e.g., 20±5km / h); trajectory features include the cluster center coordinates of the vehicle trajectories in the group and the trajectory dispersion (e.g., the Euclidean distance from the center coordinates is <8 meters); interaction features include the following frequency of vehicles in the group (e.g., 3 times of deceleration per minute) and the number of lane changes (e.g., 2 lane changes per 10 minutes). The feature distribution is presented through box plots, and the distribution range is analyzed by the spatial and temporal range of abnormal behavior. The data is presented in a structured format of "group ID-feature type-distribution interval-spatial / temporal range" to obtain the distribution range of behavioral features.

[0023] The Euclidean distance between the cluster centers of vehicle trajectories within a group and the cluster centers of normal driving trajectories on that road segment (obtained through training using 30 days of historical data) is calculated, along with the trajectory coordinate deviation. , Center of the grouped trajectory X, The X-coordinate is the reference coordinate for the normal trajectory. Y-coordinate of the group trajectory center Using the Y-coordinate as the baseline for the normal trajectory, the vehicle trajectory employs the WGS84 coordinate system, calibrated in real-time via GPS base stations within the tunnel (error ≤ 1 meter). The calibration scheme is based on Real-Time Dynamic Differential GPS (RTK-DGPS) technology, with fixed GPS base stations (base stations) deployed every 500 meters within the tunnel. The locations of these base stations are professionally measured and entered into the system; their WGS8 coordinates are known, precise values ​​(error ≤ 0.1 meters). The base stations receive satellite signals in real-time, comparing their own known precise coordinates with the satellite positioning coordinates to calculate comprehensive error parameters for the area, including satellite orbital error, ionospheric delay, and tropospheric delay. The base stations broadcast these real-time error correction parameters to vehicles traveling within the tunnel or roadside sensor nodes via 4G / 5G networks. Simultaneously, the vehicle's GPS / BeiDou devices receive satellite signals and the error correction parameters from the base stations, correcting the original positioning data in real-time and outputting calibrated, precise coordinates. All abnormal signals within a group whose trajectory deviation exceeds a preset threshold are marked as potential risk signals. If the overall deviation of a group is 200% (exceeding the 100% threshold), then all 8 abnormal signals within that group are marked as potential risks. Add data for "deviation value, threshold, and reason for exceeding the threshold" to obtain a set of potential risk signals.

[0024] Analyze vehicle behavior changes based on the set of potential risk signals. Use the sliding window method, segmenting data into 5-minute windows to observe the trend of behavioral characteristics over time. Speed ​​and clustering trends are analyzed. For example, in a potential risk signal group, the speed decreased from 25 km / h at 7:30 to 20 km / h at 7:40, and then to 15 km / h at 7:50, showing a continuous downward trend (not a temporary fluctuation). Clustering trends are also observed, such as vehicles within the group initially being dispersed within a 500-meter radius, then converging within a 200-meter radius of the intersection at 7:40, with the queue length increasing from 50 meters to 100 meters. Determine whether the behavioral changes are persistent and have a widening effect. The analysis focuses on the dispersion of vehicle behavior, retaining behaviors that exhibit continuous deceleration and clustering, indicating associated risks. Key points are the trigger points or clustering points of vehicle behavior changes. These key points can be categorized into behavioral abrupt change points and clustering centers. Behavioral abrupt change points include locations where vehicle speed suddenly drops from 30 km / h to 15 km / h (coordinates x=195, y=210), or where lane-changing frequency suddenly increases from 1 time / minute to 3 times / minute. Clustering centers include intersection entrances where 80% of vehicles in a group congregate, or road sections with the longest queues. By integrating potential risk signals, behavioral change trends, and key points, a structured input subset is formed, resulting in the correlation analysis input subset.

[0025] In step S13, surrounding environmental variables are extracted from the correlation analysis input subset, and a mapping relationship is established based on the traffic density of the surrounding environmental variables and intersection interaction data to obtain an anomaly identification embedding representation, including: Environmental variables are extracted from the input subset of the association analysis and categorized to obtain a comprehensive dataset; A mapping relationship is established based on the traffic density and intersection interaction data of the comprehensive dataset, and the feature distribution of key points in the input subset of the correlation analysis is analyzed to obtain the feature distribution. If the feature distribution deviates from the preset distribution threshold, the potential risk signal is labeled to obtain an anomaly identification embedding representation.

[0026] It should be noted that environmental variables within a 500-meter radius centered on the key points are extracted from the input subset of the correlation analysis (such as the potential risk signal at intersection A of the main road output by S12 plus key points). These variables cover three core categories: traffic density, intersection interaction data, and external interference variables. Traffic density includes vehicle flow (vehicles / minute) and vehicle density (vehicles / 100 square meters) within 500 meters of the key points. Intersection interaction data includes vehicle-traffic light interactions (e.g., number of red light violations / 5 minutes) and vehicle-to-vehicle interactions (e.g., number of intersection conflicts / 5 minutes). External interference variables include weather (e.g., heavy rain), temporary construction (yes / no), and holidays (e.g., weekdays). The extracted environmental variables are categorized by variable type to form a structured comprehensive dataset. The variable types include numerical variables and enumerable variables. Numerical variables include traffic density (80 vehicles / minute), green light duration (30 seconds), and number of conflicts (5 times / 5 minutes). Enumerable variables include weather (1=sunny, 2=rainy, 3=foggy). The final comprehensive dataset is obtained.

[0027] Based on the comprehensive dataset, a mapping relationship between traffic density and intersection interaction data is established. Traffic density is divided into three intervals (low: <50 vehicles / minute, medium: 50-100 vehicles / minute, high: >100 vehicles / minute), and intersection interaction data (such as the number of conflicts) is divided into three levels (low: <3 times / 5 minutes, medium: 3-6 times, high: >6 times). A mapping relationship between levels is established, and the feature distribution of key points in the input subset of the correlation analysis is analyzed, including spatial distribution and temporal distribution. The spatial distribution requires statistical analysis of the vehicle distribution density within 200 meters of the key point (e.g., per 100 vehicles / minute). The number of vehicles per square meter is used to generate a heat map. This heat map is then compared with historical heat maps of the same period (normal distribution). If the current heat map shows a high-density cluster at the intersection entrance (e.g., 15 vehicles per 100 square meters, compared to a historical average of 8 vehicles), the spatial distribution is considered abnormal. The temporal distribution can be calculated by statistically analyzing the vehicle throughput at key points in 5-minute windows to generate a time series curve. This curve is then compared with historical curves of the same period (normal distribution). If the current curve shows a "cliff-like drop" at 7:40 (e.g., from 60 vehicles / 5 minutes to 20 vehicles), while the historical curve is smooth (50-60 vehicles / 5 minutes), the temporal distribution is considered abnormal, and the characteristic distribution is obtained.

[0028] The preset distribution thresholds include three dimensions: spatial distribution threshold, temporal distribution threshold, and mapping relationship threshold. The spatial distribution threshold is when the vehicle density within 200 meters of a key location exceeds 50% of the historical average (e.g., if the historical average is 8 vehicles / 100 square meters, the threshold = 8 × 1.5 = 12 vehicles). The temporal distribution threshold is when the vehicle throughput within a 5-minute window fluctuates by more than 30% compared to the historical average (e.g., if the historical average is 50 vehicles, the lower threshold = 50 × 0.7 = 35 vehicles; if the current throughput is 20 vehicles < 35 vehicles, it is considered to exceed the threshold). The mapping relationship threshold is when the deviation rate between traffic density and intersection interaction data exceeds 20% (e.g., in the regression equation example above, the deviation rate is 11% < 20%, which is considered normal). If the feature distribution exceeds the preset threshold, the potential risk signal is labeled. When the spatial distribution exceeds the threshold, spatial aggregation anomaly is labeled; when the temporal distribution exceeds the threshold, traffic mutation anomaly is labeled; when the mapping relationship exceeds the threshold, interactive association anomaly is labeled. The label is vectorized to obtain an anomaly identification embedding representation, for example: embedding vector = [traffic density (80), number of conflicts (5), construction status (0), spatial anomaly (1), temporal anomaly (1), interactive anomaly (0)] (anomaly items are represented by 1 to indicate existence and 0 to indicate non-existence).

[0029] In step S14, the abnormality identification embedding representation is used for classification, and the classification result is used to determine whether it is a precursor to congestion risk and generate a corresponding warning signal, resulting in a warning signal sequence, including: The feature dimensions of the anomaly identification embedding representation are analyzed and classified to obtain classification features; Based on the comparison of the speed feature and flow feature in the classification features, if the speed feature shows continuous deceleration in multiple vehicles and the flow feature analysis shows that the flow does not decrease, it is judged as a potential congestion precursor, and a congestion risk precursor is obtained. Based on the dynamic capture of vehicle distribution trends according to the aforementioned congestion risk precursors, the characteristic distribution of abnormal behavior is obtained; A priority sequence for risk assessment is generated based on the aforementioned feature distribution; The corresponding warning signal is generated based on the priority sequence, thus obtaining the warning signal sequence.

[0030] It should be noted that, according to the aforementioned anomaly identification embedding representation, it is classified according to three dimensions: anomaly type, scope of impact, and duration. These include the anomaly type dimension, scope of impact dimension, and duration dimension. The anomaly type dimension includes spatial clustering anomalies (e.g., vehicle density exceeding a threshold around a key location); temporal abrupt change anomalies (e.g., traffic decrease exceeding 30% in 5 minutes); and interactive correlation anomalies (e.g., traffic M but conflict count H, matching degree < 50%). The scope of impact dimension (based on the number of abnormal vehicles around the key location) includes local anomalies (< 10 vehicles); regional anomalies (10-30 vehicles); and large-scale anomalies (> 30 vehicles). The duration dimension (based on the duration of the anomaly signal) includes short-term anomalies (< 5 minutes); medium-term anomalies (5-15 minutes); and long-term anomalies (> 15 minutes). These three dimensions are combined into classification labels. For example, the embedding representation "spatial anomaly 1 (yes), temporal anomaly 1 (yes), affecting 25 vehicles, lasting 10 minutes" is converted into the classification label "spatial plus temporal composite anomaly plus regional anomaly plus medium-term anomaly".

[0031] By analyzing the anomalies preceding historical congestion events, a congestion precursor feature database is established (containing congestion types, categorized as intersection queuing congestion, main road spreading congestion, and accidental congestion). Each feature includes a combination of classification labels and an occurrence probability. For example, a classification label combination of spatial clustering anomaly, regional anomaly, and medium-term anomaly results in an 85% probability of intersection queuing congestion. The current classification result is matched against the feature database to calculate the risk matching degree. If the matching degree exceeds a preset congestion threshold (calculated by matching the classification result against the precursor feature database in historical congestion events, using the lowest misclassification rate as the threshold, where the misclassification rate equals the proportion of events that are not actually congested but are classified as congested), it is identified as a congestion risk precursor. The risk matching degree is equal to the sum of the overlap between each classification label and the feature database label multiplied by its weight (the weights are set after analyzing the correlation between the precursor features of the past 300 historical congestion events and the occurrence of congestion, calculating the contribution rate of each dimension to congestion formation; spatial and temporal anomalies have a weight of 0.3, and range and duration have a weight of 0.2). The overlap degree measures the current... The quantitative index (value 0-1) of the matching degree between the classification label and the feature library label in each dimension is used to decompose the combination of the two types of labels into three independent dimensions: anomaly type, scope of influence, and duration. The matching value (0 or 1) of each dimension is calculated separately. If the current anomaly type is a subset of the feature library type or completely consistent with it, the matching value is equal to 1; otherwise, it is equal to 0. If the current scope of influence level is consistent with the feature library level, the matching value is equal to 1; otherwise, it is equal to 0. If the current duration level is consistent with the feature library level, the matching value is equal to 1; otherwise, it is equal to 0. The overlap is equal to the anomaly type matching value multiplied by 0.4, the scope of influence matching value multiplied by 0.3, and the duration matching value multiplied by 0.3. Based on risk matching and congestion type, a three-level early warning signal is generated, including the warning level, congestion type, impact range, and recommended measures. For example, Level 1 warning (risk matching ≥90%): congestion may spread to the main road within 10 minutes, affecting a range of 1 kilometer, and temporary traffic diversion is recommended; Level 2 warning (70%-89%): intersection queuing congestion may occur within 15 minutes, affecting a range of 500 meters, and signal timing adjustment is recommended; Level 3 warning (50%-69%): localized accident congestion may occur, and patrol vehicle verification is recommended. The early warning signals are arranged chronologically according to the duration of the abnormality and the trend of risk change to form a sequence, resulting in an early warning signal sequence.

[0032] In step S15, the early warning signal sequence is time-series matched to obtain an abnormal event chain, including: Obtain the historical chain of abnormal events; The warning signal sequence is scanned segment by segment along the time dimension to detect signal fluctuations and obtain fluctuation characteristics; If the fluctuation characteristic continues to deviate from the preset benchmark value, it is determined to be a potential abnormal starting point, and an abnormal starting point event is obtained; Dynamically monitor the spatial distribution trend of the abnormal initiation event, extract the diffusion range, and obtain the expansion boundary; Based on the extended boundary, the abnormal starting point event is correlated and compared with the historical abnormal event chain, and the correlation chain is extracted to obtain the abnormal event chain.

[0033] It should be noted that the historical abnormal event development chains extracted from the system's historical database must include three elements: time, space, and signal characteristics. The time element refers to the occurrence time of each node in the event chain (e.g., t0: the starting point of the anomaly, t10: the spread to 500 meters, t20: the formation of congestion); the space element refers to the influence range of each node (e.g., the starting point coordinates, the spread boundary, and the final coverage area); and the signal characteristics refer to the warning signal type (e.g., Level 3 → Level 2 → Level 1 warning) and fluctuation characteristics (e.g., the rate of change of signal strength over time) of each node.

[0034] A fixed time window with a step size is used to scan the early warning signal sequence to capture the trend of signal changes over time. The window and step size are set to a window duration of 5 minutes (consistent with the early warning signal update frequency) and a step size of 1 minute. Key parameters such as the intensity value of the early warning signal (e.g., Level 3 = 0.3, Level 2 = 0.6, Level 1 = 0.9), the range of influence, and the duration are scanned. Three types of core fluctuation characteristics are calculated from the scan results to reflect the abnormal change trend of the signal, including the rate of change of intensity, the rate of expansion of range, and the continuity of signal. The rate of change of intensity is equal to (the signal intensity of the current window minus the intensity of the previous window) divided by the time interval (e.g., from 0.3 to 0.6 within 5 minutes, the rate of change = 0.06 / minute); the rate of expansion of range is equal to (the current range of influence minus the range of the previous window) divided by the time interval (e.g., from 100 meters to 300 meters within 10 minutes, the rate of expansion = 20 meters / minute); the continuity of signal is equal to the proportion of signal intensities greater than 0 in N consecutive windows (e.g., there is a signal in 3 consecutive windows, continuity = 100%), thus obtaining the fluctuation characteristics.

[0035] Based on the fluctuation characteristics of the starting point of an anomaly in a historical chain of abnormal events, a normal fluctuation range (allowing ±20% error) is set. The intensity change rate benchmark is the average change rate within the initial 5 minutes of the historical anomaly starting point = 0.02 / minute → benchmark range = 0.016-0.024 / minute; the range expansion rate benchmark is the average expansion rate within the initial 10 minutes of the historical anomaly starting point = 5 meters / minute → benchmark range = 4-6 meters / minute. If the current fluctuation characteristics exceed the benchmark range for two consecutive windows (10 minutes), it is determined as a potential anomaly starting point, resulting in an anomaly starting point event. The spatial distribution trend of the anomaly starting point event is dynamically monitored, and its spatial diffusion is tracked in real time. Both the diffusion direction and diffusion speed need to be detected. The diffusion direction is determined by the change in the centroid coordinates of the influence range within a continuous time window (e.g., from the east entrance to the west entrance of an intersection). The diffusion speed is equal to the distance between the boundaries of the influence ranges of adjacent windows divided by the time interval (e.g., from 100 meters to 200 meters within 5 minutes, speed = 20 meters / minute). Polygonal boundary coordinates are used to describe the expansion range, ensuring that spatial characteristics can be quantified and compared.

[0036] The temporal fluctuation and spatial expansion characteristics of the current anomaly's starting point event are compared with historical anomaly event chains in multiple dimensions, including time, space, and signal dimensions. The time dimension includes the duration from the anomaly's starting point to the present and whether the rate of change in signal intensity at each stage matches the historical chain. The spatial dimension includes whether the diffusion direction, expansion speed, and boundary shape match the characteristics of the historical chain in the same segment. The signal dimension includes whether the warning signal sequence (Level 3 → Level 2 → Level 1) is consistent with the signal evolution of the historical chain. The current event is then linked with the historical event chain to complete the potential development stages, forming an anomaly event chain. Potential development stages describe subsequent states that have not yet occurred but are highly probable in the current anomaly event chain, and must include three elements: time node, spatial range, and warning signal level, extracted through the evolutionary patterns of similar historical events.

[0037] In step S16, the chain integrity of the abnormal event chain is calculated. If the chain integrity is higher than a preset integrity threshold, the type of abnormal event is analyzed to obtain the final abnormal event type.

[0038] It should be noted that chain integrity is an indicator (value 0-100%) measuring the completeness of the description of the abnormal evolution process by the current abnormal event chain. It needs to cover three key elements, including key node coverage, temporal continuity, and feature matching degree. Key node coverage is the percentage of the core stages "abnormal initiation → diffusion → peak → mitigation" included in the event chain (by traversing all nodes in the current abnormal event chain, counting the actual number of core nodes, and calculating the actual number of stages divided by the total number of stages multiplied by 100%. For example, if a complete chain should contain 4 nodes, and only 3 are actually included, then the coverage is 75%). Temporal continuity is the continuity of the timestamps of each node (by traversing the timestamps of the nodes in the current event chain, calculating the time interval between all adjacent nodes, and calculating the time interval of a single interval). The score is equal to 1 minus (the interval between adjacent nodes minus 10), then divided by 20 and multiplied by 100%. The average of all single-interval scores is taken as the final temporal coherence score. It should be noted that when the interval between adjacent nodes is greater than 30, the temporal coherence score is set to 0. The feature matching degree is the degree of agreement between the spatial features (such as diffusion direction) and signal features (such as warning level escalation) of adjacent nodes (refer to the fluctuation feature comparison logic in step S15). For each pair of adjacent nodes in the current event chain, the spatial features (diffusion direction angle) and signal features (warning level) are extracted. For each pair of adjacent nodes, the single-group score is calculated by dividing the number of matching dimensions by the total number of matching dimensions and multiplying by 100%. The average of all single-group scores of adjacent nodes is taken as the final feature matching degree score. The integrity score is equal to the key node coverage multiplied by 0.4, plus the temporal coherence score multiplied by 0.3, plus the feature matching degree score multiplied by 0.3.

[0039] Analyzing 300 complete historical abnormal event chains, the minimum completeness score that accurately predicts the final type is statistically analyzed, and the average score is taken. A completeness threshold of 70% is ultimately set (chains with a score ≥70% are considered "completely described"). When chain completeness >70% threshold, the final abnormal event type is determined through feature clustering and matching with a historical type library. Core features are extracted (identical features extracted from complete chains), including evolution speed features, spatial impact features, and trigger source features. Evolution speed features refer to the duration from the starting point to the peak (e.g., 20 minutes → rapid evolution); spatial impact features refer to the maximum coverage area (e.g., 1 km main road → large-scale impact); trigger source features refer to the initial abnormal type (e.g., interactive association abnormality → accident triggered, spatial aggregation abnormality → traffic saturation triggered). The historical type library contains known abnormal event types and their typical features. The type is determined by comparing feature similarity (obtained by calculating the cosine of the angle between two vectors) (a match greater than 80% is considered the current type), thus obtaining the final abnormal event type.

[0040] In step S17, the associated feature vector is extracted from the final abnormal event type and pushed to the pre-established traffic management platform to obtain instructions, resulting in a sequence of emergency response trigger instructions, including: Extract the associated feature vectors from the final abnormal event types, classify and process them to obtain a set of feature vectors; Dynamic data streams are obtained from a pre-established traffic management platform, and abnormal signals that match the set of feature vectors are extracted and their distribution is analyzed to obtain the distribution range. If the distribution range exceeds a preset range threshold, then emergency response confirmation information returned from the traffic management platform is obtained; Based on the emergency response confirmation information, a specific instruction sequence is generated to obtain the trigger instruction sequence.

[0041] It should be noted that, from the final abnormal event type, a correlation feature vector is extracted, and four key features are extracted: event attribute features, spatial features, temporal features, and associated environmental features. Event attribute features include event type (e.g., traffic congestion caused by a traffic accident), level (Level 1 / Extraordinary), and duration (e.g., 30 minutes); spatial features include the current boundary coordinates of the affected area (e.g., polygon vertices), core congestion points (e.g., (200, 210)), and direction of spread (East → West); temporal features include the event occurrence time (t0), expected duration (e.g., until t60), and peak period (e.g., t30); associated environmental features include weather (heavy rain), time period (morning rush hour 7:30), and surrounding infrastructure (e.g., school areas). The extracted features are quantified into vector form and classified according to the urgency of the event (urgency includes Extraordinary = 1.0, Level 1 = 0.8, Level 2 = 0.6, with higher values ​​indicating greater urgency), forming a feature vector set.

[0042] Two types of real-time data are obtained from the traffic management platform through a pre-defined interface: basic road network data and real-time resource data. Basic road network data includes speed limits, lane numbers, and traffic light control permissions for road sections; real-time resource data includes patrol vehicle locations, police force distribution, and dispatchable clearing equipment (such as tow trucks). The feature vector set is matched with the platform data to pinpoint the specific impact range of abnormal events within the road network. Using the core points and boundary coordinates in the feature vectors as a baseline, and correlated with the platform's electronic map, the specific affected road sections are determined (e.g., "Main Road A Section K0+100 to K0+500"). The total length of the affected road sections (e.g., 500 meters), the number of intersections involved (e.g., 3), and the estimated number of affected vehicles (e.g., 200 vehicles) are calculated to determine the distribution range.

[0043] Based on road segment type and traffic flow, a threshold for triggering an emergency response is set. For arterial roads, the threshold is set to an impact length > 500 meters or involving ≥ 3 intersections; for secondary arterial roads, the threshold is an impact length > 300 meters or involving ≥ 2 intersections. If the current distribution range exceeds the preset threshold, the system automatically sends an "emergency response request" to the traffic management platform, including an event feature vector (e.g., extremely congested, 500-meter impact range) and preliminary handling suggestions (e.g., requiring 2 tow trucks and 3 police officers). After receiving the request, the platform administrator confirms the response level (e.g., "activate Level 1 emergency") and returns confirmation information including "response priority and resource allocation authority." Based on the emergency response confirmation information returned by the platform, a time-ordered and clearly defined sequence of instructions is generated, resulting in the trigger instruction sequence.

[0044] In step S18, the trigger instruction sequence is executed, the feedback data stream is extracted and input into the pre-trained traffic model, and the accuracy of the model output is judged. If the accuracy meets the conditions, the optimized anomaly recognition framework is determined.

[0045] It should be noted that, based on the trigger command sequence, the "command issuance - execution tracking" module of the traffic management platform promotes the implementation of operations by various entities and records the execution process in real time. The execution tracking dimensions include time nodes (such as "whether the patrol car has reached the designated point (200,210) in t0+5 minutes" and "whether the traffic light timing has been adjusted to 40 seconds of green light for east-west traffic in t0+10 minutes"), execution effects (such as "whether the traffic speed in this lane has increased from 20km / h to 35km / h within 5 minutes after the disabled vehicle has been towed away" and "whether the queue length at the intersection has been shortened from 100 meters to 50 meters") and resource consumption (such as "whether the actual number of tow trucks called (2 vehicles) is consistent with the command requirement (2 vehicles)" and "whether the police force deployment (3 The data stream is structured to extract three types of core feedback data from the execution tracking data, including execution status data, traffic status change data, and deviation data. Execution status data includes instruction completion rate (e.g., 9 out of 10 instructions completed → 90% completion rate) and execution delay time (e.g., an instruction is completed 2 minutes late). Traffic status change data includes a comparison of traffic indicators before and after instruction execution (e.g., changes in speed, flow rate, and queue length). Deviation data includes the deviation between the actual effect and the instruction's expectation (e.g., the expected speed increased to 40 km / h, but the actual speed was only 35 km / h → 12.5% ​​deviation) and the reasons for not meeting expectations (e.g., "heavy rain caused vehicles to accelerate slowly").

[0046] Based on the feedback data, predict the "effect of the current triggering command sequence on the handling of abnormal events" and output the "prediction effect score" (0-1, such as 0.85). The pre-trained traffic model can adopt a time-series prediction model (based on LSTM+Attention), which calls the "command-effect correlation parameters" trained in the past (such as "the effect coefficient of the trailer command in rainy weather is 0.8"), and combines the "environment coefficient (1.0)" and "speed change rate (0.75)" in the current feedback data to calculate the prediction effect. The prediction effect score is directly output by the trained LSTM+Attention model. For example, combining the basic command effect (0.9), the environment correction coefficient (0.8) and the execution completion rate (0.9), the model output score is 0.648 (the actual output needs to be combined with the model's complex weights). Accuracy is the core indicator that measures the degree of agreement between the "model prediction effect" and the "actual real effect". It is calculated as a percentage of accuracy. A prediction deviation threshold is set (e.g., ≤0.1). The percentage of data whose deviation between the prediction effect score and the actual effect label is less than the threshold is counted. Accuracy is equal to (the number of data with deviation less than the threshold divided by the total number of data) multiplied by 100%.

[0047] The training process of the time series prediction model requires extracting the trigger command sequence and feedback data stream of 300 abnormal events in the past 12 months. Each data point contains input features and labels. Input features refer to command type, environmental coefficient, execution completion rate, and initial traffic state (speed, flow). Labels refer to the actual handling effect score (calculated based on traffic state recovery rate, such as speed from 20km / h to 35km / h, recovery rate equals (35-20) divided by (normal speed 40-20) = 75% → label = 0.75). The data is preprocessed. For missing values, the mean of 5 historical data points in the same scenario is used for filling. The data is standardized by performing Min-Max standardization on the input features (e.g., execution completion rate 90% → (90-0) / (100-0) = 0.9). The time series is divided into training set (70%, 210 data points), validation set (20%, 60 data points), and test set (10%, 30 data points) according to time order. The training process involves initializing the model, constructing an LSTM+Attention network with the parameters described above, and using the Adam optimizer to minimize the mean squared error (MSE) loss function. After each training round, the accuracy is evaluated using the validation set. If the validation set accuracy does not improve for 10 consecutive rounds, the learning rate is reduced by 0.5. When the validation set loss function increases for 15 consecutive rounds, training is stopped (to avoid overfitting), and the optimal model parameters are saved (85th round, validation set accuracy 89%).

[0048] Based on the practical accuracy requirements of traffic anomaly identification, a three-level judgment threshold is set, including a core threshold and a tolerance threshold. The core threshold is an accuracy rate greater than 80% (meeting this condition indicates that the model prediction is highly consistent with the actual effect, and can be used to optimize the framework); the tolerance threshold is 60% greater than the accuracy rate but less than 80% (requires supplementary feedback data for re-verification, and the framework will not be optimized at this time); the threshold not met is an accuracy rate less than 60% (the model prediction deviation is too large, and the feedback data preprocessing or model parameters need to be checked retrospectively). When the model accuracy meets the core threshold, based on the feedback data stream and model output results, if the feedback data shows that environmental features have a significant impact on the handling effect (adjust the environmental coefficient weight in the model from 0.3 to 0.4), adjust the weight of environmental features in the anomaly feature extraction; if the feedback data shows that the original speed decrease of 20% in the congestion precursor judgment has a high misjudgment rate (the actual effective threshold is 25%), then the speed threshold for congestion precursor judgment is corrected; if the feedback data shows that the handling effect of the trailer plus traffic light linkage command (recovery rate of 85%) is higher than that of the single trailer command (60%), then a combined command is added in the command generation to determine the optimized anomaly identification framework.

[0049] In summary, this invention establishes a dynamic baseline based on historical data and combines multiple features such as speed, flow rate, and spatial distribution for collaborative judgment. It uses K-means clustering to group abnormal signals, calculates trajectory deviation, and accurately filters potential risk signals, avoiding the oversensitivity of traditional solutions that trigger alarms for a single vehicle anomaly, thus reducing false alarm rates. Real-time analysis of traffic data streams automates the entire process, from extracting candidate sets of abnormal signals to generating warning sequences, eliminating the need for manual intervention and shortening response time. Through linkage with the traffic management platform, it automatically transforms abnormal feature vectors into "time-ordered and clearly defined" trigger commands and tracks execution feedback, avoiding the inefficiency of traditional solutions involving "manual dispatch and information gaps," thereby improving emergency response efficiency and early warning capabilities.

[0050] Reference Figure 2 The second embodiment of the present invention provides an abnormal traffic event identification system based on a large traffic model, comprising: The data acquisition module is used to acquire real-time traffic data streams and extract abnormal feature vectors to obtain a candidate set of abnormal signals; The data grouping module is used to group the candidate set of abnormal signals and calculate the trajectory deviation of the vehicle trajectory data within the group. If the trajectory deviation exceeds a preset deviation threshold, it is judged as a potential risk signal and used as an input subset to obtain the correlation analysis input subset. The data mapping module is used to extract surrounding environmental variables from the correlation analysis input subset, establish a mapping relationship based on traffic density and intersection interaction data in the surrounding environmental variables, and obtain anomaly identification embedding representation; The data generation module is used to classify the anomaly identification embedded representation, determine whether it is a precursor to congestion risk based on the classification result, and generate a corresponding early warning signal to obtain an early warning signal sequence. The data matching module is used to perform time-series matching on the warning signal sequence to obtain an abnormal event chain; The data calculation module is used to calculate the chain integrity of the abnormal event chain. If the chain integrity is higher than a preset integrity threshold, the type of abnormal event is analyzed to obtain the final abnormal event type. The data extraction module is used to extract the associated feature vectors from the final abnormal event type and push them to the pre-established traffic management platform to obtain instructions and get the emergency response trigger instruction sequence; The data judgment module is used to execute the trigger instruction sequence, extract the feedback data stream, input it into the pre-trained traffic model, judge the accuracy of the model output results, and determine the optimized anomaly recognition framework if the accuracy meets the conditions.

[0051] It should be noted that the abnormal traffic event identification system based on a large traffic model provided in this embodiment of the invention is used to execute all the process steps of the abnormal traffic event identification method based on a large traffic model in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0052] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as an abnormal traffic event identification program based on a large traffic model. When the processor executes the computer program, it implements the steps in the various embodiments of the abnormal traffic event identification method based on the large traffic model described above, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above system embodiments, such as the data judgment module.

[0053] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0054] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0055] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0056] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0057] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0058] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0059] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for identifying abnormal traffic events based on a large traffic model, characterized in that, include: Acquire real-time traffic data streams and extract abnormal feature vectors to obtain a candidate set of abnormal signals; The abnormal signal candidate set is grouped and the trajectory deviation of the vehicle trajectory data within the group is calculated. If the trajectory deviation exceeds a preset deviation threshold, it is judged as a potential risk signal and used as an input subset to obtain the correlation analysis input subset. The surrounding environmental variables are extracted from the input subset of the correlation analysis, and a mapping relationship is established based on the traffic density and intersection interaction data in the surrounding environmental variables to obtain the anomaly identification embedded representation. The anomaly identification embedding is used for classification, and the classification results are used to determine whether it is a precursor to congestion risk and generate a corresponding warning signal to obtain a warning signal sequence. The abnormal event chain is obtained by performing time-series matching on the warning signal sequence; Calculate the chain integrity of the abnormal event chain. If the chain integrity is higher than a preset integrity threshold, analyze the type of abnormal event to obtain the final abnormal event type. Extract the associated feature vector from the final abnormal event type and push it to the pre-established traffic management platform to obtain instructions, thus obtaining the emergency response trigger instruction sequence; The trigger command sequence is executed, the feedback data stream is extracted and input into the pre-trained traffic model, and the accuracy of the model output is judged. If the accuracy meets the conditions, the optimized anomaly recognition framework is determined.

2. The abnormal traffic event identification method based on a large traffic model according to claim 1, characterized in that, The process of acquiring real-time traffic data streams and extracting abnormal feature vectors to obtain a candidate set of abnormal signals includes: Vehicle trajectory data and traffic flow indicators are extracted from the real-time traffic data stream, and abnormal features are filtered out to obtain an abnormal feature set. Determine the individual behavior patterns of vehicle trajectory data in the abnormal feature set. When the individual behavior pattern deviates from the preset normal behavior pattern, mark it as a potential abnormal signal to obtain an initial abnormal signal set. The initial set of abnormal signals is compared and verified a second time to obtain a precise set of abnormal signals; From the precise set of abnormal signals, abnormal signals that meet the preset abnormal feature criteria are sorted out to obtain a candidate set of abnormal signals.

3. The abnormal traffic event identification method based on a large traffic model according to claim 1, characterized in that, The abnormal signal candidate set is grouped, and the trajectory deviation of vehicle trajectory data within each group is calculated. If the trajectory deviation exceeds a preset deviation threshold, it is determined to be a potential risk signal and used as an input subset to obtain the correlation analysis input subset, including: The abnormal signal candidate set is grouped and processed to obtain the behavioral characteristics of vehicle trajectory data in each group, and the distribution of the behavioral characteristics is analyzed to obtain the distribution range of the behavioral characteristics. Calculate the trajectory deviation of the behavioral features within the distribution range of the behavioral features. If the trajectory deviation exceeds a preset deviation threshold, it is marked as a potential risk signal, thus obtaining a set of potential risk signals. Analyze vehicle behavior changes based on the set of potential risk signals, and extract key points that match the vehicle trajectory data to obtain a subset of the correlation analysis input.

4. The abnormal traffic event identification method based on a large traffic model according to claim 1, characterized in that, The step of extracting surrounding environmental variables from the input subset of the correlation analysis, establishing a mapping relationship between the traffic density of the surrounding environmental variables and intersection interaction data, and obtaining anomaly identification embedding representation includes: Environmental variables are extracted from the input subset of the association analysis and categorized to obtain a comprehensive dataset; A mapping relationship is established based on the traffic density and intersection interaction data of the comprehensive dataset, and the feature distribution of key points in the input subset of the correlation analysis is analyzed to obtain the feature distribution. If the feature distribution deviates from the preset distribution threshold, the potential risk signal is labeled to obtain an anomaly identification embedding representation.

5. The abnormal traffic event identification method based on a large traffic model according to claim 1, characterized in that, The step of performing time-series matching on the warning signal sequence to obtain the abnormal event chain includes: Obtain the historical chain of abnormal events; The warning signal sequence is scanned segment by segment along the time dimension to detect signal fluctuations and obtain fluctuation characteristics; If the fluctuation characteristic continues to deviate from the preset benchmark value, it is determined to be a potential abnormal starting point, and an abnormal starting point event is obtained; Dynamically monitor the spatial distribution trend of the abnormal initiation event, extract the diffusion range, and obtain the expansion boundary; Based on the extended boundary, the abnormal starting point event is correlated and compared with the historical abnormal event chain, and the correlation chain is extracted to obtain the abnormal event chain.

6. The abnormal traffic event identification method based on a large traffic model according to claim 1, characterized in that, The step of extracting the associated feature vector from the final abnormal event type and pushing it to a pre-established traffic management platform to obtain instructions, resulting in an emergency response trigger instruction sequence, includes: Extract the associated feature vectors from the final abnormal event types, classify and process them to obtain a set of feature vectors; Dynamic data streams are obtained from a pre-established traffic management platform, and abnormal signals that match the set of feature vectors are extracted and their distribution is analyzed to obtain the distribution range. If the distribution range exceeds a preset range threshold, then emergency response confirmation information returned from the traffic management platform is obtained; Based on the emergency response confirmation information, a specific instruction sequence is generated to obtain the trigger instruction sequence.

7. The abnormal traffic event identification method based on a large traffic model according to claim 1, characterized in that, The step involves classifying the anomaly based on the embedded representation, determining whether the classification result indicates a potential congestion risk, and generating a corresponding warning signal to obtain a warning signal sequence, including: The feature dimensions of the anomaly identification embedding representation are analyzed and classified to obtain classification features; Based on the comparison of the speed feature and flow feature in the classification features, if the speed feature shows continuous deceleration in multiple vehicles and the flow feature analysis shows that the flow does not decrease, it is judged as a potential congestion precursor, and a congestion risk precursor is obtained. Based on the dynamic capture of vehicle distribution trends according to the aforementioned congestion risk precursors, the characteristic distribution of abnormal behavior is obtained; A priority sequence for risk assessment is generated based on the aforementioned feature distribution; The corresponding warning signal is generated based on the priority sequence, thus obtaining the warning signal sequence.

8. An abnormal traffic event identification system based on a large traffic model, characterized in that, include: The data acquisition module is used to acquire real-time traffic data streams and extract abnormal feature vectors to obtain a candidate set of abnormal signals; The data grouping module is used to group the candidate set of abnormal signals and calculate the trajectory deviation of the vehicle trajectory data within the group. If the trajectory deviation exceeds a preset deviation threshold, it is judged as a potential risk signal and used as an input subset to obtain the correlation analysis input subset. The data mapping module is used to extract surrounding environmental variables from the correlation analysis input subset, establish a mapping relationship based on traffic density and intersection interaction data in the surrounding environmental variables, and obtain anomaly identification embedding representation; The data generation module is used to classify the anomaly identification embedded representation, determine whether it is a precursor to congestion risk based on the classification result, and generate a corresponding early warning signal to obtain an early warning signal sequence. The data matching module is used to perform time-series matching on the warning signal sequence to obtain an abnormal event chain; The data calculation module is used to calculate the chain integrity of the abnormal event chain. If the chain integrity is higher than a preset integrity threshold, the type of abnormal event is analyzed to obtain the final abnormal event type. The data extraction module is used to extract the associated feature vectors from the final abnormal event type and push them to the pre-established traffic management platform to obtain instructions and get the emergency response trigger instruction sequence; The data judgment module is used to execute the trigger instruction sequence, extract the feedback data stream, input it into the pre-trained traffic model, judge the accuracy of the model output results, and determine the optimized anomaly recognition framework if the accuracy meets the conditions.

Images