Traffic jam intelligent identification and early warning method and system based on machine learning

Abstract

The invention discloses a traffic jam intelligent identification and early warning method and system based on machine learning, and belongs to the technical field of intelligent traffic. Comprising the steps that real-time data streams are collected through a camera and a vehicle-mounted terminal, and an initial data set is obtained; performing noise filtering and missing supplement on the initial data to obtain a clean data set; fusing the clean data by adopting a weighted average method, and triggering a standby path when the data is incomplete to obtain a fused data set; extracting traffic flow density and speed trend features from the fused data, dynamically adjusting feature weights according to trend anomalies, and determining a feature set; processing the feature set by using a convolutional neural network to obtain a preliminary congestion probability score; analyzing a time dependency relationship of the probability score sequence through a long short-term memory network, marking high-risk points, and determining a final congestion identification result; and generating and distributing a millisecond-level early warning signal by adopting a priority queue according to an identification result. According to the invention, rapid and accurate identification and early warning of urban traffic congestion are realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent transportation, and particularly relates to a method and system for intelligent identification and early warning of traffic congestion based on machine learning. Background Art

[0002] In the context of the accelerating urbanization process, the problem of traffic congestion has become a key challenge affecting the operation efficiency of cities and the quality of residents' lives. The method for intelligent identification and early warning of traffic congestion based on machine learning, as the core means to solve this problem, shows irreplaceable importance. By the real-time perception and prediction of traffic conditions through intelligent technology, it can provide timely decision-making support for management departments and travelers, relieve congestion pressure, and optimize the allocation of urban traffic resources. However, the research and application in this field still face many problems, and it is urgent to break through the technical bottleneck to improve the actual effect.

[0003] Currently, although there are various methods trying to solve the problem of traffic congestion identification and early warning, most of the solutions often expose obvious deficiencies when facing the complex and changeable real environment. Many existing technologies focus too much on the performance optimization of a certain aspect during design, ignoring the balance between different requirements, resulting in the system being difficult to operate stably in high-pressure scenarios. Especially in the urban traffic network, the ever-changing traffic flow conditions and unpredictable external disturbances make the existing methods seem powerless when dealing with diverse challenges and difficult to meet the comprehensive needs in actual operations.

[0004] Focusing on the technical difficulties, the biggest dilemma in this field lies in how to coordinate the contradiction between real-time response and system stability. The primary problem is that traffic early warning needs to complete the entire process from data acquisition to result output within an extremely short time. Especially during peak hours or emergencies, even a delay of a few seconds may lead to the failure of traffic diversion measures. What's more complicated is that this extreme pursuit of speed often conflicts with the adaptability of the system in various environments. Because to deal with special situations such as a sharp increase in traffic flow during holidays or temporary road adjustments, it is necessary to increase the computational burden of the system, which further exacerbates the time pressure. For example, in the urban expressway network, when traffic anomalies occur simultaneously at multiple intersections, the system needs to instantly judge the congestion trend and give detour suggestions. If it cannot quickly adapt to this complex situation of multi-point linkage, it may miss the best intervention opportunity.

[0005] Therefore, how to ensure the stable operation of the system while guaranteeing a millisecond-level response when facing diverse traffic scenarios has become the key problem that this research urgently needs to overcome. Summary of the Invention

[0006] The present invention proposes a method and system for intelligent identification and early warning of traffic congestion based on machine learning to solve the problems existing in the above-mentioned prior art.

[0007] To achieve the above objectives, this invention provides a machine learning-based intelligent identification and early warning method for traffic congestion, comprising the following steps: Based on the real-time data streams collected through cameras and vehicle terminals, multi-source data covering traffic flow density and speed indicators of the urban road network are obtained to obtain an initial data set; Preprocessing operations are performed on the initial data set to obtain a clean data set; Perform multi-source data fusion on the clean data set to obtain a fused data set; Feature vectors are extracted from the fused data set to obtain the traffic flow density and speed change trends. When the trend is judged to be abnormal, the feature weights are adjusted to determine the feature set used as input for the congestion identification model. The feature set is processed using a convolutional neural network. Convolutional operations are performed on image and sequence data for complex scenarios to obtain a preliminary congestion probability score. The initial congestion probability score sequence is analyzed by Long Short-Term Memory Network to obtain the time dependency relationship. When the sequence shows a rapid upward trend, it is marked as high risk to determine the final congestion identification result. A response signal is generated based on the final congestion identification result. High-risk points are sorted using a priority queue mechanism, and an early warning is issued immediately when the risk score is higher than the threshold.

[0008] Optionally, obtaining the initial data set includes: Based on the data collected by camera equipment and vehicle terminals, dynamic vehicle data for each road segment in the urban road network is obtained, resulting in a preliminary data record set. Based on the initial data record set, a preset coverage threshold is used for judgment, and when the number of data points in a road segment is lower than the coverage threshold, data is collected by cameras in adjacent road segments to obtain the initial data set.

[0009] Optionally, acquiring the cleaning data set includes: A noise filtering mechanism is used to identify and label outliers in the initial dataset, resulting in preliminary filtered data. Based on the data results after the initial screening, an outlier removal operation is performed to obtain the data group with cleaned outliers; Based on the data group with cleaned outliers, check the data missing ratio, and when the missing ratio exceeds a preset threshold, obtain the corresponding data through a backup source supplementation strategy to obtain a data combination with high integrity. For the data combination with high integrity, consistency verification and standardization processing are performed to obtain structured data adapted for multi-source fusion, and the clean data set is determined.

[0010] Optionally, obtaining the fused data set includes: Preprocess the clean dataset to obtain a pre-processed dataset. The weighted average method is used to integrate the signals of the pre-processed data group to obtain the first fused data, and it is determined whether the data meets the integrity threshold. If the first fused data does not meet the integrity threshold, an alternative path is activated to obtain supplementary data, forming the second fused data; The second fused data undergoes consistency verification and strategy adjustment to obtain optimized third fused data; Based on the third fused data, the dynamic characteristics of scene changes are analyzed to generate the fused data set that reflects the complex environment.

[0011] Optionally, determining the feature set used as input to the congestion identification model includes: Feature vectors are extracted from the fused data set, and the traffic density and speed data are standardized to obtain a preliminary set of feature vectors. Based on the preliminary set of feature vectors, time series analysis is used to analyze the changing trends of traffic flow density and speed, and to determine whether the trends show abnormal fluctuations. If the trend of change shows abnormal fluctuations, then the extreme weather-related data will be compared and analyzed to obtain the result of the abnormal trend judgment. Based on the determination result of the trend anomaly, the feature weights are dynamically adjusted to obtain the adjusted feature weight set. The feature vectors are reordered and filtered using the adjusted feature weight set to determine the final feature set used as input to the congestion identification model.

[0012] Optionally, obtaining the preliminary congestion probability score includes: The feature set is processed using a convolutional neural network, and convolution operations are performed to obtain the preliminary congestion probability score; Based on the preliminary congestion probability score, potential congestion points are identified, and potential congestion points are confirmed when the score is higher than a preset threshold. Based on the potential congestion points, obtain additional sequence data for the surrounding areas; The initial congestion probability score is fused using the additional sequence data via convolution to obtain an updated congestion probability score.

[0013] Optionally, determining the final congestion identification result includes: Long Short-Term Memory (LSTM) networks are used to process congestion probability score sequences to obtain time dependencies. Based on the time dependency, calculate the slope change of the sequence to determine whether there is a rapid upward trend; If the slope of the rapid upward trend changes beyond a preset threshold, it is marked as high risk; By combining the high-risk markers with historical congestion data, the final congestion identification result is determined.

[0014] Optionally, the step of immediately issuing an alert when the risk score exceeds a threshold includes: High-risk points are obtained from the final congestion identification results, and a priority queue is used to sort the high-risk points to obtain a sorted sequence; A risk score is determined based on the sorting sequence, and a response signal is generated when the risk score exceeds a threshold to obtain early warning data. The warning data is used to distribute response signals, and the response signals are processed using threshold judgment to obtain a distribution list. Retrieve millisecond-level alerts from the distribution list and output the millisecond-level alerts.

[0015] This invention also provides a machine learning-based intelligent traffic congestion identification and early warning system, comprising: The data acquisition module is used to obtain multi-source data covering urban road network traffic density and speed indicators based on real-time data streams collected through cameras and vehicle terminals, and to obtain an initial data set. The data preprocessing module is used to perform preprocessing operations based on the initial data set to obtain a clean data set; The data fusion module is used to perform multi-source data fusion on the clean data set to obtain a fused data set; The feature extraction module is used to extract feature vectors from the fused data set, obtain the traffic flow density and speed change trends, and adjust the feature weights when the trend is judged to be abnormal, so as to determine the feature set used for the input of the congestion identification model. The convolutional neural network processing module is used to process the feature set through a convolutional neural network, perform convolution operations on image and sequence data for complex scenarios, and obtain a preliminary congestion probability score. The Long Short-Term Memory Network Analysis Module is used to analyze the initial congestion probability score sequence through Long Short-Term Memory Network Analysis, obtain the time dependency relationship, and mark the sequence as high risk when it is judged to have a rapid upward trend, thus determining the final congestion identification result. The early warning generation module is used to generate a response signal based on the final congestion identification result, sort high-risk points using a priority queue mechanism, and issue an early warning immediately when the risk score is higher than the threshold.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This invention discloses a machine learning-based intelligent traffic congestion identification and early warning method and system. It addresses the problem of delayed and inaccurate congestion identification in urban road networks due to the susceptibility of real-time data streams to noise, missing data, and extreme weather. This problem is logically related to how to efficiently integrate multi-source data from cameras and vehicle terminals to achieve rapid early warning in complex scenarios. This invention preprocesses the initial dataset by merging it for noise filtering and missing data supplementation. It then uses a weighted average method to fuse the data to improve completeness. Feature vectors are extracted from the fused data, and weights are adjusted to handle abnormal trends. A convolutional neural network is used to process image and sequence data to obtain preliminary congestion probabilities. A long short-term memory network is then used to analyze time dependencies and mark high-risk points. Finally, a millisecond-level response signal is generated based on a priority queue, thus solving the problems of incomplete data and delayed early warning. The technical advantages of this invention are significantly improved accuracy and real-time performance in congestion identification, balanced computational complexity, and efficient urban traffic management. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a system structure diagram of an embodiment of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0020] Example 1 like Figure 1 As shown, this embodiment provides a machine learning-based intelligent traffic congestion identification and early warning method, including the following steps: Based on the real-time data streams collected through cameras and vehicle terminals, multi-source data covering traffic flow density and speed indicators of the urban road network are obtained to obtain an initial data set; Preprocessing operations are performed on the initial data set to obtain a clean data set; Perform multi-source data fusion on the clean data set to obtain a fused data set; Feature vectors are extracted from the fused data set to obtain the traffic flow density and speed change trends. When the trend is judged to be abnormal, the feature weights are adjusted to determine the feature set used as input for the congestion identification model. The feature set is processed using a convolutional neural network. Convolutional operations are performed on image and sequence data for complex scenarios to obtain a preliminary congestion probability score. The initial congestion probability score sequence is analyzed by Long Short-Term Memory Network to obtain the time dependency relationship. When the sequence shows a rapid upward trend, it is marked as high risk to determine the final congestion identification result. A response signal is generated based on the final congestion identification result. High-risk points are sorted using a priority queue mechanism, and an early warning is issued immediately when the risk score is higher than the threshold.

[0021] Specifically, the following steps are included: Step S101: Real-time data streams are collected through cameras and vehicle terminals to obtain multi-source data to cover traffic density and speed indicators in the urban road network, and an initial data set is obtained for subsequent fusion processing.

[0022] Specifically, real-time data collection of vehicle dynamics in the urban road network is performed using camera equipment and vehicle-mounted terminals. At least one data point is acquired for each road segment, resulting in a preliminary data record set. Based on this preliminary data record set, a preset threshold is used to determine the data coverage. If the number of data points for a certain road segment is lower than the threshold, data is collected from camera equipment on adjacent road segments to determine a complete road network monitoring dataset. For this complete road network monitoring dataset, data cleaning tools are used to remove outliers and duplicates, extracting multi-source data that meets quality requirements to form an initial set for subsequent processing. Data fusion tools are used to integrate the multi-source data in the initial set, combining it with traffic density and speed indicators to determine the traffic status of each road segment, resulting in the final urban road network traffic condition dataset.

[0023] For example, in urban road network traffic monitoring scenarios, real-time data collection on vehicle dynamics can be achieved through camera equipment deployed at intersections and vehicle-mounted terminal equipment. Camera equipment is primarily responsible for capturing the number of vehicles and their movement trajectories on the road segment, while the vehicle-mounted terminal uploads speed and location information through the vehicle's own positioning system. Assuming a city has 100 main road segments, at least one data point needs to be collected from each segment, initially forming a record set containing 100 data points.

[0024] For example, when determining data coverage, a preset threshold can be set to at least three data points for each road segment. If only one data point is collected for a road segment, which is below the threshold, the system will automatically call upon camera devices on adjacent road segments to supplement the data collection. For instance, if road segment A has insufficient data points, the system may detect that the camera coverage of adjacent road segment B extends to the end of road segment A. Therefore, it will collect additional data through the devices on road segment B, ultimately ensuring that road segment A has at least three data points, forming a complete road network monitoring dataset. This approach effectively improves the comprehensiveness of data coverage and reduces monitoring blind spots.

[0025] For example, in the data cleaning process, data cleaning tools are used to remove outliers and duplicates. Data can be filtered by setting speed ranges. Assuming the normal speed range for vehicles is 10 to 80 kilometers per hour, if a data point shows a speed of 150 kilometers per hour, it is considered an outlier and removed. Similarly, if multiple data points are recorded repeatedly at the same time and location, redundant data will be deleted to ensure a high-quality initial dataset. This step significantly improves data reliability and lays the foundation for subsequent analysis.

[0026] For example, in data fusion processing, by integrating multi-source data and combining traffic density and speed indicators, the traffic status of each road segment can be accurately determined. Suppose a road segment has a traffic density of 50 vehicles per minute and an average speed of 20 kilometers per hour; the system can determine that the segment is congested based on preset rules. Another road segment, with a traffic density of 30 vehicles per minute and a speed of 60 kilometers per hour, is judged to be in a smooth-flowing state. Through this fusion analysis, a comprehensive dataset of urban road network traffic conditions is ultimately formed, providing a scientific basis for traffic management.

[0027] For example, from a business perspective, the above methods can help transportation departments monitor road network conditions in real time and adjust traffic light timings or issue congestion warnings promptly. For instance, during peak hours, if the system detects persistent congestion on a certain road segment, it can analyze the dataset to determine if the congestion is caused by excessive traffic density, thereby optimizing traffic management plans for surrounding road segments. This technology not only improves traffic management efficiency but also effectively alleviates urban congestion and enhances the travel experience for citizens.

[0028] For example, in extended applications, a complete traffic condition dataset can also be used to predict future traffic flow trends. By comparing historical data with the current dataset, the system can predict potential congestion on a particular road segment within the next hour, allowing for proactive measures. This predictive capability further enhances the system's practical value and provides strong technical support for the construction of smart urban transportation systems.

[0029] Step S102: Perform preprocessing operations based on the initial data set, use a noise filtering mechanism to remove outliers, and supplement data from backup sources if the number of missing data exceeds a preset threshold, thereby determining a clean data set to support multi-source data fusion.

[0030] Specifically, the process begins by retrieving information from the initial dataset, performing preprocessing to initially organize the data, and identifying and marking outliers using a noise filtering mechanism to obtain pre-filtered data results. For these pre-filtered data results, outlier removal is implemented; if a data point deviates from a preset threshold, it is removed, identifying the data group with cleaned outliers. Data missingness is then checked from this cleaned data group; if the missing percentage exceeds a preset threshold, corresponding data is obtained through a backup source supplementation strategy, resulting in a data combination with higher integrity. Based on this high-integrity data combination, a clean data generation process is used to verify the consistency of the supplemented data. By comparing field integrity, it is determined whether the data meets the requirements for multi-source fusion, resulting in a validated dataset. For this validated dataset, data standardization is performed, using format unification to obtain structured data suitable for multi-source fusion, determining the final fusionable data units. Starting from these final fusionable data units, multi-source fusion logic is applied using pre-established mapping rules to associate and integrate data units from different sources, resulting in a unified fused data result.

[0031] For example, in the business domain of urban road network traffic data processing, for the stage of obtaining information to be processed from the initial dataset, a hierarchical filtering approach can be used to initially classify the data collected by cameras and vehicle terminals according to timestamps and road segment identifiers. Assuming a road segment has 1000 data points during peak hours, a noise filtering mechanism is first used to identify and mark the 50 outliers that deviate from the normal range, obtaining the initial filtering results. This method helps to quickly locate problematic data during subsequent processing.

[0032] For example, for outlier removal, a speed threshold range can be set, such as 10 to 80 kilometers per hour. If a data point shows a speed of 120 kilometers per hour, it is considered an outlier and removed. In practice, historical data trends can be used as a supplementary basis for judgment, such as the average speed of the same road segment over the past week being 40 kilometers per hour, to ensure the rationality of the removal.

[0033] For example, when checking for missing data, if the missing data ratio for a certain road segment is 30%, exceeding the preset threshold of 20%, then a backup source supplementation strategy is activated to obtain relevant data from cameras on adjacent road segments. The specific supplementation logic could prioritize data from the nearest road segment and combine this with time-time matching to ensure the relevance of the supplemented data.

[0034] For example, data consistency verification can be performed by comparing field completeness to check whether the supplementary data includes necessary fields, such as road segment number and collection time. If a data entry is missing the time field, it is marked as unqualified and needs to be retrieved again. This verification method effectively ensures the smoothness of subsequent fusion processing.

[0035] For example, in the data standardization process, data formats from different sources can be unified into a standard time format, such as adjusting "2023-10-01 08:00" and "202310010800" to a consistent format. Simultaneously, the speed unit is standardized to kilometers per hour, ensuring the data structure adapts to integration requirements. This operation significantly improves data integration efficiency.

[0036] For example, when applying multi-source fusion logic, mapping rules can be pre-established to associate and integrate road segment identifiers from camera data and vehicle terminal data. Suppose a road segment has 100 data points from two sources; by matching them according to time and location using mapping rules, 50 unified fused data points can be formed. This approach effectively improves the accuracy of comprehensive data analysis.

[0037] For example, as an extension, a weighting mechanism can be introduced during the fusion process, prioritizing the reliability of camera data during peak hours and assigning it a higher weight, such as 0.7, while giving vehicle terminal data a weight of 0.3. This differentiated processing can better reflect actual traffic conditions and provide a more reliable basis for subsequent analysis.

[0038] Step S103: Perform multi-source data fusion on the clean data set, and use a weighted average method to integrate camera images and vehicle terminal signals. Determine whether the fused data reaches the integrity threshold. If it is lower than the threshold, trigger the backup fusion path to obtain the fused data set that reflects the complex changes in the scene.

[0039] Specifically, the clean dataset undergoes preprocessing to obtain a pre-processed data set. A weighted average method is used to integrate the signals from this pre-processed data set, generating first fused data. The first fused data is then assessed to determine if it meets an integrity threshold; if not, the missing data is recorded. If the first fused data still doesn't meet the integrity threshold, a backup path is activated to obtain supplementary data, forming second fused data. Consistency checks are performed on the second fused data to determine its usability. Based on the results of the second fused data check, and considering the impact of scenario changes and complex environments, the data integration strategy is adjusted to obtain optimized third fused data. Based on the third fused data, the dynamic characteristics of scenario changes are analyzed to form a feature data set reflecting the complex environment. The feature data set is then assessed to determine if it meets the expected standards. If the feature data set meets the expected standards, the final dataset is generated. This final dataset is then stored and formatted to obtain structured data suitable for subsequent analysis.

[0040] For example, in the signal integration stage, when using a weighted average method, weights can be assigned based on the reliability of the data sources. Assuming the camera data has a weight of 0.6 and the vehicle terminal data has a weight of 0.4, the first fused data is generated through a weighted average. This method can balance the contributions of data from different sources, and especially in complex environments, it can effectively improve the overall reliability of the data.

[0041] For example, when assessing the completeness of the first fused data, a threshold of 85% can be set. If the completeness of the fused data is only 70%, missing data is recorded, such as the discovery of lost camera data within a certain time period. In this case, activating a backup path becomes particularly important. The backup path can be another set of backup cameras, historical data cache, or existing map software data. By supplementing the missing parts, a second fused data is formed, and its consistency is verified to ensure that the accuracy of the supplemented data matches the original data.

[0042] For example, when adjusting data integration strategies, the impact of scene changes and complex environments must be considered. Suppose a vehicle enters a tunnel, causing a decrease in camera data quality. In this case, the weight of the vehicle terminal data can be increased to 0.7, while the weight of the camera data can be decreased to 0.3, generating optimized third-party fused data. This dynamic adjustment strategy can adapt to environmental changes and ensure the stability of data fusion.

[0043] For example, when analyzing the dynamic characteristics of scene changes, key information, such as vehicle speed change trends or road condition features, can be extracted from third-party fusion data. Suppose the data shows that the speed drops from 60 km / h to 20 km / h in a short period, which may indicate congestion or obstacles ahead. The resulting feature data set can provide important information for subsequent decision-making.

[0044] For example, during the generation and storage of the final dataset, the data can be formatted into a uniform structured format, such as using timestamps as indexes to record speed and image feature information per second. This formatting facilitates subsequent analysis, while storage optimization reduces data redundancy, improves access efficiency, and supports real-time application scenarios.

[0045] Step S104: Extract feature vectors from the fused data set, obtain the traffic flow density and speed change trends, determine whether the trends indicate the impact of extreme weather, and if the trends are abnormal, adjust the feature weights to determine the feature set for input to the congestion identification model.

[0046] Specifically, feature vectors are extracted from the fused dataset. Data preprocessing is performed on changes in traffic density and speed, and the data is normalized using standardization methods to obtain a preliminary set of feature vectors. Based on this preliminary set, the changing trends of traffic density and speed are analyzed. Time series analysis is used to decompose the trends and determine if they exhibit abnormal fluctuations. If abnormal fluctuations are found, comparative analysis is performed with data related to extreme weather conditions. A preset threshold is used to determine whether extreme weather has an impact, resulting in a trend anomaly assessment. Based on this assessment, if extreme weather is identified, the feature weights are dynamically adjusted, increasing the weight ratio of feature vectors corresponding to the abnormal trend, thus determining the adjusted feature weight set. Using this adjusted feature weight set, the feature vectors are reordered and filtered to obtain a subset of features highly correlated with congestion identification, thus determining the final feature determination result.

[0047] For example, extracting feature vectors is a crucial step in processing fused datasets. Feature vectors are typically the core information extracted from multi-source data, used to characterize dynamic features such as traffic density and speed changes. For traffic density, vehicle quantity data captured by cameras can be combined with the total number of vehicles passing through a road segment within a time period to form a density index. For instance, if 50 vehicles pass through a road segment per minute during peak hours, but only 10 vehicles per minute during off-peak hours, this difference can serve as the basic input for the feature vector. For speed changes, average speed data recorded by in-vehicle terminals can be used. For example, if the average speed on a road segment is 60 km / h under normal conditions, but drops to 20 km / h during congestion, this forms a feature value for speed change. Subsequently, these data are normalized using standardization methods, unifying data from different dimensions into the range of 0 to 1, facilitating subsequent analysis.

[0048] For example, when analyzing trends in traffic density and speed, time series analysis can help decompose the periodicity and anomalous fluctuations in the data. Taking traffic density as an example, suppose the density data of a city's main road shows a regular peak during weekday morning rush hour, but on a certain day the data suddenly exceeds the normal range—for example, the normal peak density is 50 vehicles per minute, but on a certain day it reaches 80 vehicles per minute—this may indicate an anomalous fluctuation. Regarding speed changes, if the speed data for a certain road segment suddenly drops from 60 km / h to 10 km / h in a short period, it may also indicate an anomaly. Through time series decomposition, long-term trends and short-term fluctuations can be separated, thus determining whether the anomaly is an isolated event.

[0049] For example, when determining whether abnormal fluctuations are influenced by extreme weather, meteorological data can be used for comparative analysis. Suppose that on a certain day, traffic density is abnormally high, and meteorological data indicates a rainstorm warning with rainfall reaching 30 millimeters per hour. Combined with a preset threshold, such as rainfall exceeding 20 millimeters potentially affecting traffic, it can be inferred that the abnormal fluctuations are related to extreme weather. This comparative analysis helps to accurately identify the causes of anomalies, providing a basis for subsequent strategy adjustments.

[0050] For example, if an anomaly is determined to be influenced by extreme weather, dynamically adjusting feature weights becomes particularly important, such as through linear weighting. For feature vectors showing changes in traffic density and speed, the weight proportion related to abnormal trends can be increased. For instance, under normal circumstances, the weights of density and speed features are each 0.5, but under the influence of heavy rain, the weight of the density feature is increased to 0.7, and the weight of the speed feature is decreased to 0.3, to highlight the importance of density data in congestion identification. This adjusted weight set can better reflect traffic conditions under extreme conditions.

[0051] For example, by reordering and filtering feature vectors using an adjusted set of feature weights, we can focus on a subset of features highly relevant to congestion identification. Assuming that after adjustment, density features are prioritized due to their higher weights, while some secondary features, such as vehicle distance data, are filtered out due to lower relevance, the resulting feature subset is more representative of the congestion scenario. This filtering process ensures the relevance of data analysis and helps improve the accuracy of subsequent congestion predictions.

[0052] Step S105: Use a convolutional neural network to process the feature set, perform convolution operations on image and sequence data for complex scenarios, and obtain a preliminary congestion probability score to quantify potential congestion points.

[0053] Specifically, a convolutional neural network is used to process the feature set, performing convolution operations on complex scenes to obtain a preliminary congestion probability score. Based on this preliminary congestion probability score, potential congestion points are identified; if the preliminary congestion probability score is higher than a preset threshold, a potential congestion point is confirmed. Based on the potential congestion point, supplementary sequence data of the surrounding area is obtained. Using this supplementary sequence data, a convolution operation is performed to fuse the preliminary congestion probability score, resulting in an updated congestion probability score. Based on the updated congestion probability score, congestion points are identified; if the updated congestion probability score is higher than a preset threshold, a congestion point is confirmed.

[0054] For example, the step of using a convolutional neural network to process the feature set and perform convolution operations can be understood as decomposing the acquired image data into multiple feature layers, such as vehicle density and lane occupancy. By analyzing these feature layers one by one, the system can preliminarily determine whether there are signs of congestion on a certain road segment. Suppose that at an intersection, the system identifies a high vehicle density and, combined with speed sequence data, finds that vehicles are moving slowly, ultimately outputting a preliminary congestion probability score of 0.75, indicating a potential risk of congestion.

[0055] For example, when identifying potential congestion points, if the preset threshold is 0.7, the initial score of the aforementioned intersection of 0.75 exceeds the threshold and is therefore marked as a potential congestion point. This method can quickly identify areas that require further attention.

[0056] It should be noted that the threshold setting can be flexibly adjusted according to the traffic characteristics of different road sections. For example, the threshold can be appropriately increased on main roads with high traffic volume to reduce false alarms.

[0057] For example, regarding the step of acquiring additional sequence data from the surrounding area, it can be envisioned that after identifying a potential congestion point, the system would access real-time data from other nearby sensors, such as traffic speed and density information for adjacent road segments. Suppose that within a 1-kilometer radius of the aforementioned intersection, sensors show that vehicle speeds are generally below normal levels, further supporting the possibility of congestion.

[0058] For example, in the step of fusing additional sequence data and updating the congestion probability score, the system combines the new data with the preliminary results to recalculate the score. Suppose that after fusion, the score rises from 0.75 to 0.82, indicating a further increase in the likelihood of congestion. This multi-data source fusion method can improve the reliability of the judgment.

[0059] For example, when finally identifying a congestion point, if the updated score of 0.82 is still higher than the threshold of 0.7, then the intersection is identified as a congestion point.

[0060] It should be noted that this layered verification method can effectively avoid misjudgments caused by a single data source, and at the same time provide accurate basis for subsequent traffic management, such as timely adjustment of traffic light duration or issuance of congestion warnings.

[0061] For example, from the overall process perspective, each of the above steps is interconnected, forming a complete analysis chain from data collection to final judgment.

[0062] It should be noted that this method is particularly suitable for complex urban traffic scenarios, and can quickly respond to potential problems in a changing environment, providing strong support for alleviating traffic congestion.

[0063] Step S106: Analyze the preliminary congestion probability score sequence through a long short-term memory network to obtain the time dependency relationship, determine whether there is a rapid upward trend in the sequence, and mark it as high risk if it exists, and determine the final congestion identification result to reduce early warning lag.

[0064] Specifically, real-time traffic flow data is collected through traffic sensors to obtain a preliminary congestion probability score sequence. A Long Short-Term Memory (LSTM) network is used to process the score sequence and obtain its time dependencies. Based on these time dependencies, the slope change of the sequence is calculated to determine if a rapid upward trend exists. If the slope change of a rapid upward trend exceeds a preset threshold, it is marked as high-risk. Using these high-risk markers, historical congestion data is integrated to determine the final congestion identification result. Based on the final congestion identification result, the real-time traffic model is updated to obtain an adjusted probability score sequence.

[0065] For example, in traffic flow analysis, collecting real-time traffic data through traffic sensors is a crucial step. Suppose multiple sensors are deployed on a city's main roads, recording the number of vehicles passing through every minute, resulting in a preliminary congestion probability score sequence. These scores reflect the likelihood of road congestion at different times; for instance, during the morning rush hour, a certain road segment might have a probability score of 0.75, indicating a high risk of congestion.

[0066] For example, processing score sequences in Long Short-Term Memory (LSTM) networks to obtain time dependencies can be understood as capturing patterns in congestion probability scores over time by analyzing the correlation between historical and current data. Suppose that the probability score of a certain road segment gradually increases from 0.3 to 0.75 in the past hour; the network will recognize this continuously rising trend and determine that congestion may be forming. This method is particularly suitable for capturing periodic changes in traffic flow, such as the regular congestion during weekday morning and evening rush hours.

[0067] For example, when calculating changes in the slope of a sequence to identify a rapid upward trend, this can be achieved by analyzing the rate of change of the probability score over a short period. If, within 10 minutes, the probability score jumps from 0.4 to 0.8, with the slope change significantly exceeding a preset threshold of 0.03, the system will mark that road segment as a high-risk area. This marking helps to provide early warnings and prevent further deterioration of congestion.

[0068] For example, integrating historical congestion data to determine the final congestion identification result can improve accuracy by comparing current data with past congestion records for the same time period. Suppose historical data shows that a certain road segment frequently experiences congestion at 8:00 AM on Mondays, and current data also shows a high-risk marker; the system will combine both to confirm that the road segment does indeed have a potential for congestion. This approach effectively reduces false positives and improves the reliability of the identification.

[0069] For example, updating real-time traffic models and adjusting probability score sequences allows for dynamic optimization of subsequent predictions by feeding the latest identification results back into the model. If a road segment is identified as a congestion point, the model will adjust the probability scores of surrounding road segments to indicate potentially affected areas, such as increasing the score of an adjacent road segment from 0.5 to 0.6. This dynamic adjustment helps to more accurately reflect traffic conditions and supports subsequent decision-making.

[0070] For example, from a business perspective, the above methods can help traffic management departments understand road conditions more promptly and respond quickly to potential problems. For instance, after a high-risk sign appears, traffic light durations can be adjusted in advance or detour suggestions can be issued, thereby alleviating congestion. This proactive management approach not only improves road utilization efficiency but also enhances the travel experience for citizens.

[0071] For example, as an extension, congestion identification can be further optimized by combining weather data or special event information with the core method. Suppose that during rainy days or large events, the system adjusts the probability score threshold based on additional factors, such as lowering the threshold from 0.7 to 0.6, to more sensitively capture congestion risks. This diversified data fusion approach allows the system to adapt to more complex scenarios, enhancing its practicality.

[0072] Step S107: Generate a response signal based on the final congestion identification result, sort high-risk points using a priority queue mechanism, and immediately distribute the risk points if the risk score is higher than the threshold to obtain millisecond-level early warning output, balancing model accuracy and computational complexity.

[0073] Specifically, high-risk points are identified from congestion results and sorted into a priority queue to obtain a ranking sequence. Risk scores are determined based on the ranking sequence; if a risk score exceeds a threshold, a response signal is generated to obtain early warning data. Response signals are distributed using the early warning data, and a distribution list is obtained by processing the response signals using a threshold judgment. Millisecond early warnings are obtained from the distribution list, and the balance between model accuracy and computational complexity is assessed to obtain an optimization index. A balancing mechanism is determined for the optimization index; if the balancing mechanism meets the conditions, an early warning signal is output to obtain the final sequence. Immediate distributions are obtained from the final sequence, and the immediate distributions are processed using the response signals to obtain the output result.

[0074] For example, in the field of traffic congestion early warning, extracting high-risk points from congestion data and prioritizing them is a crucial step. Suppose that real-time traffic data analysis identifies three high-risk points on a city's main roads: road segment A, segment B, and segment C. Based on congestion probability and historical data, these are assigned risk values ​​of 80, 65, and 90, respectively. After prioritizing them using a priority queue, the resulting sequence is C, B. This sequence provides the basis for subsequent risk score assessment, ensuring that the most severely congested points are addressed first.

[0075] For example, a threshold of 75 can be set for determining the risk score and generating the response signal. If the risk score of segment C in the sorted sequence exceeds the threshold (90), the system will generate a response signal, forming early warning data. This process can be understood as a rapid response to potential congestion crises. The early warning data may contain information such as the specific location of segment C and predicted congestion levels, providing a basis for subsequent distribution.

[0076] For example, in the early warning data distribution phase, the system generates a distribution list based on response signals. Assume the distribution list includes multiple targets, such as traffic management departments, navigation applications, and traffic broadcasting platforms. By using threshold judgments, signals requiring priority processing are selected; for instance, information on road segment C with a risk score exceeding 85 is distributed only to all targets, while information on other road segments is distributed only to a subset of targets. This tiered distribution mechanism effectively allocates resources and ensures timely delivery of critical information.

[0077] For example, for millisecond-level alerts in the distribution list, the system needs to balance model accuracy and computational complexity. Suppose that millisecond-level alerts require a response time of less than 200 milliseconds, while the current model takes 300 milliseconds to compute. Therefore, the algorithm needs to be optimized or some accuracy needs to be reduced to meet the timeliness requirements. Optimization metrics might include reducing the response time to 200 milliseconds while maintaining 90% prediction accuracy. Determining this balancing mechanism helps improve the system's real-time performance.

[0078] For example, once the balancing mechanism meets the conditions, the system outputs a warning signal, forming the final sequence. Assuming the final sequence includes emergency warning information for section C, the system will immediately distribute it to the traffic management department through response signal processing, notifying them to take traffic control measures. This process ensures rapid information transmission and effective response.

[0079] For example, in the specific implementation of obtaining immediate distribution from the final sequence, assuming that after the warning signal for section C is distributed, the traffic management department adjusts the traffic light duration according to the signal content and pushes detour suggestions in the navigation application. This multi-party collaborative output can effectively alleviate congestion pressure, improve road traffic efficiency, and provide drivers with timely road condition guidance, which has significant practical value.

[0080] For example, the implementation of each technical aspect in the overall process is closely centered around the core business of traffic congestion early warning, forming a complete closed-loop system from the identification of high-risk points to final distribution. This logically rigorous approach not only improves the timeliness of early warnings but also ensures the efficient operation of the system through hierarchical management and resource optimization, providing strong support for urban traffic management.

[0081] like Figure 2 As shown, this embodiment provides a machine learning-based intelligent traffic congestion identification and early warning system, including: The data acquisition module is used to obtain multi-source data covering urban road network traffic density and speed indicators based on real-time data streams collected through cameras and vehicle terminals, and to obtain an initial data set. The data preprocessing module is used to perform preprocessing operations based on the initial data set to obtain a clean data set; The data fusion module is used to perform multi-source data fusion on the clean data set to obtain a fused data set; The feature extraction module is used to extract feature vectors from the fused data set, obtain the traffic flow density and speed change trends, and adjust the feature weights when the trend is judged to be abnormal, so as to determine the feature set used for the input of the congestion identification model. The convolutional neural network processing module is used to process the feature set through a convolutional neural network, perform convolution operations on image and sequence data for complex scenarios, and obtain a preliminary congestion probability score. The Long Short-Term Memory Network Analysis Module is used to analyze the initial congestion probability score sequence through Long Short-Term Memory Network Analysis, obtain the time dependency relationship, and mark the sequence as high risk when it is judged to have a rapid upward trend, thus determining the final congestion identification result. The early warning generation module is used to generate a response signal based on the final congestion identification result, sort high-risk points using a priority queue mechanism, and issue an early warning immediately when the risk score is higher than the threshold.

[0082] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent identification and early warning of traffic congestion based on machine learning, characterized in that, The method comprises the following steps: According to the real-time data stream collected by the camera and the vehicle terminal, the multi-source data covering the traffic density and speed indicators of the urban road network is obtained, and an initial data set is obtained; According to the initial data set, a preprocessing operation is performed to obtain a clean data set; Multi-source data fusion is performed on the clean data set to obtain a fusion data set; From the fusion data set, a feature vector is extracted, the traffic density and speed change trend are obtained, and the feature weight is adjusted when the trend is abnormal to determine the feature set for the congestion recognition model input; The convolutional neural network is used to process the feature set, the image and sequence data convolution operation is performed for complex scenes, and a preliminary congestion probability score is obtained; The long short-term memory network is used to analyze the preliminary congestion probability score sequence, the time dependence is obtained, and the sequence is marked as high risk when a rapid upward trend is judged, and the final congestion recognition result is determined; According to the final congestion recognition result, a response signal is generated, a priority queue mechanism is used to sort the high-risk points, and a warning is immediately given when the risk score is higher than the threshold.

2. The method of claim 1, wherein, The initial data set is obtained by: According to the collection results of the camera device and the vehicle terminal, the vehicle dynamic data of each road section in the urban road network is obtained, and a preliminary data record set is obtained. According to the preliminary data record set, a preset coverage threshold is used for judgment, and when the number of road segment data points is lower than the coverage threshold, adjacent road segment cameras are used to supplement data collection to obtain the initial data set.

3. The method of claim 1, wherein, The clean data set is obtained by: A noise filtering mechanism is used to identify and mark abnormal values in the initial data set to obtain a preliminary filtered data result. An abnormal value removal operation is performed on the preliminary filtered data result to obtain a data group with cleaned abnormal values. According to the data group with cleaned abnormal values, the proportion of missing data is checked, and when the proportion of missing data exceeds the preset threshold, the corresponding data is obtained through a backup source supplement strategy to obtain a high-integrity data combination. The high-integrity data combination is subjected to consistency checking and standardization processing to obtain structured data suitable for multi-source fusion, and the clean data set is determined.

4. The method of claim 1, wherein, The fusion data set is obtained by: The clean data set is preprocessed to obtain a preliminary processed data group. The weighted average method is used to integrate the signals of the preliminary processed data group to obtain first fusion data, and it is judged whether the first fusion data meets the integrity threshold. If the first fusion data does not meet the integrity threshold, activate the backup path to obtain supplementary data to form second fusion data. The second fusion data is subjected to consistency checking and strategy adjustment to obtain optimized third fusion data. According to the third fusion data, the dynamic characteristics of scene changes are analyzed to generate the fusion data set reflecting complex environments.

5. The method of claim 1, wherein, The feature set for the congestion recognition model input is determined by: From the fusion data set, a feature vector is extracted, and the traffic density and speed data are standardized to obtain a preliminary feature vector set. According to the preliminary feature vector set, a time series analysis method is used to analyze the change trend of the traffic density and speed, and whether the trend presents abnormal fluctuation is judged; If the change trend presents abnormal fluctuation, comparison analysis is performed in combination with extreme weather related data to obtain a trend abnormality judgment result; According to the trend abnormality judgment result, the feature weight is dynamically adjusted to obtain an adjusted feature weight set; Through the adjusted feature weight set, the feature vector is reordered and filtered to determine a feature set finally used for input of a congestion identification model.

6. The method of claim 1, wherein, The obtaining of the preliminary congestion probability score includes: The feature set is processed by a convolutional neural network to perform convolution operation to obtain the preliminary congestion probability score; According to the preliminary congestion probability score, a potential congestion point is judged, and the potential congestion point is determined when the score is higher than a preset threshold; According to the potential congestion point, additional sequence data of a surrounding area is obtained; Through the additional sequence data, the preliminary congestion probability score is fused by convolution operation to obtain an updated congestion probability score.

7. The method of claim 1, wherein, The determination of the final congestion identification result includes: The congestion probability score sequence is processed by a long short-term memory network to obtain a time dependence relationship; According to the time dependence relationship, the slope change of the sequence is calculated to judge whether there is a rapid rising trend; If the slope change of the rapid rising trend exceeds a preset threshold, it is marked as high risk; Through the high risk marking, the final congestion identification result is determined by fusing historical congestion data.

8. The method of claim 1, wherein, The immediate warning when the risk score is higher than the threshold includes: The high risk point is obtained from the final congestion identification result, and the high risk point is sorted by using a priority queue to obtain a sorting sequence; According to the sorting sequence, a risk score is determined, and a response signal is generated when the risk score exceeds a threshold to obtain warning data; Through the warning data distribution response signal, the response signal is processed by using a threshold value judgment to obtain a distribution list; The millisecond level warning is obtained from the distribution list, and the millisecond level warning is output.

9. A machine learning based intelligent identification and early warning system for traffic congestion, characterized in that, It includes: A data acquisition module is configured to obtain a plurality of source data covering traffic density and speed indicators of a city road network from an initial data set according to real-time data streams collected by a camera and a vehicle terminal; A data preprocessing module is configured to perform preprocessing operations on the initial data set to obtain a clean data set; A data fusion module is configured to perform multi-source data fusion on the clean data set to obtain a fusion data set; A feature extraction module is configured to extract feature vectors from the fusion data set, obtain traffic density and speed change trends, and adjust feature weights when the trends are abnormal to determine a feature set for input of a congestion identification model; A convolutional neural network processing module is configured to process the feature set by a convolutional neural network, perform convolution operation on image and sequence data for complex scenes, and obtain a preliminary congestion probability score; A long short-term memory network analysis module is configured to analyze the preliminary congestion probability score sequence by a long short-term memory network, obtain a time dependence relationship, mark as high risk when the sequence has a rapid rising trend, and determine a final congestion identification result. The early warning generation module is configured to generate a response signal according to the final congestion identification result, sort high-risk points using a priority queue mechanism, and immediately perform early warning when the risk score is higher than a threshold value.

Images