Traffic situation awareness data fusion analysis method in five-post-in-one mode

Abstract

The invention provides a traffic situation awareness data fusion analysis method in a five-post-in-one mode. The method comprises the following steps: S1, five-post data acquisition; s2, data aggregation; s3, data preprocessing; s4, performing multi-algorithm fusion processing; s5, performing post data analysis; and S6, comprehensive research and judgment. According to the traffic situation awareness method in the'five-in-one 'mode, closed-loop fusion of'data-feature-association-decision' is realized through algorithm collaboration, and the limitation of a traditional mode is broken through.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation Internet of Things and data processing technology, and in particular to a method for fusion and analysis of traffic situation perception data under a five-position integration model. Background Technology

[0002] In the traditional "five-post" management model for highways (highway police post, traffic enforcement post, monitoring and operation post, road property maintenance post, and collaborative management post), there are significant limitations in data collection, analysis, and application for each post: Data silos: Data from different positions is stored and processed independently (e.g., traffic accident data from highway police posts and traffic flow data from toll collection operations posts are stored in different systems), lacking a unified aggregation mechanism. For example, information about "damaged guardrail at K150" discovered by road maintenance posts cannot be synchronized to the accident investigation system of highway police posts in real time, which may lead to the omission of key clues in the analysis of accident causes; Collaboration gaps: Inter-departmental collaboration relies on manual data transmission (such as telephone reporting), which can easily lead to information delays or inaccuracies. For example, information on "overweight vehicles" identified by traffic enforcement officers needs to be manually communicated to road maintenance officers before it can be linked to "road damage risk analysis." If this information is missed, road maintenance plans will lack focus. Lagging situational awareness: In the traditional model, the analysis of situations such as traffic congestion and equipment failure relies on "manual summarization and experience judgment", which cannot capture dynamic changes in real time (for example, after a sudden congestion on a certain road section, it takes more than 1 hour to complete the data collection and analysis of each post), resulting in a lag in decision response. Summary of the Invention

[0003] This invention provides a traffic situation perception data fusion and analysis method under the "five-position integration" model. The traffic situation perception method under the "five-position integration" model achieves closed-loop fusion of "data-feature-association-decision" through algorithm collaboration, breaking the limitations of the traditional model.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: A method for fusion and analysis of traffic situation perception data under a five-post integration model, including: S1: Data collection for five positions, namely highway police, traffic enforcement, monitoring and operation, road property maintenance, and collaborative management. The collected data includes traffic accident data from the highway police, public opinion monitoring data from the traffic enforcement, traffic flow data from the monitoring and operation, rescue resource data from the road property maintenance, and equipment operation data from the collaborative management. S2: Data aggregation, which aggregates all the data collected in S1 by job position and data type to form the original dataset; S3: Data preprocessing, which cleans and standardizes the original dataset from S2, removes redundant information, unifies the data format, and obtains preprocessed data; S4: Multi-algorithm fusion processing. Based on preprocessed data, it sequentially transforms unstructured text into topic feature vectors through natural language processing, extracts traffic flow temporal features through convolutional neural networks, calculates data association probabilities through Bayesian networks, and completes spatiotemporal gaps through surface fitting. The final output includes a fusion dataset, a set of association probabilities, and a data association chain. S5: Job-specific data analysis. Each job conducts a specific analysis based on the fusion results of S4 to obtain job-specific analysis results; S6: Comprehensive analysis and judgment, the monitoring and operation staff integrates the job-specific analysis results of S5 to generate a traffic situation awareness report.

[0005] The traffic situation awareness data fusion and analysis method under the "five-post integration" model also includes the following in this manual: S7: Instruction distribution, issuing execution instructions to the five posts based on the traffic situation awareness report from S6; S8: Feedback Update, the fifth post feeds back the new data after the instruction is executed to S2, updating the original dataset; S9: Iterative optimization, evaluate the effectiveness of S1-S8 based on the feedback from S8, and adjust the algorithm parameters of S4.

[0006] In this specification, the topic feature vector output by the natural language processing step of S4 and the traffic time-series feature output by the convolutional neural network step are both used as input evidence for the Bayesian network. The conditional probability table of the Bayesian network is generated based on the historical correlation data of the five posts, and the update frequency of the conditional probability table is consistent with the feedback update frequency of S8 (1 hour / time).

[0007] In this specification, in the surface fitting step of S4, the sampling point weights are obtained by normalizing the correlation probability output by the Bayesian network. If the fitting deviation of a certain type of data (such as the correlation between rescue resources and congestion relief) exceeds 30%, the fusion rules for that type of data will be adjusted in S9. The adjusted rules need to be updated synchronously to the fusion logic of S4.

[0008] In this manual, the fusion dataset output by S4 contains the correlation between the data of the five positions (such as the correlation between accident data of the highway police position and toll traffic data of the collaborative management position). This correlation serves as the priority basis for the position-specific analysis of S5 (such as prioritizing the analysis of events with a correlation probability ≥ 50%), and the determination of the correlation is entirely based on the actual business logic of the five positions (such as "accidents causing a decrease in toll traffic" is a fixed correlation).

[0009] In this manual, the fusion dataset output by S4 must include three elements: data source (name of the five positions), processing time, and correlation probability. Each data entry must correspond to the original data number collected by S1 to ensure traceability to the specific collection time of the specific position (e.g., "equipment operation data collected by the collaborative management position at 09:00 on 20XX-07-28").

[0010] In this manual, the traffic situation awareness report of S6 must explicitly reference the fusion dataset of S4 and the job-specific analysis results of S5. The report content includes situation data before the execution of the instruction, situation data after the execution, and the reasons for the deviation. If the deviation exceeds 30%, the fusion rules of S4 should be adjusted in S9 (such as adding the "correlation dimension between rescue resources and congestion"). The adjustment results need to be verified in the next S1-S8 cycle.

[0011] In this manual, the equipment operation data of the collaborative control post is integrated by S4 and used as the basis for the road property maintenance post to allocate rescue resources (such as the distribution of rescue resources related to equipment failure data). This data is fed back to the original dataset in S8 for the effect evaluation of S9.

[0012] In this manual, if the correlation probability calculation rule of the Bayesian network in S4 needs to be adjusted, it must be based on the actual deviation data fed back in S8 (such as "probability deviation of traffic accident correlation with congestion"). The adjusted rule must record the adjustment time and reason in S9.

[0013] In this manual, the closed-loop termination condition for S4 to S9 is: in three consecutive cycles (3 hours), the deviation between the correlation probability output by S4 and the actual situation is ≤10%. At this time, the current fusion rule is deemed valid, the rule adjustment of S9 is stopped, and S1-S8 continue to be executed at the normal frequency.

[0014] In summary, the present invention has at least the following beneficial effects: 1. Improved Data Collaboration Efficiency: Compared to the traditional manual data transmission model, this method eliminates redundant steps such as manual data entry and telephone reporting through "automatic algorithm aggregation-association-analysis" (e.g., S1 data collection directly drives S4 fusion), achieving real-time data sharing across all five positions. For example, rescue information from the road maintenance post can be synchronized to all relevant positions within one minute, avoiding the problem of "information transmission taking more than 10 minutes" in the traditional model.

[0015] 2. Improved Accuracy of Situational Awareness: Traditional methods rely on data from a single role to judge the situation (e.g., predicting congestion solely based on traffic flow, easily overlooking the impact of accidents). This method, however, integrates data from multiple roles (e.g., combining accident location, traffic flow changes, and equipment status), making situational analysis more realistic. For example, judging "congestion caused by a traffic accident" no longer relies solely on traffic flow data but also integrates multi-dimensional information such as accident cleanup progress and rescue resource allocation, avoiding the problem of "excessive discrepancies between predictions and reality" in traditional methods.

[0016] 3. Improved rationality in resource allocation: In the traditional model, the allocation of rescue resources by road maintenance posts often relies on experience (such as "send two clearing vehicles as usual"), which may lead to insufficient or wasted resources. This method dynamically adjusts resource allocation based on integrated data (such as real-time traffic flow after an accident and the distribution of surrounding rescue forces). For example, it prioritizes the dispatch of rescue forces in road sections with increased congestion and reduces redundant resources in road sections with stable traffic flow, making resource utilization more in line with actual needs.

[0017] 4. Improved smoothness of job collaboration: In the traditional model, collaboration between jobs requires manual coordination to initiate (e.g., highway police officers need to actively contact road maintenance officers to obtain accident scene data); this method automatically links data from each job through algorithms (e.g., accident data automatically triggers rescue resource analysis), reducing manual coordination costs and transforming job collaboration from "passive response" to "proactive linkage". Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the traffic situation perception data fusion and analysis method under the five-post integration model involved in this invention.

[0019] Figure 2 This is a schematic diagram of the natural language processing (NLP) process involved in this invention.

[0020] Figure 3 This is a schematic diagram of the convolutional neural network (CNN) process involved in this invention.

[0021] Figure 4 This is a schematic diagram of the Bayesian network process involved in this invention.

[0022] Figure 5 This is a schematic diagram of the surface fitting process involved in this invention. Detailed Implementation

[0023] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] like Figure 1 As shown in the figure, this embodiment provides a method for fusion and analysis of traffic situation perception data under the five-post integration model, including: S1: Data collection for five positions, namely highway police, traffic enforcement, monitoring and operation, road property maintenance, and collaborative management (including electromechanical maintenance and toll collection). The collected data includes traffic accident data from the highway police, public opinion monitoring data from the traffic enforcement, traffic flow data from the monitoring and operation, rescue resource data from the road property maintenance, and equipment operation data from the collaborative management. S2: Data aggregation, which aggregates all the data collected in S1 by job position and data type to form the original dataset; S3: Data preprocessing involves cleaning and standardizing the original dataset from S2, removing redundant information, and unifying the data format (e.g., traffic flow data is standardized to "vehicles / 5 minutes", and time is standardized to "hour:minute:second") to obtain preprocessed data. S4: Multi-algorithm fusion processing. Based on preprocessed data, it sequentially transforms unstructured text into topic feature vectors through natural language processing, extracts traffic flow temporal features through convolutional neural networks, calculates data association probabilities through Bayesian networks, and completes spatiotemporal gaps through surface fitting. The final output includes a fusion dataset, a set of association probabilities, and a data association chain. S5: Job-specific data analysis. Each job conducts a special analysis based on the fusion results of S4 (e.g., highway police job sets investigation priorities based on accident correlation probability) to obtain job-specific analysis results; S6: Comprehensive analysis and judgment, the monitoring and operation staff integrates the job-specific analysis results of S5 to generate a traffic situation awareness report.

[0025] In some embodiments, during S2 data aggregation, newly added external data (such as meteorological data, major event arrangements) is obtained by the monitoring and operations staff through government public channels or cooperation units. The external source and acquisition time are marked during the initial aggregation; this marking is directly used in subsequent cycles. External data introduced in S2 (such as meteorological data) needs to be updated at a frequency of "3 hours / time," automatically obtained by the monitoring and operations staff through an interface with the meteorological department. In the event of a sudden meteorological event (such as short-term heavy rainfall), a temporary data update is triggered. During the first update, "Sudden External Data: Heavy Rainfall, Acquisition Time XX:XX" is marked in S2 and immediately incorporated into the S4 fusion calculation as temporary evidence for the Bayesian network. The original S2 dataset needs to be backed up to physical storage media at a frequency of "1 hour / time." The backup file naming format is "Five-Post Data-Backup Time (Year-Month-Day-Hour:Minute)," and the backup file must contain key parameters of the S4 fusion dataset (such as the association probability matrix). If the master data is lost, it can be restored to the most recent complete S4 output state through the backup file.

[0026] In some embodiments, the cycle frequency of S1 to S8 is "1 hour / time". The data collected by the five posts in S1 must be aggregated to S2 within 5 minutes. After aggregation in S2, the preprocessing of S3 must be completed within 10 minutes to ensure that the fused data in S4 is available within 20 minutes after collection in S1. This real-time requirement applies to all data types of the five posts (including equipment operation data of the collaborative management post). If the transmission delay exceeds 10 minutes, "transmission delay and reason" must be marked in the feedback of S8. When the S7 instruction is distributed, the accident handling instruction of the highway police post must include "specific road section (e.g., 'K150 Chengdu direction'), handling target (e.g., 'complete the towing of the accident vehicle within 30 minutes'), and cooperating post (e.g., 'road property maintenance post cooperates in clearing obstacles')". This instruction must be synchronized from the monitoring and operation post to the traffic law enforcement post (for public opinion prediction) and the collaborative management post (for equipment scheduling) to ensure that the instruction reception of the five posts is completely synchronized without any delay.

[0027] In some embodiments, if abnormal data is found during S3 data preprocessing (such as traffic flow time-series data deviating from the historical same period by more than 50%), an abnormality marking process must be executed: the monitoring and operation personnel record the "abnormal data value, corresponding position (such as 'highway police post'), and collection time", and simultaneously notify the corresponding position to review the original data; if the data is confirmed to be valid after review (such as traffic flow changes caused by special weather), it is marked as "abnormal but valid" during S4 fusion and included in normal calculation; if it is invalid data, it is removed and the reason for removal is recorded. The S3 data preprocessing stage needs to perform "three-level verification": the first level verifies the data format (such as whether the station number retains two decimal places), the second level verifies the data integrity (such as whether the time-series data contains "start time - end time"), and the third level verifies the data correlation (such as "whether the accident data is associated with the corresponding road segment traffic flow"). If a certain level of verification fails, "verification failure item and position source (such as 'highway police post accident data missing timestamp')" must be marked in S3, and the corresponding verification rules should be adjusted first during S9 optimization.

[0028] In some embodiments, the fused dataset output by S4 needs to be verified by the "historical data comparison method": select similar data from the past 3 months (traffic flow of the same road segment at the same time). If the fusion result deviates from the actual value by more than 20%, then trace the fusion rules of S4 (such as whether the correlation between "toll station traffic flow" and "mainline traffic flow" is missing). After adjusting the rules, new data needs to be used for re-verification, and the verification results are recorded in the optimization report of S9. When road maintenance personnel allocate rescue resources based on the fusion results of S4, they need to refer to "distance from rescue point to accident point (obtained from continuous features of surface fitting)" and "correlation probability of accident causing congestion (output by Bayesian network)". The closer the distance and the higher the correlation probability, the higher the priority of rescue resources (e.g., "distance from K150 rescue point to K152 accident point is 2 kilometers, correlation probability is 70%" is preferred over "distance is 5 kilometers, correlation probability is 60%"). When the collaborative management and control post (electromechanical maintenance post) handles equipment failures, the fused data in S4 must include "traffic flow 30 minutes before the failure (provided by the monitoring and operation post)" and "the scope of the failure's impact (station congestion data reported by the toll station operation post)." After the failure is repaired, "equipment operating status after repair" and "corresponding road segment traffic flow changes" must be reported in S8 to ensure traceability of the entire "failure-handling-effect" chain. If, after the traffic enforcement post's public opinion data is fused by S4, the correlation probability between "congestion-related public opinion volume" and "actual congestion duration" is less than 40%, then the fusion rules of S4 should be adjusted in S9 (e.g., adding a correlation dimension of "time difference between public opinion release time and congestion start time"). The effect of the adjustment needs to be verified in the next cycle.

[0029] In some embodiments, all data from S2 to S8 are stored in a directory named “Job Title-Data Type-Timestamp” (e.g., “Collaborative Management and Control Post-Toll Flow-20XX07280900”), where the “Data Type” is completely consistent with the data types of the five posts collected in S1 (e.g., “Traffic Flow Time Series Data”, “Equipment Failure Record”). The storage format is a structured table containing three columns: “Original Value”, “Processing Result”, and “Association Probability”. The iterative optimization stopping condition for S9 is: in 5 consecutive cycles (5 hours), the deviation between the S4 fusion result and the actual situation is ≤10%, and the feedback from the five posts on the execution effect of the instructions is “effective” (e.g., the highway police post reports “accident handling efficiency improved”). At this time, the cycle frequency can be extended to “2 hours / cycle”, but the data transmission link must still be maintained. The priority of S7 instructions is determined based on the correlation probability output by S4: events with a correlation probability ≥ 60% (such as "accident causing congestion") are classified as Level 1 instructions (to be executed within 10 minutes), events with a correlation probability ≤ 30% and < 60% are classified as Level 2 instructions (to be executed within 30 minutes), and events with a correlation probability < 30% are classified as Level 3 instructions (to be executed within 1 hour). The priority label must be clearly marked in the S7 instructions and simultaneously fed back to the result evaluation stage of S8.

[0030] In some embodiments, when the traffic flow data at the toll station of the collaborative management and control post (toll operation post) is merged in S4, a fixed association rule of "toll station traffic × 0.8 = main line associated traffic" needs to be established with the main line traffic flow data of the monitoring and operation post (based on historical data verification, error ≤ 15%). If the actual deviation exceeds 15%, the association coefficient is adjusted in S9 (e.g., changed to 0.75). After adjustment, "corrected association rule and reason" needs to be re-annotated in S4. The spatial boundary of the S4 surface fitting is the entire range of the road sections under the jurisdiction of the five posts (e.g., "K0+000 to K300+000"), and the time boundary is "00:00 to 24:00". If it is necessary to expand the boundary (e.g., add adjacent road section data), "new boundary and acquisition post (e.g., 'collaborative management and control post connects with adjacent road section operators')" needs to be annotated when the data is aggregated in S2. The first fusion after expansion needs to verify the compatibility with the original data. When traffic enforcement personnel's public opinion data is integrated into S4, the "public opinion release time" must be added. If the interval between the release of public opinion and S4 integration exceeds 2 hours, the evidence weight of that public opinion in the Bayesian network must be reduced (e.g., from 0.8 to 0.5). The weight adjustment rules must be clearly recorded in S4 and linked with the optimization logic of S9. When road maintenance personnel allocate rescue resources based on S4 integrated data, they must meet the matching formula "number of rescue personnel = estimated congestion time (minutes) × 0.1 + basic configuration of 2 people" (e.g., if the estimated congestion is 30 minutes, then configure 5 people). The coefficients (0.1, 2) of this formula need to be dynamically adjusted through S9 effect evaluation (e.g., if the actual queuing time exceeds expectations, the coefficient is increased to 0.15).

[0031] In some embodiments, when the equipment operation data of the collaborative management and control post (electromechanical maintenance post) is integrated into S4, the "fault risk index = continuous equipment running time (hours) × 0.02 + historical fault count × 0.3" needs to be calculated. When the index is ≥ 0.8, an early warning is triggered (e.g., the information board needs to be inspected in advance). The warning result serves as the basis for the S7 instruction. When this index is first introduced, the "calculation logic and five-post verification record" must be noted in S4. When the call data of the monitoring and operation post is integrated into S4, an association rule of "every additional 10 congestion inquiries = traffic exceeding normal by 20%" needs to be established with the traffic data of the corresponding road segment. If the call volume and traffic deviation exceed 30% in a certain period (e.g., 10 inquiries correspond to traffic exceeding 10%), then in S9, it is necessary to trace whether the "time difference between the inquiry period and the traffic peak" association dimension was omitted in S4. The results of the S5 job-specific analysis need to be stored in a three-dimensional structure of "job-data type-analysis conclusion" (e.g., "highway police post-accident data-accident investigation priority"). The conclusions in this structure can be directly reused during the S6 comprehensive assessment without recalculation. If the analysis results of a certain job are invalid for 3 consecutive times, the analysis rules of that job will be checked first in S9 (e.g., whether the highway police post has missed the correlation between accidents and weather).

[0032] The technical concept of this invention is as follows: A traffic situation awareness data fusion and analysis method under the five-post integration model includes: S1 Multi-Source Data Acquisition: Provides "Adaptive" Raw Data for the S4 Algorithm Core objective: To collect multi-source data from five different sources according to the input requirements of the S4 algorithm (NLP, CNN, Bayesian network, surface fitting), ensuring that the original data contains the core elements required by the algorithm (such as the "time-location-request" of text and "fixed frequency sampling" of time-series data).

[0033] 1.1 Data Collection Content and Specifications for Each Position (In line with S4 Algorithm Requirements) 1.1.1 Highway Police Post Collect traffic accident and traffic control related data, which should include: Traffic accidents: time (accurate to the minute), location (station number + direction, such as "K150+200 Chengdu direction"), accident type (rear-end collision / single-vehicle accident), casualties, lane occupancy status (number of lanes occupied), and mark "whether it may cause congestion" (yes / no, providing initial association clues for the S4.3 Bayesian network); Traffic control: control type (closure / restriction), scope (start-end station number), start and end time, reason for control, and mark "whether it may cause complaints" (yes / no).

[0034] 1.1.2 Traffic Enforcement Post Data collected on public opinion monitoring, road construction, and overloaded vehicles must include: Public opinion monitoring: source of public opinion (Weibo / Douyin / 12345 hotline), release time, core content (must include "road section + problem", such as "K200 traffic jam for 3 hours"), and dissemination scope (number of reposts / comments); Road construction: construction location (chainage), construction period (start-end time), number of lanes occupied, and whether it has been publicized. Overloaded vehicles: license plate, overload percentage, inspection time, and direction of travel.

[0035] 1.1.3 Monitoring and Operations Position Collect call volume, complaint / report, and cross-section traffic data, including: Call volume / complaint / report: call time, request type (help / inquiry / complaint), core content (must specify "road segment + request", such as "K180 wants to know when the congestion will ease"), contact information (optional); Cross-section traffic: sampled at a fixed frequency of 5 minutes / time (matching the CNN time-series window of S4.2), and record timestamps (such as "9:00-9:05"), station number, traffic flow (vehicles / 5 minutes), and average vehicle speed (km / h).

[0036] 1.1.4 Road maintenance post Data collected on road property damage, clearing and rescue, and emergency response must include: Road property damage: location of damage (chainage), type (guardrail / ditch), extent of damage, and time of discovery; Clearing and rescue: time of rescue, location (chainage), type of disabled vehicle, number of rescue personnel, and whether lanes are occupied; Emergency response: location of natural disaster / accident, scope of impact, and rescue resources deployed (number of personnel / equipment).

[0037] 1.1.5 Collaborative Management and Control Position Electromechanical Maintenance Position: Location, type, discovery time, and repair status of equipment malfunctions (information board / surveillance camera); Network security (attack time, attack type, and interception results); Toll collection and operation staff: number of vehicles entering and exiting the toll station (sampled every 10 minutes), toll revenue, passenger flow in the service area (statistics every hour), and the status of service area facilities (such as whether the restrooms are open).

[0038] 1.2 Data Acquisition Methods and Quality Requirements Data collection tools: Road patrol recorders (road property, law enforcement), automatic export from the system backend (traffic, charges), and manual entry (calls, public opinion). All data is synchronized to the "Five Posts in One" joint operation platform. Quality requirements: Textual data (public opinion, complaints) must not be vague (e.g., "traffic jam ahead" should be changed to "traffic jam at K150 Chengdu direction"); Time-series data (traffic flow, number of vehicles) must not be missing (if missing data is found, it must be marked "missed" and completed in S3); Location information should be uniformly expressed as "station number + direction" (e.g., "K120 Chongqing direction"), avoiding vague descriptions such as "middle section of XX Expressway".

[0039] S2 Data Aggregation: Categorized and integrated according to "S4 Algorithm Type" Core objective: To categorize and aggregate the raw data collected by S1 according to "algorithm input type", establish a "data-algorithm" mapping relationship, and ensure that S4 can directly call the corresponding data, avoiding low calling efficiency caused by mixed data.

[0040] 2.1 Data Classification and Labeling System (corresponding to S4 Algorithm) Add category labels to each data point. The correspondence between the labels and the S4 algorithm is as follows:

[0041] 2.2 Aggregation Operation Process 1. Data access: The joint operations platform automatically synchronizes the data collected by each position every 5 minutes (real-time data), and summarizes the historical data of the day at 24:00 every day; 2. Tag matching: The system automatically identifies data types and adds tags (e.g., "K150 traffic jam" → "text type"; "9:00-9:05 K150 traffic flow 200 vehicles" → "time series type"). Mislabeled data can be corrected manually. 3. Create an index table: Generate a "Data-Algorithm Index Table" to record the "number, tag, storage path, and corresponding S4 step" for each data entry. For example: Number D20XX0728001: Tag "Text", Storage path "Public Opinion Database / 20XX0728", corresponding to S4.1; 4. Data verification: Verify the classification accuracy daily (≥95%). If the mislabeling rate of a certain type of data is >5%, the automatic labeling rules need to be optimized (e.g., increase the recognition weight of keywords such as "traffic jam" and "congestion").

[0042] S3 Data Preprocessing: Outputs standardized data for S4 algorithm. Core objective: To address the "data-sensitive points" of each algorithm in S4 (such as NLP's sensitivity to redundant words and CNN's sensitivity to outliers), clean and standardize the data aggregated in S2 to ensure that the data stream input to S4 is free from interference and has a uniform format.

[0043] 3.1 Text Data Preprocessing (Adapted for S4.1 NLP) 1. Redundant information filtering: Automatically remove non-business words (such as "Hello", "Please", "Thank you", "Trouble") and punctuation marks. For example: Original complaint: "Hello, may I ask if there is a traffic jam on K180 towards Chongqing?" → After processing: "K180 towards Chongqing is congested"; Filter duplicate expressions (such as "It's congested, it's really congested" → After processing: "Traffic jam").

[0044] 2. Synonym normalization: Unify semantically similar expressions into standard terms (based on the five-position business dictionary): "traffic jam", "blocked", "slow traffic" → unified as "congestion"; "no one at the toll station" and "toll collector not here" → unified as "toll station staff shortage".

[0045] 3. Element completion: If the text is missing "time" or "location" (such as "traffic jam, don't know where"), it will automatically associate the caller's IP location (approximate road segment) or the call time (default current time) and mark it with the word "complete".

[0046] 3.2 Temporal Data Preprocessing (Adapted to S4.2CNN) 1. Outlier Detection and Repair: Outlier Definition: Values ​​exceeding "historical average ± 3 standard deviations" (e.g., historical average traffic volume at K1509:00 was 200 vehicles, standard deviation was 50 vehicles; therefore, >350 vehicles or <50 vehicles are outliers). Repair Rules: If a sampling point is an outlier, replace it with the average of the two adjacent normal sampling points (e.g., ...). Exception, use (Replace); if three or more consecutive anomalies occur, mark "data anomaly" and notify the monitoring and operations staff for manual verification.

[0047] 2. Time series alignment: Ensure that the timestamps of all time series data are consistent (e.g., "9:00-9:05" "9:05-9:10"). If there is 10-minute data in "9:00-9:10", split it into two 5-minute data (distributed according to the average).

[0048] 3.3 Preprocessing of Discrete Sampling Data (Adapted to S4.4 Surface Fitting) 1. Coordinate standardization: station number : Keep two decimal places (e.g., "K150.5" → "150.50"); Time Convert to "hour + decimal" format (e.g., "9:06" → 9.1, "10:30" → 10.5).

[0049] 2. Data deduplication: If duplicate sampling occurs at the same location and time (e.g., the road maintenance post and the monitoring and operation post simultaneously record the rescue information for K150), the record with more complete data (e.g., the record containing "number of rescue personnel") will be retained.

[0050] 3.4 Event-based data preprocessing (adapted to S4.3 Bayesian networks) 1. Binarization conversion: Convert the event state to "1 (occurred) / 0 (not occurred)" (e.g., "Traffic accident confirmed" → 1, "not occurred" → 0); 2. Time matching: Ensure that the "time of occurrence" of the event is accurate to the minute (e.g., "around 9:00" → "9:00"), so as to facilitate correlation with the time dimension of other data.

[0051] S4 Data Fusion: Multi-dimensional Fusion Based on Multi-algorithm Collaboration The "five-in-one" traffic situation awareness model requires the integration of multi-source data from five departments: highway police, traffic enforcement, monitoring operations, road asset maintenance, and collaborative management. This includes both structured data (such as traffic flow and accident records) and unstructured data (such as public opinion texts and transcripts of complaint recordings). The core objective of data fusion is to break down "data silos" and transform scattered data into "situational features" that can directly support decision-making through algorithmic collaboration. This step (S4), as the core link of data fusion, achieves a complete process of "unstructured to structured → structured feature extraction → feature association modeling → related data completion" through deep interaction of four algorithms: Natural Language Processing (NLP), Convolutional Neural Networks (CNN), Bayesian Networks, and surface fitting. (Input: "Standardized dataset" output from S3, containing structured and unstructured data submitted by the five departments; Output: "fused dataset", containing data association probability, continuous spatiotemporal features, and complete data association chains) The following details the operational details, examples, and algorithmic collaboration logic of each step.

[0052] S4.1 Natural Language Processing (NLP): Transforming Unstructured Text Data into Structured Data (process as follows) Figure 2 (As shown) Unstructured text generated during the work of the five positions (such as "public opinion monitoring records" from traffic enforcement, "12122 hotline complaint records" from monitoring operations, and "rescue scene descriptions" from road maintenance) cannot be directly used in numerical analysis and need to be transformed into structured features through NLP. This step adopts a combination of "TF-IDF keyword extraction + LDA topic model" to ensure that the transformation results are consistent with the business scenarios of the five positions.

[0053] 1. Definitions of Terms and Symbols A collection of unstructured text data, covering all text data from the five positions. Where m is the total number of text entries (e.g., 5000 monthly public opinion reports and complaints). Represents the i-th text (e.g., "The lights were dim in the tunnel at K300 on the XX Expressway, and we almost had a rear-end collision"). A collection of business keywords, compiled jointly by five departments (combining frequently used terms from daily work). n=200 (e.g., "congestion", "accident", "tunnel lights", "service area toilets", etc.). Represents the k-th keyword; Keywords In the text Word frequency in, i.e. exist The number of times it appears in the text (e.g., if "tunnel lights" appears once in the example text above, then TF=1); Keywords The inverse document frequency reflects the word's "discriminative power" in the text (high-frequency but indiscriminate words, such as "high speed," have low weight). The calculation formula is: ;in For inclusion The number of text entries (denominator + 1 to avoid calculation errors when no text is included); Keywords In the text The weight in the text (term frequency × inverse document frequency) is such that a higher value indicates that the term is more "thematically representative" of the text. Text topic feature vectors, combined with the business requirements of five positions, pre-set 5 types of topics (covering core work scenarios). ,in: Related to road congestion (e.g., "traffic jam" or "slow-moving traffic"); Traffic accident related (such as "rear-end collision" or "rollover"); : Related to facility malfunctions (such as "lights are broken" or "information board is black"); Service quality related (e.g., "no one was at the service area" or "the toll collector had a bad attitude"). Other (e.g., "asking for directions", "weather information"); vector satisfies , This represents the probability that the text belongs to the j-th topic category.

[0054] 2. Model building and training (aligned with the business scenarios of the five positions) Training data preparation: 30,000 unstructured texts from the past 6 months were collected from five different positions. These texts were jointly tagged with themes by the highway police (responsible for accident-related data), traffic enforcement (responsible for public opinion-related data), and monitoring operations (responsible for complaint-related data) personnel (each text corresponds to a unique theme). The data was then divided into a training set (24,000 texts) and a validation set (6,000 texts) at an 8:2 ratio.

[0055] For example: A traffic enforcement post reported, "The K100 section of the XX Expressway was congested from 7 a.m. to 9 a.m., and the emergency lane was blocked." → marked "Road congestion ( )”; The complaint from the monitoring and operations staff, “The service area toilets are not clean and no one cleans them”, is marked as “Service quality ( )".

[0056] TF-IDF Keyword Selection: Calculate the TF-IDF value of all keywords in the training set text, retain the top 150 keywords by weight (filter out words with no distinguishing power such as "highway" and "road section"), and form a "core keyword table" (such as "congestion", "rear-end collision", "lights", "sanitation" etc.).

[0057] LDA topic model training: Using core keywords as input, the number of topics is set to 5 (corresponding to preset topics), and training is performed iteratively through Gibbs sampling. Initial iterations (1-200): Randomly assign topics and calculate the probability that each keyword belongs to a certain topic (e.g., "congestion" belongs to a certain topic). The probability is 0.8). Intermediate iterations (201-800): Adjust text topic allocation based on keyword-topic probabilities, gradually converging (e.g., "emergency lane blocked" is stably allocated to...). ); Final iteration (801-1000): Training stops when the topic classification accuracy on the validation set is ≥85%, and the "topic-keyword correspondence rule" is output (e.g., Corresponding keywords: lighting, information board, surveillance, guardrail, etc.

[0058] 3. Real-time data conversion (full text coverage for all five job roles) For unstructured text in S3 standardized data, convert it according to the following steps (covering all guard post text types): Step 1: Text Preprocessing Text from different positions is cleaned uniformly: punctuation and redundant words (such as "Hello" and "Thank you") are removed, and abbreviations are standardized (e.g., "Highway Police" is standardized to "Highway Police Post"). For example: The road maintenance post's description of the rescue scene was "A truck broke down at K200+500, occupying the slow lane, and vehicles behind were moving slowly" → After cleaning, it became "A truck broke down at K200+500, occupying the slow lane and vehicles behind were moving slowly". Step 2: Keyword Extraction and Weight Calculation The Jieba word segmentation tool was used to extract keywords from the cleaned text, which were then matched against a core keyword list to calculate the TF-IDF value. For example, the keywords extracted from the above road property text were "truck breakdown", "occupying the slow lane", and "slow traffic behind". Among them, the TF of "slow traffic behind" was 1, and the IDF was log(5000 / (1+300))≈log(16.1)≈2.78. Therefore, TF-IDF=1×2.78=2.78. Step 3: Output topic probability Input keywords and weights into the trained LDA model, and it outputs topic feature vectors. For example: The above text from the road maintenance post → model determines "road congestion ( The probability of "traffic slowdown" is 0.7 (due to "slow traffic behind"), and the probability of "traffic accident" is 0.7 (due to "slow traffic behind"). The probability of "0.2" is due to "truck breaking down". =(0.7,0.2,0,0,0.1); The service area feedback from the collaborative control post was "The gas station's fuel dispenser is broken, I waited for half an hour" → the keyword "fuel dispenser broken" was extracted, and the model determined "facility malfunction". The probability of ") is 0.9, therefore =(0,0,0.9,0,0.1).

[0059] S4.2 Convolutional Neural Networks (CNN): Feature Extraction from Structured Temporal Data (process as follows) Figure 3 (As shown) The structured time-series data from the five monitoring and operation positions (such as "5-minute traffic flow at a cross-section" for the monitoring and operation position, "10-minute vehicle count at toll booths" for the toll station operation position, and "traffic flow changes 1 hour after an accident" for the highway police position) need to have trend features (such as "peak growth" and "congestion dissipation") extracted before they can be correlated with other data. This step uses 1D-CNN (one-dimensional convolutional neural network) because it excels at capturing the local continuous features of time-series data, which aligns with the "continuous change" characteristic of traffic flow.

[0060] 1. Definitions of Terms and Symbols T: Time series sampling window, combined with the five-post linkage response timeliness (fastest response in 5 minutes), set T=12 (sampling once every 5 minutes, covering 1 hour, which is enough to capture short-term trends). V: Traffic time-series vector ,in Let t be the traffic flow at the t-th sampling point (unit: vehicles / 5 minutes), such as This represents the traffic flow from 9:00 to 9:05. : Convolution kernel set (A total of 8 convolutional kernels, covering different trend features), each convolutional kernel A 3D vector (covering 3 consecutive sampling points, corresponding to a 15-minute local trend), such as... =(0.2,0.5,0.3) focuses on capturing "growth trends"; Convolution output vector (Total 10 values), among which , It is a ReLU activation function (filtering out negative values). This is a bias term (to avoid outputting 0); F: Trend feature vector after pooling (Compressed into 4 core trends, tailored to transportation scenarios), among which: Traffic volume increases during the morning rush hour (7:00-9:00 AM). Peak hours are stable (flow fluctuations are small from 10:00 to 16:00); Traffic congestion (traffic surge followed by sustained high traffic volume, vehicle speed ≤20km / h); : Resumption of congestion (flow gradually decreases to normal after congestion); vector satisfies , This represents the probability that the time series data belongs to the k-th trend.

[0061] 2. Model Building and Training (Based on Historical Traffic Data from Five Positions) Training data preparation: Collect time-series data from five positions over the past three months (monitoring and operations personnel provide traffic flow across the entire road network, and toll collection and operations personnel provide traffic flow at the entrances and exits of each toll station), and label the data with trend tags according to "road segment + time period": Morning rush hour growth: For example, "On the main line of XX Expressway at K100, the traffic flow increased from 100 vehicles / 5 minutes to 300 vehicles / 5 minutes between 7:00 and 8:00"; Traffic congestion status: such as "XX Expressway ramp K50, from 14:00 to 15:00 the traffic flow is maintained at 250 vehicles / 5 minutes (normal is 150), and the speed is 15km / h"; A total of 100,000 time-series vectors were collected and divided into a training set (80,000 vectors) and a validation set (20,000 vectors) in an 8:2 ratio.

[0062] 1D-CNN model architecture design: Input layer: Received traffic time sequence vector V (dimension 12×1); Convolutional layer: 8 3D convolutional kernels, stride = 1 (slides 1 sampling point at a time), extracting local trends (e.g.) (growth rate) Activation layer: ReLU function ( ), filtering out meaningless negative features; Pooling layer: Max pooling (takes the maximum value of every 4 convolution outputs), compressing the 10 convolution results into 4 (corresponding to 4 trends). Output layer: softmax function, outputs the probabilities of the four trends ( ).

[0063] Training optimization: With cross-entropy loss function ( , The true value of the label, such as the morning rush hour sample. Using the Adam algorithm as the target, iterative training is performed: Learning rate = 0.001 (to avoid iteration oscillation); Iterate for 50 rounds (evaluate with a validation set after each training round, and stop when the accuracy reaches 92%).

[0064] 3. Real-time feature extraction (covering time-series data types across five job roles) For time-series data in the S3 standardized data, extract features according to the following steps (all guard post time-series data need to be processed): Step 1: Timing Vector Extraction Extract time-series data for a one-hour window from the data of each position, forming a vector V. For example: The "Traffic Flow at Section K150 of XX Expressway from 9:00 to 10:00" provided by the monitoring and operation team is: V=(180,220,250,280,300,290,260,230,200,180,160,150) (unit: vehicles / 5 minutes). The toll station operation staff provided the following data: "Number of vehicles exiting XX toll station from 9:00 to 10:00": V=(50,60,70,65,60,55,50,45,40,35,30,25); Step 2: Convolution operation (extracting local trends) Each convolutional kernel slides to compute local features. Using the first convolutional kernel... =(0.2,0.5,0.3) (Capturing growth trends) Calculate V for the monitoring operations role: ; ; And so on, we get ; Step 3: Pooling and Trend Probability Output Output of 8 convolution kernels Pooling is performed separately (taking the maximum value from every four values), and the final trend feature vector is obtained through the output layer. For example: Monitoring operations personnel's V (increase then decrease) → Model determination "recovery dissipation ( The probability of "0.1, 0.1, 0, 0.8" is 0.8, therefore F = (0.1, 0.1, 0, 0.8). The V (continuously decreasing) value for the toll collection and operation position → model determination of "peak stability" The probability of "0.7" is 0.7 (due to the small decrease), so F = (0, 0.7, 0, 0.3).

[0065] S4.3 Bayesian Network: Calculation of the probability of data association in five positions (process as follows) Figure 4 (As shown) The data from the five posts have natural causal relationships (e.g., "traffic accidents at the highway police post" may lead to "congestion at the monitoring and operation post", "rescue at the road property maintenance post", and "public opinion at the traffic law enforcement post"). It is necessary to quantify the probability of the association through Bayesian networks to clarify "who affects whom" and the degree of influence.

[0066] 1. Definitions of Terms and Symbols U: A set of network nodes (covering the five core data types), U={A,B,C,D,E}, where each node is a binary variable (1=occurred, 0=not occurred): A (Highway Police Post): Traffic Accident (A=1: Occurred; A=0: Not Occurred); B (Traffic Enforcement Post): Traffic Control (B=1: Implemented; B=0: Not Implemented); C (Monitoring and Operations): Traffic Congestion (C=1: Occurred; C=0: Not Occurred); D (Road Property Management Post): Clearance and Rescue (D=1: Activated; D=0: Not Activated); E (Collaborative Management Position): Public opinion complaints (E=1: exist; E=0: do not exist); Node relationships (causal chains): Directed edges are set based on the business logic of the five positions: A→C (accidents cause congestion), B→C (control measures affect congestion), C→D (congestion triggers rescue), C→E (congestion causes complaints); P(X): Prior probability (the probability of X occurring alone), such as P(A=1) being the traffic accident rate; P(X|Y): Conditional probability (the probability that X occurs when Y occurs), such as P(C=1|A=1) which is the probability of congestion after an accident occurs; P(X|Y=1): Posterior probability (the probability that X is the cause when Y occurs), such as P(A=1|C=1) which is the probability that congestion is caused by an accident.

[0067] 2. Model Building and Training (Based on Historical Association Data of Five Positions) Conditional probability table (CPT) construction: Analyze the data from the five-position linkage event over the past year (a total of 5000 events) and calculate the prior probability and conditional probability: Prior probability (probability of occurrence alone): P(A=1)=0.05 (250 accidents occurred out of 5000 events); P(B=1)=0.03 (150 times of implementation of control); P(C=1)=0.1 (congestion occurs 500 times); Conditional probability (probability of causal association): P(C=1|A=1)=0.8 (200 out of 250 accidents caused traffic congestion); P(C=1|B=1)=0.9 (135 out of 150 traffic control measures caused congestion). P(D=1|C=1)=0.7 (rescue was initiated 350 times out of 500 traffic jams); P(E=1|C=1)=0.6 (300 out of 500 traffic jams resulted in complaints). Model validation: Validated using 20% ​​of historical data (1000 events): Given known node states (e.g., A=1, B=0), the predicted probability of C=1 is 0.8, which matches the actual probability by 91%, meeting the accuracy requirements.

[0068] 3. Real-time correlation probability calculation (five-position data linkage application) Input real-time data from S4.1 (public opinion complaints) and S4.2 (congestion trends) to calculate the correlation probability of each node: Step 1: Input real-time evidence The evidence is "known events," derived from the outputs of S4.1 and S4.2: From S4.1 (public opinion text), we get E=1 (the public opinion complaint is about congestion, with a probability of 0.9 ≥ 0.5, so it is determined to have occurred). From S4.2 (traffic trend), we get C=1 (the probability of congestion trend is 0.8 ≥ 0.5, so it is determined to occur). Step 2: Calculate the posterior probability (core association) Based on Bayes' theorem (P(Y) is a normalization constant), calculate the "cause of congestion (C=1)" and the "impact of congestion": ① Calculate the total probability P(C=1) of C=1: P(C=1)=P(C=1|A=1)P(A=1)+P(C=1|B=1)P(B=1)+P(C=1|A=0,B=0)P(A=0)P(B=0); where P(C=1|A=0,B=0)=0.05, which is the probability of congestion without accidents or traffic control. Substituting the values, we get: P(C=1)=0.8×0.05+0.9×0.03+0.05×0.95×0.97≈0.04+0.027+0.046=0.113; ② Calculate the probability that congestion is caused by an accident: ; ③ Calculate the probability that congestion is caused by traffic control: ; ④ Calculate the probability of traffic congestion triggering a rescue: P(D=1|C=1)=0.7 (directly taken from the CPT table); Output results (five-position correlation conclusion): Currently, there is a 35% probability that traffic congestion is caused by traffic accidents (highway police posts) and a 24% probability that it is caused by traffic control (traffic enforcement posts); there is a 70% probability that congestion will require clearing and rescue (road property maintenance posts) and a 60% probability that it will trigger more public opinion (traffic enforcement posts).

[0069] S4.4 Surface Fitting: Spatiotemporal Completion of Continuous Data in Five Positions (process as follows) Figure 5 (As shown) The continuous data of the five positions (such as "cross-sectional flow" of the monitoring and operation position and "distribution of rescue personnel" of the road property maintenance position) have sampling intervals (such as once every 5 minutes). It is necessary to fill in the blank time period / location data through surface fitting, and the completion should prioritize the fitting of highly correlated data (such as the data "highly correlated with congestion" in S4.3).

[0070] 1. Definitions of Terms and Symbols Space-time coordinates The station number for the road section (unit: km, e.g., K150 is recorded as K150). =150), Time (unit: hour, e.g., 9:00) =9); Discrete sampled data values, i.e. Actual measured data at the location: Monitoring and Operations Position: = Traffic flow (vehicles / 5 minutes); Road property maintenance post: =Number of rescue personnel (people); Quadratic surface fitting function (output arbitrary) (continuous values ​​at the location), in the form of: ; in ~ These are the fitting coefficients (required solution); The weight of the i-th sampling point (the higher the correlation, the greater the weight) is obtained by normalizing the posterior probability in S4.3. ; in This represents the probability of association between the data at this point and the current core event (such as congestion) (taken from S4.3).

[0071] 2. Model Building and Training (Based on Continuous Data from Five Job Positions) Training data preparation: Discrete sampling points from five positions were collected over the past three months (e.g., 1000 traffic flow sampling points from the monitoring and operation position and 500 rescue personnel sampling points from the road maintenance position), and matched with the correlation probabilities calculated by S4.3. (e.g., the correlation probability between traffic sampling points and congestion).

[0072] Solving for the fitting coefficients: The goal is to minimize the "weighted error" (the error of highly weighted sampling points has a greater impact): ; Will Substituting the values ​​into the system of linear equations, we can solve for the coefficients using the least squares method. ~ (For example, using Python's numpy.linalg.lstsq function).

[0073] Model validation: Use 20% sampling points for verification to ensure that the average absolute error between the fitted value and the measured value is ≤5% (e.g., if the measured flow rate is 200 vehicles, the fitted value is 190-210 vehicles).

[0074] 3. Real-time data completion (covering continuous data types for five job positions) For continuous discrete sampling points in S3 standardized data, complete the data using the following steps: Step 1: Matching sampling points with weights Extract sampling points And obtain its association probability from S4.3. Calculate weights .For example: Traffic sampling points for monitoring operations personnel: , =0.35 (associated with congestion); , =0.4; Weight , =0.53; Step 2: Solve for the fitting coefficients Substituting the sampling points and weights into the weighted least squares formula, we can obtain the coefficients (example): =50, =2, =10, =0.01, =0.5, =0.2; Step 3: Fill in the blank data use Calculate the value for the blank time period / location. For example: Complete =150, =Flow rate of 9.1 (9:06): (The units here are intermediate calculated values; actual values ​​need to be adjusted according to the units.) Supplementing the rescue personnel at the road maintenance post: It is known that 9:00 ( =9) There are 3 people, 9:20 ( =9.33) There are 5 people, complete 9:10 ( =9.17) The number of people is approximately 4 (due to the high correlation weight, the fit is closer to reality).

[0075] S4.5 Algorithm Collaborative Interaction and Fusion Result Output The four algorithms work together in a deep collaboration of "data transfer and parameter feedback" to ultimately output a fusion dataset covering all five job positions, providing a unified input for subsequent job-specific analysis.

[0076] 1. Algorithm Collaboration Logic (End-to-End Interaction) NLP → Bayesian Network: The public opinion topic features output by S4.1 (such as E=1) serve as evidence nodes of the Bayesian network and directly participate in the calculation of posterior probability (such as P(C=1|E=1)). CNN → Bayesian Network: The traffic trend features output by S4.2 (e.g., C=1) serve as the core evidence for the Bayesian network, and the prior probability is adjusted (e.g., P(C=1) is increased from 0.1 to 0.8). Bayesian Network → Surface Fitting: Association Probabilities Output by S4.3 Weights converted to surface fitting Ensure that highly relevant data (such as traffic near the incident site) dominates the completion process; Surface fitting → Fusion result: The completed continuous data is integrated with the structured features and correlation probabilities of S4.1-S4.3 to form a complete data chain.

[0077] 2. Output of merged datasets (including application scenarios for five job positions) The fused dataset contains three core types of information, each labeled with its source post: Association probability table (example):

[0078] Continuous spatiotemporal characteristics (example): “Continuous traffic flow distribution on the K150 section of XX Expressway from 9:00 to 10:00 (including completion): 9:00 (200 vehicles) → 9:06 (210 vehicles) → 9:12 (215 vehicles) → ... → 10:00 (150 vehicles)” (Source: Monitoring and Operations Staff + Fitting and Completion); Data association chain (example): "Highway Police Post (Traffic Accident A=1) → Monitoring and Operation Post (Congestion C=1) → Road Property Management Post (Rescue D=1) → Traffic Enforcement Post (Public Opinion E=1)" (the probabilities on the chain are 35%, 70%, and 60% respectively.

[0079] The final output of S4 (the fused dataset) can be formally represented by an expression: 1. Merge dataset overall expression Let the fused dataset be ,but: ,in: Five-position data association probability set (from S4.3 Bayesian network); Continuous spatiotemporal characteristic functions (from S4.4 surface fitting); Data association chain (based on) (Constructed causal chain).

[0080] 2. Set of Association Probabilities

[0081] The core correlation probability among the five data points is expressed as: ,in: P(C) represents the posterior probability that traffic congestion ($C) is caused by a traffic accident (A, highway police post) (e.g., 35%); P(B|C) represents the posterior probability that traffic congestion (C) is caused by traffic control (B, traffic enforcement post) (e.g., 24%); P(D|C) represents the conditional probability that traffic congestion (C) triggers road clearing and rescue (D, road maintenance post) (e.g., 70%); P(E|C) represents the conditional probability that traffic congestion (C) triggers public opinion complaints (E, collaborative management post) (e.g., 60%). All probabilities are derived from the Bayesian network calculation results P(X|Y=1) of S4.3. 3. Continuous spatiotemporal characteristics

[0082] For the surface fitting function, we directly use the quadratic surface expression from S4.4. The output values ​​are defined according to the data type: Monitoring and Operations: Traffic flow (vehicles / 5 minutes); Road Maintenance and Management: Number of rescue personnel (people); Toll Collection and Operations: Number of vehicles entering and exiting the toll station (vehicles / 10 minutes).

[0083] 4. Data Association Chain

[0084] The probability chain for the causal association of the five job data is expressed as follows: , where: symbol " " indicates "triggered by probability P"; Traffic accidents (A) at highway police posts cause congestion (C) at monitoring and operation posts with probability P(A|C). The congestion at the monitoring and operation post (C) triggers a rescue operation at the road maintenance post (D) with probability P(D|C). The congestion (C) at the monitoring and operation post triggers public opinion (E) at the collaborative control post with probability P(E|C); the symbol " "" indicates that two related branches exist in parallel.

[0085] Through the above expression, the final output of S4 achieves a formal integration of "probability quantization + continuous features + correlation logic".

[0086] S5 Job Role Data Analysis: In-depth Application Based on S4 Fusion Results Core objective: To determine the "association probability" of each position based on the output of S4. "Continuous spatiotemporal characteristics ()" "Data association chain ()" Based on the core principle of "data analysis", we conduct targeted analysis in conjunction with our own responsibilities, replacing the traditional analysis model that "only relies on raw data".

[0087] 5.1 Highway Police Post: Accident Handling Priority Analysis Based on Association Probability 1. Core input: S4 output P(A|C) (probability of an accident causing congestion), (Accident → Traffic Congestion → Rescue); 2. Analysis Logic: If P(A|C)≥50% (high correlation): the accident is determined to be the main cause of the current congestion, and police forces are dispatched to the scene with priority (e.g., "the probability of the K150 accident causing congestion is 60%" → arrive and handle the situation within 10 minutes). If 30% ≤ P(A|C) < 50% (intermediate correlation): Combine Check if a rescue has been initiated (D=1). If a rescue has been initiated, postpone it and focus on tracking unrescued incidents. If P(A|C) < 30% (low correlation): reduce priority and prioritize handling high correlation incidents.

[0088] 3. Output: Accident Handling Priority List (including "Road Section, Priority, and Suggested Handling Time").

[0089] 5.2 Traffic Enforcement Posts: Risk Prevention and Control Analysis Based on Public Opinion Probability 1. Core input: S4 output P(E|C) (the probability of congestion triggering public opinion), (Continuous characteristics of congestion diffusion); 2. Analysis Logic: If P(E|C)≥60%: anticipate the potential spread of public opinion and release "cause of congestion + estimated relief time" in advance through the information board (e.g., "K200 is congested due to an accident, expected to ease in 1 hour"). Combination Predict the congestion spread path (e.g., "Currently congested at K150, it will spread to K160 in 1 hour") and issue early warning information on the K160 information board in advance; For data on overloaded vehicles, combined with (Overweight → Probability of accidents) Prioritize intercepting vehicles that are "overweight by ≥30% and pass through accident-prone road sections".

[0090] 3. Output: "Suggestions on Public Opinion Prevention and Control and Over-limit Interception".

[0091] 5.3 Monitoring and Operations Position: Full-process status quo tracking based on relational chains 1. Core input: S4 output (Accident → Traffic Congestion → Rescue → Public Opinion) (Continuous flow distribution); 2. Analysis Logic: Track the status of each link in the chain: If “Accident (A=1) → Congestion (C=1)” has occurred, but “Rescue (D=0)” has not been initiated (and P(D|C)=70%), immediately remind the road maintenance post to initiate rescue; based on Analyze traffic recovery trends: If the traffic flow on a congested road section decreases for three consecutive sampling points (e.g., from 280 vehicles → 250 vehicles → 220 vehicles), predict that the congestion will ease in one hour, and update the response guidelines for call inquiries accordingly.

[0092] 3. Output: Road Network Status Tracking Report (including "Current Status, Abnormal Factors, and Recovery Prediction").

[0093] 5.4 Road Maintenance Post: Analysis of Rescue Resource Allocation Based on Continuous Characteristics 1. Core input: S4 output (Spatiotemporal distribution of rescue demand), P(D|C) (probability of congestion triggering rescue); 2. Analysis Logic: according to Identify areas with high traffic volume (e.g., "K150-K160 traffic volume ≥ 250 vehicles / 5 minutes") and preposition clearing vehicles (e.g., K155 on standby) around these areas. If P(D|C)≥70% and Traffic volume continues to rise: Increase rescue personnel (e.g., from 2 to 4) to avoid insufficient rescue resources; Combining road property loss data with If "K180 guardrail is damaged" and The system indicates that traffic flow on this section of road will increase in 1 hour, and repairs will be prioritized (to avoid secondary accidents).

[0094] 3. Output: "Rescue Resource Allocation and Road Property Repair Plan".

[0095] 5.5 Collaborative Management and Control Position: Service and Equipment Assurance Analysis Based on Continuous Characteristics Mechanical and electrical maintenance position: 1. Core Input: (Peak traffic sections) (Equipment failure → probability of congestion); 2. Analysis Logic: Increase the inspection frequency of information boards and monitoring equipment on peak traffic sections (such as K150) (from once every 2 hours to once every hour); prioritize the repair of equipment with a "fault → congestion probability ≥ 40%" (such as mainline information boards).

[0096] Fee collection and operation position: 1. Core Input: (Station traffic forecast) (Congestion → Slow traffic at toll booths); 2. Analytical Logic: If If it is predicted that "the inbound traffic at toll station XX will increase by 50% in one hour", the backup lane will be opened in advance; if "congestion has already caused slow traffic at the toll station" ( (As shown), traffic police were coordinated to guide vehicles to adjacent toll stations.

[0097] S6 Comprehensive Analysis: Integrating S4 fusion results to generate a dynamic situation report. Core objective: To monitor and operate the integration of analysis results from various positions in S5, and combine them with the fused data from S4 (correlation probability, continuous features) to generate a comprehensive report that includes "risk diffusion, resource matching, and decision-making recommendations," replacing the traditional model of "only summarizing the conclusions of each position."

[0098] 6.1 Assessing Core Inputs S4 output: (Probability of full association) (Continuous spatiotemporal characteristics) (Full-link association); S5 Outputs: Analysis reports for each position (such as "Accident Handling Priority List" and "Rescue Resource Allocation Plan").

[0099] 6.2 Analysis of Key Modules 1. Risk Diffusion Prediction Module based on and For example, "Current K150 accident (A=1) → Congestion (C=1)". The display shows that the congestion will spread to K160 in 1 hour, and P(E|C)=60% → it is predicted that public opinion will be affected at K160. Output "Risk Spread Path and Time Sequence": K150 (Current Congestion) → K160 (Congestion 1 hour later) → K160 Public Opinion (1.5 hours later).

[0100] 2. Resource Matching Degree Assessment Module Compare the resource requirements of each position with the existing resources: For example, "Road property maintenance position requires 4 rescue teams (based on P(D|C)=70%), but currently only 2 teams are available → resource gap of 2 teams"; Combination Priority: Resources should be allocated to areas with "high correlation risk (P≥60%)" (e.g., first meet the needs of K150 rescue, then coordinate support for adjacent road sections).

[0101] 3. Decision Recommendation Module In response to the risk of spread: it is recommended to "set up traffic guidance signs in advance at K160 to divert vehicles to the branch expressway"; Regarding the resource shortage: It is recommended to "dispatch two rescue teams from K120 to K150, and have them arrive within one hour".

[0102] 6.3 Output: Traffic Situation Dynamics Analysis Report It includes "current road network status, risk warning (including spread prediction), resource matching status, and decision-making suggestions", and is updated every 30 minutes, synchronized to the five posts and the joint command center.

[0103] S7 Decision Distribution and Execution: Precise Command Issuance Based on S4 Fusion Results Core objective: The Joint Logistics Command Center, based on the S6 assessment report and combined with the correlation probability of S4 (… ) and continuous features ( ), formulate instructions that are "clearly prioritized and precisely allocated to resources" to avoid traditional "generalized instructions" (such as "strengthen patrols").

[0104] 7.1 Instruction Priority Hierarchy (based on) )

[0105] 7.2 Quantification of Resource Allocation (based on) ) Rescue forces: According to The distribution of peak traffic points (e.g., "K150-K160 traffic peak at K155" → rescue forces deployed at K155). Equipment investment: For road sections with consistently high traffic volume (such as K150 with an average daily traffic volume of ≥300 vehicles / 5 minutes), add one set of temporary information boards; Personnel allocation: Based on the toll collection and operation staff... Based on the traffic flow forecast at the toll station (e.g., "Traffic flow at XX toll station will reach its peak at 10:00"), two additional toll collectors will be deployed in advance.

[0106] 7.3 Instruction Execution Trace The joint operations platform displays the execution status of instructions in real time (not started / in execution / completed). For instructions that "exceed the execution time limit and have a priority of ≥ Level 2" (such as a Level 1 instruction that has not started within 10 minutes), it automatically reminds the instruction receiving position and superior. If new situations arise during execution (such as new incidents), it will be combined with S4 real-time updates. Adjustment instructions (such as upgrading the original level 2 instruction to level 1).

[0107] S8 Execution Feedback: Synchronously collects feedback from both "algorithm performance" and "instruction execution" dimensions. Core objective: Not only to provide feedback on the execution results of instructions, but also to collect "accuracy feedback" of the S4 algorithm output (such as whether the correlation probability matches reality), so as to provide a basis for the optimization of the S9 algorithm and the improvement of data quality.

[0108] 8.1 Instruction Execution Feedback After executing the instructions, each staff member should provide feedback: Execution results (e.g., "Arrived at K150 to handle the accident" or "The backup lane at XX toll station has been activated"); Actual results (e.g., "After the accident was handled, the traffic flow on K150 decreased from 280 vehicles to 200 vehicles" and "Traffic efficiency increased by 30% after the backup lane was opened"). Reasons for incomplete (e.g., "Rescue vehicles could not arrive on time due to heavy rain").

[0109] 8.2 S4 Algorithm Performance Feedback (Core New Feature) Association probability accuracy: comparison Compared with the actual situation (e.g., "P(A|C)=60%, the actual congestion was indeed caused by an accident → accurate; P(A|C)=60%, the actual congestion was caused by traffic control → too low"). Accuracy of continuous features: Comparison Predicted vs. Actual values ​​(e.g., "Predicted traffic flow at K15010:00: 220 vehicles, actual: 215 vehicles → error 2.3%; predicted: 220 vehicles, actual: 150 vehicles → error 31.8%"). Linkage chain integrity: verification Are there any omissions of actual connections (e.g., "the actual 'accident → congestion → public opinion' has occurred, but...")? (Not including public opinion → omission).

[0110] 8.3 Feedback Data Storage All feedback data is stored in a sequence of "instruction number + algorithm output ID", for example: The instruction Z20XX0728001 corresponds to the output ID456 of S4 with P(A|C)=60%, and the feedback is "accurate"; This facilitates S9's ability to trace the historical accuracy of an algorithm's output.

[0111] S9 Iterative Optimization: Feedback-based Full-Chain Optimization of Algorithm-Data-Process Core objective: Based on the feedback from S8, optimize the algorithm parameters of S4 and the data acquisition and preprocessing rules of S1-S3 in a targeted manner to form a closed loop of "data → fusion → application → optimization".

[0112] 9.1 S4 Algorithm Parameter Optimization (Based on Algorithm Performance Feedback) Bayesian Network: If P(A|C) underestimates the actual correlation multiple times (e.g., "predicted 30%, actual 60%" 3 out of 5 feedbacks), adjust the conditional probability table P(C|A) (e.g., increase it from 0.8 to 0.9). Surface fitting: If If the error consistently exceeds 10% on a certain road segment (e.g., K150), increase the weight of the sampling points on that road segment. The sampling frequency was increased from 0.47 to 0.6, and the sampling frequency was adjusted (from once every 5 minutes to once every 3 minutes). CNN temporal features: If the accuracy of identifying a certain trend (such as "morning rush hour growth") is less than 85%, increase the historical training data of that trend (from 8,000 to 10,000).

[0113] 9.2 Optimization of S1-S3 Data Rules (Based on Algorithm Performance Feedback) S1 Collection: If text data causes S4.1 topic identification errors due to "missing locations" (e.g., "traffic jam" without road segment label → large topic probability error), strengthen the collection specifications (mandatory requirement that "text must contain road segment", otherwise it cannot be submitted); S3 Preprocessing: If the time series data still has large errors after outlier correction (e.g., the deviation between the replacement value and the actual value is >20%), optimize the correction rules (replace "2-point average" with "weighted average of the three points before and after").

[0114] 9.3 Process Optimization (Based on Instruction Execution Feedback) If a certain type of instruction (such as "rescue force deployment") is repeatedly not completed on time due to "inaccurate resource location", combined with Optimize resource pre-positioning rules (e.g., "deploy rescue forces to the predicted peak point 2 hours before the traffic peak"); The feedback data is summarized and feedback is collected every quarter, and an "Optimization Report" is output. The "Five-Post Joint Operation Manual" (including algorithm parameters, data rules, and instruction flow) is updated.

[0115] S1-S3 provide "algorithm-adaptive" data input for S4, S5-S7 achieve accurate analysis and decision-making based on the fusion results of S4, and S8-S9 optimize the entire process through feedback to form a complete closed loop.

Claims

1. A traffic situation awareness data fusion analysis method in a five-hill-in-one mode, characterized by, Comprise: S1: five post data collection, the five post is high speed public security post, traffic law enforcement post, monitoring operation post, road property management post, collaborative control post, the collected data includes traffic accident data of high speed public security post, public opinion monitoring data of traffic law enforcement post, traffic flow data of monitoring operation post, rescue resource data of road property management post, equipment operation data of collaborative control post; S2: data aggregation, all data collected in S1 is classified and aggregated according to post + data type to form an original data set; S3: data preprocessing, the original data set of S2 is cleaned, standardized, redundant information is removed, and the data format is unified to obtain preprocessed data; S4: multi algorithm fusion processing, based on the preprocessed data, the unstructured text is converted into topic feature vector through natural language processing, traffic flow time sequence feature is extracted through convolutional neural network, data correlation probability is calculated through Bayesian network, and space blank data is completed through surface fitting, finally the fusion result including fusion data set, correlation probability set and data correlation chain is outputted; S5: post data analysis, each post carries out special analysis based on the fusion result of S4 to obtain post analysis result; S6: comprehensive research and judgment, the monitoring operation post integrates the post analysis result of S5 to generate traffic situation awareness report.

2. The method according to claim 1, wherein, Also include: S7: instruction distribution, execute instruction is issued to five post based on traffic situation awareness report of S6; S8: feedback update, the new data after instruction execution is fed back to S2 to update the original data set; S9: iterative optimization, the effectiveness of S1-S8 is evaluated according to the feedback of S8, and the algorithm parameters of S4 are adjusted.

3. The method according to claim 2, wherein, The topic feature vector outputted by the natural language processing step of S4 and the traffic flow time sequence feature outputted by the convolutional neural network step are both input evidence of the Bayesian network, the conditional probability table of the Bayesian network is generated based on the historical correlation data of five post, and the update frequency of the conditional probability table is consistent with the feedback update frequency of S8.

4. The method according to claim 2, wherein, In the surface fitting step of S4, the sampling point weight is obtained by normalizing the correlation probability outputted by the Bayesian network, if the fitting deviation of a certain type of data is more than 30%, the fusion rule of this type of data is adjusted in S9, and the adjusted rule needs to be updated to the fusion logic of S4.

5. The method according to claim 1, wherein, The fusion data set outputted by S4 contains the correlation relationship of five post data, which is used as the priority basis for post analysis of S5, and the determination of the correlation relationship is completely based on the actual business logic of five post.

6. The method according to claim 1, wherein, The fusion data set outputted by S4 needs to contain data source, processing time and correlation probability, and each data entry corresponds to the original data number collected by S1, which ensures that it can be traced back to the specific collection time of the specific post.

7. The method according to claim 2, wherein, The traffic situation awareness report of S6 needs to clearly quote the fusion data set of S4 and the post analysis result of S5, the report content includes pre instruction execution situation data, post instruction execution situation data, deviation reason, if the deviation is more than 30%, the fusion rule of S4 is adjusted in S9, and the adjustment result needs to be verified in the next S1-S8 cycle.

8. The method according to claim 2, wherein, The equipment operation data of the collaborative control post is fused by S4, and serves as the basis for the allocation of rescue resources by the road property management and protection post. The data is fed back to the original data set in S8, and used for effect evaluation in S9.

9. The method according to claim 2, wherein, If the correlation probability calculation rule of the Bayesian network in S4 needs to be adjusted, it needs to be based on the actual deviation data fed back by S8. The adjusted rule needs to record the adjustment time and reason in S9.

10. The method according to claim 2, wherein, The closed-loop end condition of S4 to S9 is that in the last 3 consecutive cycles, the deviation of the correlation probability output by S4 from the actual situation is ≤10%. At this time, it is determined that the current fusion rule is effective, the rule adjustment in S9 is stopped, and S1-S8 continues to be executed at the normal frequency.

Images