Mask-sensing large-scale model multi-agent traffic state diagnosis method

By constructing a mask-sensing large-scale model for multi-agent traffic state diagnosis, the reliability of sparse traffic data and the problem of multi-agent collaboration are solved, thereby improving the reliability and computational efficiency of traffic state analysis and ensuring the safety and real-time performance of traffic management.

CN123119682APending Publication Date: 2026-11-17HUAIYIN INSTITUTE OF TECHNOLOGY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611435997.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-16
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing traffic condition analysis methods struggle to distinguish between real traffic condition changes and uncertainties caused by insufficient observation. They lack effective utilization of sparse traffic data, and large-scale traffic analysis lacks multi-agent task collaboration and selective scheduling mechanisms, leading to wasted computational resources and unstable diagnostic results.

Method used

A multi-agent traffic state diagnosis method based on a mask perception big model is proposed. This method constructs a regional traffic state matrix and an observation mask matrix by acquiring basic data, and utilizes multiple agents to perform data quality assessment, regional traffic state analysis, and spatial correlation analysis. Combined with a diagnostic priority selective triggering mechanism, the big model is dynamically invoked for diagnosis.

Benefits of technology

It improves the reliability and interpretability of sparse traffic state analysis, reduces computational resource consumption, enhances the comprehensiveness and real-time processing capability of traffic state diagnosis, and ensures the traceability and reliability of diagnostic results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123119682A_ABST
    Figure CN123119682A_ABST
Patent Text Reader

Abstract

This invention discloses a mask-sensing large-scale model multi-agent traffic state diagnosis method, belonging to the fields of intelligent transportation systems, traffic big data analysis, and artificial intelligence intelligent diagnosis technology. This method is designed for traffic observation data acquired from ride-hailing vehicles, taxis, and other mobile sensing devices. Under conditions of spatiotemporal sparsity, uneven coverage, and data gaps in traffic state observations, it constructs a regional traffic state matrix and an observation mask matrix, integrating mask-sensing traffic state estimation, large-scale model multi-agent collaborative analysis, and a selective intelligent invocation mechanism to achieve the identification of regional traffic state changes, anomaly analysis, and structured output of diagnostic results. This invention can be applied to scenarios such as urban road traffic operation monitoring, traffic state assessment, congestion analysis, and intelligent traffic management decision-making.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation systems, traffic big data analysis and artificial intelligence intelligent diagnosis technology, and in particular to a method for diagnosing traffic conditions using a masked perception large-scale model with multiple agents. Background Technology

[0002] With the continuous increase in the number of urban motor vehicles and transportation demand, the operational status of urban road traffic exhibits dynamic changes. Intelligent Transportation Systems (ITS) typically rely on traffic condition perception data to conduct road operation monitoring, condition assessment, and decision support. Currently, traffic condition perception data mainly comes from fixed detection equipment, video surveillance equipment, and mobile sensing equipment. Among these, trajectory data generated by mobile probe vehicles such as ride-hailing vehicles and taxis has the characteristics of wide coverage, low acquisition cost, and strong spatial expansion capability, and is gradually becoming an important data source for urban road traffic condition perception.

[0003] However, influenced by factors such as the number of mobile probe vehicles, their operating trajectories, and time-of-day distribution, mobile sensing traffic data typically exhibits significant spatiotemporal sparsity and uneven coverage, with some areas and time periods lacking effective observations. Existing traffic state analysis methods usually interpolate, complete, or predict missing data before conducting state assessments based on the processed data, making it difficult to distinguish between actual traffic state changes and uncertainties caused by insufficient observation. Furthermore, existing traffic prediction methods primarily focus on the accuracy of future traffic state predictions, often lacking a comprehensive analysis of the data support for the prediction results, observation coverage, and missing patterns. This makes it difficult to directly form reliable and interpretable diagnostic criteria based on the prediction results.

[0004] In recent years, large-scale models have been increasingly applied to intelligent traffic analysis due to their strong information understanding, integration, and reasoning capabilities. However, existing large-scale model traffic analysis methods typically employ a single agent or single-invocation approach, making it difficult to allocate tasks based on traffic data quality, regional traffic conditions, and spatial relationships. Furthermore, the analytical basis and data support for the output results are not easily traceable. In addition, uniformly invoking large models across all regions and time windows incurs high computational costs, while some areas with stable traffic conditions do not require complex reasoning.

[0005] In summary, existing sparse traffic state analysis techniques still have the following shortcomings: First, it is difficult to use missing observation information to characterize the data support level of traffic state; second, there is a lack of effective connection between traffic state prediction and intelligent diagnosis, making it difficult to comprehensively utilize state prediction, data reliability, and spatial correlation information; third, large-scale traffic analysis lacks multi-agent task collaboration and structured diagnostic mechanisms; and fourth, there is a lack of selective scheduling mechanisms for large-scale model calls, making it difficult to balance diagnostic effectiveness and computational efficiency. Summary of the Invention

[0006] Purpose of the invention: To address the above problems, the purpose of this invention is to provide a method for diagnosing traffic conditions in a large-scale mask-sensing model with multiple agents.

[0007] Technical solution: The mask-sensing large-scale multi-agent traffic state diagnosis method of the present invention includes the following steps:

[0008] Step 1: Obtain the basic data required for the sparse traffic condition diagnosis task. The basic data includes mobile sensing traffic observation data, road network data, traffic area division data, and method operation parameter information.

[0009] Step 2: Based on the acquired regional traffic status data, construct the regional traffic status matrix and its corresponding observation mask matrix to obtain the temporal coverage features, regional historical coverage features, continuous missing features, and spatial correlation matrix.

[0010] Step 3: Estimate the regional traffic status for the target time period under sparse traffic observation conditions to obtain the predicted traffic status for the target area at the next time step.

[0011] Step 4: Based on the predicted traffic conditions, traffic condition change characteristics, and observation reliability information, generate rule-based initial traffic condition evaluation results;

[0012] Step 5: Construct a structured evidence set for traffic condition diagnosis;

[0013] Step 6: Construct multiple large language model agents; the large language model agents include at least a data quality assessment agent, a regional traffic status diagnosis agent, and a spatial correlation analysis agent;

[0014] Step 7: Construct a selective triggering mechanism based on diagnostic priority. Dynamically call the large model multi-agent to perform diagnosis according to the degree of state change in different traffic areas, data reliability, and analysis needs, and obtain traffic state diagnosis results.

[0015] Step 8: Generate a structured traffic condition diagnosis report based on the traffic condition diagnosis results, and apply the diagnosis results to traffic operation monitoring, condition assessment, and intelligent traffic management systems.

[0016] Preferably, step 2 includes:

[0017] for Traffic analysis areas and Each discrete time step organizes the regional traffic state data into a regional traffic state matrix. ,in, This indicates the number of time steps in the traffic analysis. Indicates the number of traffic analysis areas. Indicates the first The time step, the first Traffic status values ​​corresponding to each traffic analysis area;

[0018] According to the regional traffic status matrix To determine whether each location has valid motion sensing observation information, construct the corresponding observation mask matrix. ,in, Indicates the first The time step, the first Each region has effective mobile sensing observation data. This indicates that there were no effective observations in the area during the corresponding time period;

[0019] The first is calculated based on the observation mask matrix. Overall regional observation coverage within each time step , used to indicate the first The effective observation ratio of all traffic analysis areas within a time step is used as a time coverage feature.

[0020] For any region Based on historical time windows Calculate the historical observation coverage of the region , used to represent regions At the target time The proportion of effective observations within the previous historical window serves as a characteristic of regional historical coverage.

[0021] The length of the missing time within the historical window is calculated based on the observation mask matrix. , used to represent regions At the target time The number of time steps lacking effective observations within the previous historical window is used as a feature of continuous missing data.

[0022] Construct a regional spatial correlation matrix based on the spatial relationships between transportation areas. , of which elements The value can be:

[0023] .

[0024] Preferably, step 3 includes:

[0025] Set target area In time The previous history input window length is The corresponding historical traffic state input sequence is represented as: ;

[0026] The corresponding historical observation mask sequence is represented as follows: ;

[0027] The two types of information mentioned above are fused to form a mask-aware input. ,in, Used to indicate the target area at the prediction time Previous comprehensive historical status input;

[0028] Input the mask sensing data Input the traffic state estimation model to obtain the predicted traffic state of the target area at the next time step. ,in, Indicates the first Time step region Predicted traffic conditions This represents a parametric traffic state estimation model. Indicates model parameters.

[0029] Preferably, when the mask sensing input is... Before inputting the traffic state estimation model and obtaining the predicted traffic state of the target area at the next time step, the traffic state estimation model is trained. The training objective is:

[0030] ;

[0031] Among them, when When, it indicates that the current location has valid real observations and can participate in supervised model training; when When the time is specified, it indicates that the current location lacks real observations and is not included in the model error calculation.

[0032] Preferably, step 4 includes:

[0033] For any region In the predicted time The traffic conditions are defined as the basic evaluation feature set. ,in, This indicates the magnitude of the change in the predicted state relative to the most recent valid observation state;

[0034] Construct the initial state scoring function, expressed as:

[0035] ;

[0036] in, Indicates the area In time Initial state score, Represents the normalized mapping function for the corresponding feature; Indicates the corresponding feature weights;

[0037] Thresholds are divided according to preset states. Generate the initial traffic state category, represented as:

[0038] ;

[0039] in, Indicates a stable traffic situation. This indicates a state that requires further analysis. This indicates a region where a significant change in state exists.

[0040] Build data reliability metrics ,in, This indicates the reliability of the data used for traffic condition analysis in the target area. This represents a data reliability mapping function constructed based on historical observation coverage and the degree of continuous missing data.

[0041] Preferably, step 5 includes:

[0042] For any region and diagnosis time Define the corresponding structured diagnostic evidence. ,in, Indicates the traffic status characteristics of the target area; Represents information about spatial relationships within a neighborhood;

[0043] The traffic state characteristics Includes: predicting traffic conditions Recent effective observation status Traffic condition change range And historical traffic condition change sequences;

[0044] The neighborhood spatial association information is represented as follows: ,in, Indicates the neighboring area that has a spatial relationship with the target area;

[0045] Organize the data according to a preset format to form standardized traffic condition diagnosis cases. ,in, Indicates the target traffic area. Indicates the time of diagnosis.

[0046] Preferably, in step 6, the data quality assessment agent is used to analyze the reliability of the data in the current traffic condition diagnostic cases and determine whether the current traffic condition has sufficient observational support; the input of the data quality assessment agent is reliability evidence. The output is ,in, Indicates the data reliability level. Indicates the mode of missing observations. This describes the uncertainty caused by insufficient observation;

[0047] The regional traffic condition diagnostic agent is used to analyze the current traffic operation status of the target area; the input of the regional traffic condition diagnostic agent is state evidence. The output is ,in, Indicates the category of regional traffic status. Indicates the confidence level of the diagnosis. Explanation of the basis for state judgment;

[0048] The spatial correlation analysis agent is used to analyze the state correlation between the target region and its neighboring regions; the input of the spatial correlation analysis agent is... The output is ,in, Indicates the degree of spatial correlation. Indicates the related areas that may be affected. This indicates the basis for spatial correlation analysis;

[0049] The final diagnostic result obtained by fusing the outputs of multiple agents is represented as follows:

[0050] ;

[0051] in, This represents the fusion function of results from multiple agents;

[0052] The final structured traffic condition diagnosis result is represented as follows:

[0053] ;

[0054] in, This indicates diagnostic interpretation information.

[0055] Preferably, step 7 includes:

[0056] For any region and time The diagnostic priority score is defined as follows:

[0057] ;

[0058] in, Indicates the area In time Diagnostic priority scoring, Indicates the weight of the corresponding factor;

[0059] Based on the calculated diagnostic priority scores, all regions to be analyzed are sorted, as follows:

[0060] ;

[0061] in, This function represents the sorting of a set of regions according to their diagnostic priority scores from highest to lowest. Represents the sorted set of regions;

[0062] Based on the preset large model call quantity parameter Select the first sorted items Each region enters the large-scale model multi-agent diagnostic process, represented as:

[0063] ;

[0064] in, This represents the set of regions for which a large model multi-agent analysis needs to be invoked. This indicates the number of regions that can be analyzed using the large model within each time period;

[0065] For belonging to The region is input into a large model multi-agent for traffic state diagnosis, generating structured diagnostic results that include state categories, reliability evaluations, and explanatory information;

[0066] For those who have not entered The region is represented by a low-complexity state evaluation method as follows:

[0067] ;

[0068] Employ a multi-level triggering strategy: when When the target area is identified as a high-priority area, a large-scale multi-agent model is invoked for in-depth analysis; when At that time, the decision to call a model is made based on whether the current number of large model calls has reached the preset call limit; when In this case, only the rule-based state evaluation results are used; among them, Indicates the high-priority trigger threshold; This indicates a low-priority trigger threshold;

[0069] Alternatively, a non-learning-based ranking strategy can be adopted, which prioritizes data based on initial state scores, the magnitude of changes in traffic conditions, the degree of continuity of missing data, and the extent of observation coverage.

[0070] Preferably, step 8 includes:

[0071] Using data reliability evaluation results as an important component of traffic condition diagnosis results, when data reliability indicators... Greater than or equal to the preset data reliability threshold When the data reliability index is [value missing], the reliability level of the diagnostic data for the target area is determined to be reliable; when the data reliability index is [value missing], the reliability level of the diagnostic data for the target area is determined to be reliable. Less than the preset data reliability threshold At that time, the reliability level of the diagnostic data in the target area is determined to be low reliability;

[0072] Based on the output of the spatial correlation analysis agent, record the state influence relationship between the target region and the associated region;

[0073] Diagnostic reports are stored using a standardized data structure, and the structured report is defined as follows:

[0074] ;

[0075] in, Indicates the diagnostic region number. Indicates the time of diagnosis; Indicates the target region in time Structured traffic condition diagnostic results, including at least traffic condition categories. Diagnostic confidence Data reliability level Spatial association information and diagnostic interpretation information .

[0076] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:

[0077] 1. This invention can distinguish between real traffic state changes and uncertainties caused by insufficient observation, thereby improving the reliability of traffic state analysis under sparse motion sensing conditions.

[0078] 2. This invention avoids the problems of information redundancy and analysis instability caused by directly inputting complex raw traffic data into large language models by constructing a structured traffic diagnostic evidence set. Through the structured evidence expression method, the subsequent diagnostic process has clear data sources and analytical basis, improving the traceability and interpretability of the traffic condition diagnosis process.

[0079] 3. This invention divides the traffic condition diagnosis task into three different functional modules: data quality assessment, regional traffic condition analysis, and spatial correlation analysis. This allows different agents to analyze different types of traffic evidence. This invention enables the collaborative processing of data reliability judgment, regional condition identification, and spatial correlation analysis, reducing the uncertainty of diagnosis results in complex traffic environments and improving the comprehensiveness of traffic condition analysis.

[0080] 4. By constructing a selective triggering mechanism based on diagnostic priority, this invention can reduce the number of unnecessary large language model calls, reduce computational resource consumption, and improve the real-time processing capability of the traffic condition diagnosis system.

[0081] 5. This invention establishes a degradation processing mechanism for low-reliability diagnostic results by combining diagnostic results with data reliability evaluation. When there is insufficient observation, insufficient historical state support, or high diagnostic uncertainty in a traffic area, the priority of using this result in subsequent traffic management decisions can be reduced, avoiding the direct impact of low-reliability state judgments on traffic management and guidance decisions, and improving the operational safety of the intelligent transportation system. Attached Figure Description

[0082] Figure 1 This is a flowchart of the present invention;

[0083] Figure 2 A schematic diagram illustrating the construction of structured diagnostic evidence for mask-sensing traffic state estimation.

[0084] Figure 3 A schematic diagram of a multi-agent collaborative diagnostic structure for a large language model;

[0085] Figure 4 Flowchart of selective triggering and structured diagnostic output based on diagnostic priority. Detailed Implementation

[0086] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the embodiments of the present invention, and not all structures.

[0087] In the following description, specific details such as target system architecture and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0088] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0089] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0090] Furthermore, in the description of this application and the appended claims, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0091] References to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include the target features, structures, or characteristics described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0092] The flowchart of the mask perception large-scale multi-agent traffic state diagnosis method described in this embodiment is as follows: Figure 1 As shown, the method includes the following steps:

[0093] S1, Obtain the basic data required for the sparse traffic condition diagnosis task, including mobile sensing traffic observation data, road network data, traffic area division data, and method operation parameter information.

[0094] In one example, mobile sensing traffic observation data refers to vehicle trajectory data collected by ride-hailing vehicles, taxis, floating cars, vehicle positioning terminals, or other mobile sensing devices, after preprocessing. Vehicle trajectory data includes at least vehicle number, sampling time, longitude coordinates, and latitude coordinates.

[0095] First, the acquired vehicle trajectory data undergoes data cleaning to remove data with abnormal positioning, time, and speed. Then, based on road network data, the vehicle trajectories are spatially matched to map the trajectory points to corresponding road locations. Finally, the road vehicle operating status is aggregated according to a preset time granularity to obtain road-level traffic condition observation data.

[0096] Road-level traffic condition observation data is obtained by processing mobile sensing traffic observation data through map matching and time aggregation. It is used to describe the traffic operation status of each road unit at different time steps.

[0097] In one implementation, the continuous traffic operation time is divided into multiple discrete time units, assuming the traffic observation time range includes... At each time step, the traffic network after road matching includes... If there are multiple road units, then the road-level traffic condition observation data is represented as follows: ,in, Indicates the number of time steps in traffic observation; Indicates the number of road units; Indicates the first The time step, the first Traffic condition observations corresponding to each road unit.

[0098] When a road unit has a valid mobile sensing vehicle observation at a certain time step, the traffic state value is recorded at the corresponding location; when the road unit has no valid observation at the corresponding time, the location is recorded as empty.

[0099] Furthermore, to reduce the instability caused by spatial sparsity in the analysis of traffic conditions at the single road level, this invention divides the road network into multiple traffic analysis areas based on road spatial connectivity, traffic operation similarity, or preset area division rules.

[0100] The traffic zone division data is used to describe the correspondence between road units and traffic analysis areas, including road numbers, zone numbers, and spatial relationship information.

[0101] Suppose the road network is divided into If there is a traffic analysis area, then the regional traffic status data is represented as follows: ,in, Indicates the number of traffic analysis areas. Indicates the first The time step, the first Traffic status values ​​corresponding to each traffic analysis area.

[0102] The method of dividing the area can be completed based on road topology, spatial clustering methods, traffic state similarity analysis methods, or manually set rules. This example does not limit this method.

[0103] Furthermore, the method operating parameter information used for the traffic condition diagnosis process is obtained, including:

[0104] (1) Historical state analysis window length, used to determine the historical time range for participating in traffic state estimation;

[0105] (2) Predict the time step, which is used to determine the time location of the target traffic state to be analyzed;

[0106] (3) Spatial neighborhood range, used to determine the range of areas around the target area that participate in spatial correlation analysis;

[0107] (4) State level classification rules, used to generate initial state categories;

[0108] (5) Large model call strategy parameters are used to control the subsequent selective triggering process.

[0109] Through the above steps, this example obtains basic data for sparse traffic condition diagnosis and converts the original mobile sensing trajectory data into traffic condition data with temporal continuity and spatial regional attributes, providing a data foundation for the subsequent construction of regional traffic condition matrix and observation mask matrix.

[0110] S2. Based on the acquired regional traffic status data, construct the regional traffic status matrix and its corresponding observation mask matrix to obtain time coverage features, regional historical coverage features, continuous missing features, and spatial correlation matrix.

[0111] Further, step S2 includes:

[0112] for Traffic analysis areas and Each discrete time step organizes the regional traffic state data into a regional traffic state matrix. ,in, This indicates the number of time steps in the traffic analysis. Indicates the number of traffic analysis areas. Indicates the first The time step, the first Traffic status values ​​corresponding to each traffic analysis area;

[0113] The traffic state value can be the regional average speed, regional average travel time, regional congestion index, or other indicators that can characterize the traffic operation status. In a preferred embodiment, the regional average speed is used as the traffic state representation.

[0114] Because mobile sensing data is affected by the number of vehicles, spatial distribution, and temporal variations, some traffic areas may lack effective observations at certain time steps. Therefore, this example further uses the regional traffic state matrix... To determine whether each location has valid motion sensing observation information, construct the corresponding observation mask matrix.

[0115] According to the regional traffic status matrix To determine whether each location has valid motion sensing observation information, construct the corresponding observation mask matrix. ,in, Indicates the first The time step, the first Each region has effective mobile sensing observation data. This indicates that there were no effective observations in the area during the corresponding time period;

[0116] Unlike traditional traffic data processing methods that only interpolate to fill in missing locations, this example preserves the original missing observation state and uses the observation mask matrix. As a crucial input in the traffic condition diagnosis process, the observation mask matrix is ​​used not only to distinguish between valid and missing observation locations, but also to describe the level of data support in the traffic condition estimation process.

[0117] The first is calculated based on the observation mask matrix. Overall regional observation coverage within each time step , used to indicate the first The effective observation ratio of all traffic analysis areas within a given time step, used as a time coverage feature, is calculated using the following mathematical expression:

[0118] ;

[0119] in, Indicates the first The effective observation ratio of all traffic analysis areas within a given time step.

[0120] For any region Based on historical time windows Calculate the historical observation coverage of the region , used to represent regions At the target time The proportion of effective observations within the previous historical window, as a characteristic of regional historical coverage, is calculated using the following expression:

[0121] ;

[0122] The coverage of historical observations in a region is used to describe whether the target area has continuous and effective traffic condition observation support in the near future.

[0123] The length of the missing time within the historical window is calculated based on the observation mask matrix. , used to represent regions At the target time The number of time steps lacking valid observations within the previous historical window, as a feature of continuous missing data,

[0124] Construct a regional spatial correlation matrix based on the spatial relationships between transportation areas. , of which elements The value can be:

[0125] .

[0126] Spatial correlation conditions can be determined based on road topology connectivity, regional adjacency, spatial distance thresholds, or traffic operation status correlations. For example:

[0127] (1) There is a direct road connection between the two areas;

[0128] (2) The spatial distance between the two regions is less than a preset threshold;

[0129] (3) The changes in historical traffic conditions in the two regions are related.

[0130] Based on the regional spatial correlation matrix, the observation coverage of the associated regions surrounding the target region can be further calculated for subsequent spatial correlation analysis.

[0131] Through the above steps, the present invention completes the regional traffic state matrix. and observation mask matrix The construction of the time coverage features was further obtained. Regional historical coverage characteristics Continuous missing features and spatial correlation matrix This provides a data foundation for subsequent mask-sensing traffic state estimation and large-scale multi-agent traffic state diagnosis.

[0132] Compared to methods that rely solely on traffic state numerical values ​​for analysis, this example explicitly preserves observation validity information, enabling the traffic state diagnosis process to simultaneously consider traffic state changes and the reliability of traffic data, thereby improving the reliability of state analysis results in sparse traffic environments.

[0133] S3 estimates the regional traffic status for the target time period under sparse traffic observation conditions, and obtains the predicted traffic status of the target area at the next time step.

[0134] Existing traffic condition prediction methods typically only use historical traffic condition sequences as model input, that is, they use historical traffic condition change patterns to predict future traffic conditions. However, under conditions of sparse mobile sensing data, there are many missing positions in the historical traffic condition sequences. If the missing data-filled traffic sequence is directly used as input, the model cannot distinguish between the actual traffic condition changes and the estimation bias caused by data gaps.

[0135] Therefore, in this example, historical traffic state information and corresponding observation mask information are input into the estimation model during the traffic state estimation process, so that the model can simultaneously perceive the traffic state change pattern and the effectiveness of historical observations.

[0136] Combination Figure 2 As shown, step S3 further includes:

[0137] Set target area In time The previous history input window length is The corresponding historical traffic state input sequence is represented as: ;

[0138] The corresponding historical observation mask sequence is represented as follows: ;

[0139] The two types of information mentioned above are fused to form a mask-aware input. ,in, Used to indicate the target area at the prediction time Previous comprehensive historical status input;

[0140] Mask-aware input Input the traffic state estimation model to obtain the predicted traffic state of the target area at the next time step. ,in, Indicates the first Time step region Predicted traffic conditions This represents a parametric traffic state estimation model. Indicates model parameters.

[0141] In one example, the traffic state estimation model can be implemented using a recurrent neural network model, a gated recurrent neural network model, a spatiotemporal neural network model, or other models with time series modeling capabilities.

[0142] Furthermore, to avoid incorrect supervision of model training due to missing observation locations, this example adopts a loss calculation method based on observation mask constraints, which only uses locations with valid observation labels to calculate the model training error.

[0143] Furthermore, when inputting mask sensing... Before inputting the traffic state estimation model and obtaining the predicted traffic state of the target area at the next time step, the traffic state estimation model is trained. The training objective is:

[0144] ;

[0145] Among them, when When, it indicates that the current location has valid real observations and can participate in supervised model training; when When the time is specified, it indicates that the current location lacks real observations and is not included in the model error calculation.

[0146] S4 generates rule-based initial traffic state evaluation results based on predicted traffic conditions, traffic state change characteristics, and observation reliability information.

[0147] Further, step S4 includes:

[0148] For any region In the predicted time The traffic conditions are defined as the basic evaluation feature set. ,in, This indicates the magnitude of the change in the predicted state relative to the most recent valid observation state;

[0149] Construct the initial state scoring function, expressed as:

[0150] ;

[0151] in, Indicates the area In time Initial state score, Represents the normalized mapping function for the corresponding feature; Indicates the corresponding feature weights;

[0152] Thresholds are divided according to preset states. Generate the initial traffic state category, represented as:

[0153] ;

[0154] in, Indicates a stable traffic situation. This indicates a state that requires further analysis. This indicates a region where a significant change in state exists.

[0155] Build data reliability metrics ,in, This indicates the reliability of the data used for traffic condition analysis in the target area. This represents a data reliability mapping function constructed based on historical observation coverage and the degree of continuous missing data.

[0156] Data reliability metrics As a quantitative basis for the reliability of data in subsequent traffic diagnosis processes, it together with predicted traffic conditions, the magnitude of changes in conditions, and spatial correlation information to form structured traffic diagnosis evidence. On the other hand, it serves as input to the data quality assessment agent and is used to calculate the diagnostic priority during selective triggering and to determine the reliability of structured diagnostic results.

[0157] Data reliability metrics can be used to characterize the degree of data support for traffic state analysis in target areas during subsequent structured diagnostic evidence construction, multi-agent diagnosis using large language models, and selective triggering processes. In one example, it is defined as: ,in, Indicates the first Target area at each time step Data reliability metrics; This indicates the observation coverage of the target area within the historical input window; This indicates the number of consecutive missing time steps within the historical input window; Indicates the length of the history input window; Weight parameters representing coverage and consecutive missing values, and .

[0158] Since 0≤ ≤1 and 0≤ ≤1, therefore .when The larger the value, the more sufficient the historical traffic observation support for the target area; when The smaller the value, the less support there is from historical traffic observations of the target area.

[0159] S5, Construct a structured evidence set for traffic condition diagnosis.

[0160] Based on the mask-sensing traffic state estimation results and the initial traffic state evaluation results, this example further constructs a structured traffic state diagnostic evidence set, which is used to convert sparse traffic observation information, traffic state change information, data reliability information and spatial correlation information into standardized inputs suitable for large-scale multi-agent analysis.

[0161] Existing large-model-based traffic analysis methods typically input raw traffic data, textual descriptions, or large amounts of statistical information directly into the model, allowing it to interpret traffic conditions. However, raw traffic data often contains a large amount of irrelevant information, and the data structures of different sources vary, which can easily lead to the model's inability to accurately extract key traffic state information. Furthermore, in sparse traffic perception environments, missing observations are inherently significant; directly deleting missing information during data transformation will prevent the effective analysis of traffic state reliability.

[0162] Therefore, in this example, instead of directly inputting the original vehicle trajectory data into the large model, a structured traffic condition diagnosis evidence set for traffic condition diagnosis tasks is constructed based on traffic condition estimation and initial state evaluation.

[0163] Further, step S5 includes:

[0164] For any region and diagnosis time Define the corresponding structured diagnostic evidence. ,in, Indicates the traffic status characteristics of the target area; Represents information about spatial relationships within a neighborhood;

[0165] Traffic condition characteristics Includes: predicting traffic conditions Used to describe the level of traffic operation in a future time period; the most recently observed state. This is used to represent the recent actual observation situation in the target area; the magnitude of changes in traffic conditions. The traffic condition change trend is described by the historical traffic condition change sequence, which is used to describe the recent operating mode of the target area.

[0166] Data reliability characteristics Including: Current time coverage Used to describe overall observational coverage; regional historical coverage. Used to describe the historical observational support level of the target area; continuous missing time Used to describe the degree of missing historical observations; historical observation mask sequence , used to preserve observation missing patterns.

[0167] Initial traffic condition assessment results Includes: initial state category; state score This is used to indicate the degree of change in traffic conditions; it also provides information on data reliability.

[0168] Neighborhood spatial association information is represented as This is used to describe changes in traffic conditions in other areas that are spatially related to the target area; among which, The target area is defined as the neighboring area that has a spatial relationship with the target area. The set of related regions is .

[0169] Organize the data according to a preset format to form standardized traffic condition diagnosis cases. ,in, Indicates the target traffic area. Indicates the time of diagnosis.

[0170] Traffic condition diagnosis cases serve as a unified input for subsequent large-scale model agents. Different agents do not need to re-parse the original traffic data; they can complete their corresponding tasks based solely on diagnostic evidence in a unified format, thereby reducing the complexity of the input data and improving the stability of the traffic condition diagnosis process.

[0171] Through the steps described above, this example establishes a data conversion mechanism from mask-sensing traffic state estimation results to large-scale multi-agent diagnostic inputs. Compared to existing methods that directly call large models to analyze traffic information, this invention explicitly expresses traffic state, data reliability, and spatial correlations through structured evidence construction, enabling subsequent agents to conduct traffic state analysis based on explicit evidence, thereby improving the interpretability and traceability of the diagnostic process.

[0172] Combination Figure 3As shown, in step S6, multiple large language model agents are constructed; the large language model multi-agent includes at least a data quality assessment agent, a regional traffic status diagnosis agent, and a spatial correlation analysis agent.

[0173] Existing traffic analysis methods based on large language models typically employ a single-agent approach, where traffic state information is input into a single large language model, which directly generates the analysis results. However, in sparse traffic perception scenarios, traffic state diagnosis requires not only judging changes in traffic operation status but also considering the reliability of observation data, regional state change trends, and spatial correlations. A single agent must simultaneously complete multiple analysis tasks of different natures, which can easily lead to confusion in analytical criteria, untraceable reasoning processes, and insufficient stability of results.

[0174] Therefore, in this example, the traffic condition diagnosis task is functionally divided, multiple large language model agents with different analysis responsibilities are constructed, and collaborative analysis is carried out through a structured traffic condition diagnosis evidence set.

[0175] Furthermore, in step S6, the data quality assessment agent analyzes the reliability of the data in the current traffic condition diagnostic case and determines whether the current traffic condition has sufficient observational support; the input to the data quality assessment agent is reliability evidence. The output is ,in, Indicates the data reliability level. Indicates the mode of missing observations. This describes the uncertainty caused by insufficient observation;

[0176] The regional traffic condition diagnostic agent is used to analyze the current traffic operation status of the target area; the input of the regional traffic condition diagnostic agent is state evidence. The output is ,in, Indicates the category of regional traffic status. Indicates the confidence level of the diagnosis. Explanation of the basis for state judgment;

[0177] The spatial association analysis agent is used to analyze the state associations between the target region and its neighboring regions; the input of the spatial association analysis agent is... The output is ,in, Indicates the degree of spatial correlation. Indicates the related areas that may be affected. This indicates the basis for spatial correlation analysis;

[0178] The final diagnostic result obtained by fusing the outputs of multiple agents is represented as follows:

[0179] ;

[0180] in, This represents the fusion function of results from multiple agents;

[0181] The final structured traffic condition diagnosis result is represented as follows:

[0182] ;

[0183] in, This indicates diagnostic interpretation information.

[0184] In one implementation, the multi-agent result fusion process integrates the outputs of the data quality assessment agent, the regional traffic condition diagnosis agent, and the spatial correlation analysis agent according to a preset field correspondence. Specifically, the outputs of the regional traffic condition diagnosis agent... The traffic state category of the target area is used to evaluate the agent's output based on data quality. As a measure of data reliability for the target region, the output of the spatial correlation analysis agent... and As spatial correlation information, the explanations of the causes of each agent's output, the descriptions of the missing observation patterns, and the basis for spatial correlation analysis are combined to form diagnostic and explanatory information.

[0185] For the diagnostic confidence score, in one implementation, the final diagnostic confidence score is obtained by weighting the confidence scores output by the three agents respectively:

[0186] ;

[0187] in, This represents the confidence level of the data quality assessment agent's output. This indicates the confidence level of the output of the regional traffic condition diagnosis agent. This represents the confidence level of the spatial correlation analysis agent's output. , , Let represent the confidence fusion weights for the three agents, respectively, and satisfy the following:

[0188] ;

[0189] Based on this, the final structured traffic state diagnosis result is expressed as follows:

[0190] ;

[0191] Furthermore, diagnostic interpretation information This is generated based on structured traffic diagnostic evidence and the output results of each agent. In one implementation, diagnostic interpretation information is generated according to preset interpretation field combination rules, based on the state category and cause description output by the regional traffic state diagnostic agent, the data reliability level and observation missing pattern description output by the data quality assessment agent, and the spatial correlation state and analysis basis output by the spatial correlation analysis agent.

[0192] When the regional traffic state diagnosis agent determines that there is a significant change in the state of the target area, and the data quality assessment agent determines that the data reliability level meets the preset reliability requirements, the diagnostic explanation information shall at least include the traffic state change and its data support. When the target area has a low observation coverage or a long continuous period of missing data, the diagnostic explanation information shall add a description of insufficient observation support or diagnostic uncertainty. When the spatial correlation analysis agent determines that there is a consistent traffic state change between the target area and one or more neighboring areas, the diagnostic explanation information shall further include the basis for the spatial correlation of the neighboring areas.

[0193] Therefore, the final structured traffic condition diagnosis result is not an independently generated text result, but is formed by combining traffic condition characteristics, data reliability characteristics, spatial correlation information and the diagnosis results of each intelligent agent, and is used to explain the main sources of evidence corresponding to the final traffic condition diagnosis result.

[0194] Through the above steps, a traffic state diagnosis method based on a large language model and multiple agents was established. Compared with the method of a single agent directly generating analysis results, this example separates the tasks of data reliability analysis, regional state judgment and spatial correlation analysis, so that different agents can carry out analysis around a clear goal, which improves the interpretability, stability and engineering application capability of the traffic state diagnosis process in sparse traffic environment.

[0195] S7 constructs a selective triggering mechanism based on diagnostic priority, dynamically calling large model multi-agents to perform diagnosis according to the degree of state change in different traffic areas, data reliability, and analysis needs, and obtaining traffic state diagnosis results.

[0196] Based on the initial traffic state evaluation results and traffic state change characteristics, this step further constructs a selective triggering mechanism based on diagnostic priority. This mechanism is used to dynamically determine whether to call the large model multi-agent for further diagnosis based on the degree of state change in different traffic areas, data reliability, and analysis needs.

[0197] Existing intelligent traffic analysis methods typically assume that all monitored objects undergo the same processing flow, meaning that complex model calculations are performed on all regions and all time windows. However, in actual urban traffic monitoring, traffic conditions vary significantly across different regions, with many areas operating in a stable state and not requiring complex inference analysis. Calling large models on all regions would lead to wasted computational resources, increased response time, and higher system operating costs.

[0198] Therefore, this example proposes a selective triggering mechanism based on diagnostic priority. By comprehensively evaluating the degree of state change, observation reliability, and analysis needs of different regions, only regions with high diagnostic value are selected to enter the large model multi-agent traffic state diagnosis process.

[0199] Combination Figure 4 As shown, step S7 further includes:

[0200] For any region and time The diagnostic priority score is defined as follows:

[0201] ;

[0202] in, Indicates the area In time Diagnostic priority scoring, Indicates the weight of the corresponding factor;

[0203] A higher initial state score indicates a more significant change in the state of the target area; a greater change in traffic status indicates a more significant change in operational status; a longer period of continuous missing data indicates insufficient data support for the current area, requiring further analysis; and a lower historical coverage indicates that the current analysis results require more intelligent reasoning assistance.

[0204] Based on the calculated diagnostic priority scores, all regions to be analyzed are sorted, as follows:

[0205] ;

[0206] in, This function represents the sorting of a set of regions according to their diagnostic priority scores from highest to lowest. Represents the sorted set of regions;

[0207] Based on the preset large model call quantity parameter Select the first sorted items Each region enters the large-scale model multi-agent diagnostic process, represented as:

[0208] ;

[0209] in, This represents the set of regions for which a large model multi-agent analysis needs to be invoked. This indicates the number of regions that can be analyzed using the large model within each time period;

[0210] For belonging to The region is input into a large model multi-agent for traffic state diagnosis, generating structured diagnostic results that include state categories, reliability evaluations, and explanatory information;

[0211] For those who have not entered The region is represented by a low-complexity state evaluation method as follows:

[0212] ;

[0213] Employ a multi-level triggering strategy: when When the target area is identified as a high-priority area, a large-scale multi-agent model is invoked for in-depth analysis; when At that time, the decision to call a model is made based on whether the current number of large model calls has reached the preset call limit; when In this case, only the rule-based state evaluation results are used; among them, Indicates the high-priority trigger threshold; This indicates a low-priority trigger threshold;

[0214] Alternatively, a non-learning-based ranking strategy can be adopted, which prioritizes data based on initial state scores, the magnitude of changes in traffic conditions, the degree of continuity of missing data, and the extent of observation coverage.

[0215] In the multi-level triggering strategy, the region to be analyzed is divided into high-priority, medium-priority, and low-priority regions based on the comparison between the diagnostic priority score and preset high and low priority thresholds. High-priority regions directly invoke the large language model multi-agent for diagnosis; medium-priority regions determine whether to invoke it based on the available computing resources of the current system, the current number of large language model calls, and a preset resource occupancy threshold; low-priority regions use rule-based state evaluation results. For example, when the high-priority triggering threshold is 0.8 and the low-priority triggering threshold is 0.5, regions with a diagnostic priority score of 0.8 or higher directly invoke the large language model multi-agent; regions with a score of 0.5 or higher but lower than 0.8 are invoked only when the current number of large language model calls has not reached a preset upper limit; and regions with a score lower than 0.5 use rule-based state evaluation results. This allows for hierarchical scheduling of large language model calls based on diagnostic needs and computing resources.

[0216] Furthermore, to ensure the stability of the selective triggering process, a non-learning sorting strategy can be adopted, which prioritizes traffic based on: initial state score; magnitude of traffic state change; degree of continuous missing data; and observation coverage.

[0217] The ranking strategy requires no additional training data and can be directly applied to different urban areas and different mobile sensing data conditions, thus improving the method's adaptability.

[0218] Through the above steps, a selective intelligent analysis mechanism based on diagnostic priorities is established. Compared with the traditional method of uniformly performing complex calculations on all areas, this invention can dynamically allocate computing resources according to the degree of traffic state change and data reliability, and only call large-scale multi-agent models for areas with high analysis needs. This reduces computational overhead and improves the real-time operation capability of the system while ensuring the effectiveness of traffic state diagnosis.

[0219] S8 generates a structured traffic condition diagnosis report based on the traffic condition diagnosis results and applies the diagnosis results to traffic operation monitoring, condition assessment, and intelligent traffic management systems.

[0220] Existing traffic state analysis methods typically only output a single state category or prediction result, lacking descriptions of the basis for state judgment, data reliability, and spatial correlation. When traffic monitoring data exhibits significant sparsity, simply outputting a traffic state category cannot demonstrate whether the result is adequately supported by observations, nor can it assist managers in assessing the reliability of the analysis results.

[0221] Therefore, this example integrates multi-agent diagnostic results, traffic state estimation results, and data reliability evaluation results to form a structured traffic state diagnostic report for traffic management applications.

[0222] Further, step S8 includes:

[0223] Using data reliability evaluation results as an important component of traffic condition diagnosis results, when data reliability indicators... Greater than or equal to the preset data reliability threshold When the data reliability index is [value missing], the reliability level of the diagnostic data for the target area is determined to be reliable; when the data reliability index is [value missing], the reliability level of the diagnostic data for the target area is determined to be reliable. Less than the preset data reliability threshold At that time, the reliability level of the diagnostic data in the target area is determined to be low reliability;

[0224] Based on the output of the spatial correlation analysis agent, record the state influence relationship between the target region and the associated region;

[0225] Diagnostic reports are stored using a standardized data structure, and the structured report is defined as follows:

[0226] ;

[0227] in, Indicates the diagnostic region number. Indicates the time of diagnosis; Indicates the target region in time Structured traffic condition diagnostic results, including at least traffic condition categories. Diagnostic confidence Data reliability level Spatial association information and diagnostic interpretation information .

[0228] Furthermore, the diagnostic report can be output to traffic operation monitoring platforms, traffic condition assessment platforms, traffic guidance systems, or intelligent traffic management systems.

[0229] High-reliability anomaly diagnosis results can serve as an important reference for traffic management systems, assisting in traffic operation monitoring, supporting traffic guidance information dissemination, and helping traffic managers formulate control strategies. For low-reliability diagnosis results, a degradation processing mechanism is provided. In such cases, the following actions can be taken: reduce the priority of diagnostic result display, prohibit its direct use as a basis for automatic control, wait for subsequent observation data to be supplemented before re-analyzing, and revert to using historical statistical status as reference information. This indicates the data reliability threshold.

[0230] Furthermore, statistical analysis can be performed on long-term low-coverage areas based on historical diagnostic results. For areas that have long suffered from low observation coverage, high uncertainty, and insufficient data support, perception optimization suggestions can be generated, such as increasing the deployment of fixed traffic detection equipment, optimizing the scheduling of mobile sensing equipment, and adjusting traffic monitoring strategies.

[0231] Through the above steps, a complete technical process is completed, from sparse traffic observation data acquisition, masked perception state estimation, large-scale multi-agent traffic state diagnosis, selective intelligent invocation to structured result output.

[0232] Compared with traditional methods that only output traffic state predictions or state categories, this invention can not only determine the changes in traffic area state, but also simultaneously explain the reliability of the data corresponding to the state determination, spatial correlation, and diagnostic basis, making the traffic state analysis results traceable, interpretable, and valuable for engineering applications.

[0233] This invention explicitly represents the missing observation states in mobile sensing data by constructing a regional traffic state matrix and an observation mask matrix. The observation mask serves as crucial input information in the traffic state diagnosis process, enabling the system to simultaneously consider changes in traffic operation status and the reliability of data observations. Compared to traditional methods that analyze solely based on traffic state values, this invention can distinguish between actual traffic state changes and uncertainties caused by insufficient observations, thus improving the reliability of traffic state analysis under sparse mobile sensing conditions.

[0234] This invention inputs historical traffic state information and corresponding observation mask information into the state estimation model, enabling the model to simultaneously learn the correlation between traffic operation change patterns and missing observation patterns. Furthermore, by employing a loss calculation method based on observation mask constraints, invalid observation locations are avoided from participating in model training, improving the accuracy and stability of state estimation results under sparse traffic conditions and providing data-supported traffic state information for subsequent intelligent diagnosis.

[0235] This invention constructs a structured traffic diagnostic evidence set, which uniformly organizes predicted traffic conditions, traffic condition change characteristics, observation coverage, missing patterns, spatial neighborhood information, and initial state evaluation results. This avoids the information redundancy and analytical instability problems caused by directly inputting complex raw traffic data into large language models. The structured evidence presentation method provides subsequent diagnostic processes with clear data sources and analytical basis, improving the traceability and interpretability of the traffic condition diagnostic process.

[0236] This invention constructs a multi-agent collaborative diagnostic method based on a large language model. It divides the traffic condition diagnosis task into three functional modules: data quality assessment, regional traffic condition analysis, and spatial correlation analysis. This allows different agents to analyze different types of traffic evidence. Compared to methods where a single agent directly generates diagnostic results, this invention enables collaborative processing of data reliability judgment, regional condition identification, and spatial correlation analysis, reducing the uncertainty of diagnostic results in complex traffic environments and improving the comprehensiveness of traffic condition analysis.

[0237] This invention employs a standardized output structure to represent diagnostic results as a structured output that includes state category, diagnostic confidence level, data reliability level, spatial correlation, and explanatory information. This ensures that traffic state diagnostic results not only contain state judgments but also provide corresponding data support and explanatory information. The structured output facilitates data interaction with traffic operation monitoring platforms, traffic state assessment systems, and intelligent traffic management systems, enhancing the method's engineering application capabilities.

[0238] This invention constructs a selective triggering mechanism based on diagnostic priority. It determines the diagnostic priority of different regions according to the degree of traffic state change, initial state evaluation results, observation gaps, and data coverage. Only regions with high analytical needs are further analyzed using a large language model multi-agent system. Compared to methods that uniformly perform large model inference on all regions, this invention reduces unnecessary large language model calls, lowers computational resource consumption, and improves the real-time processing capability of the traffic state diagnostic system.

[0239] This invention establishes a degradation processing mechanism for low-reliability diagnostic results by combining diagnostic results with data reliability evaluation. When there is insufficient observation, insufficient historical state support, or high diagnostic uncertainty in a traffic area, the priority of using this result in subsequent traffic management decisions can be reduced, avoiding the direct impact of low-reliability state judgments on traffic management and guidance decisions, and improving the operational safety of the intelligent transportation system.

[0240] This invention forms a complete technical process from sparse motion sensing data acquisition, observation mask construction, traffic state estimation, diagnostic evidence generation, multi-agent collaborative analysis of large language models, intelligent triggering and scheduling to structured state output. It can be applied to traffic state diagnosis tasks under different data coverage conditions and different urban traffic environments, and has good versatility, scalability and engineering application value.

Claims

1. A masked perception large model multi-agent traffic state diagnosis method, characterized in that, Includes the following steps: Step 1: Obtain the basic data required for the sparse traffic condition diagnosis task. The basic data includes mobile sensing traffic observation data, road network data, traffic area division data, and method operation parameter information. Step 2: Based on the acquired regional traffic status data, construct the regional traffic status matrix and its corresponding observation mask matrix to obtain the temporal coverage features, regional historical coverage features, continuous missing features, and spatial correlation matrix. Step 3: Estimate the regional traffic status for the target time period under sparse traffic observation conditions to obtain the predicted traffic status for the target area at the next time step. Step 4: Based on the predicted traffic conditions, traffic condition change characteristics, and observation reliability information, generate rule-based initial traffic condition evaluation results; Step 5: Construct a structured evidence set for traffic condition diagnosis; Step 6: Construct multiple large language model agents; the large language model agents include at least a data quality assessment agent, a regional traffic status diagnosis agent, and a spatial correlation analysis agent; Step 7: Construct a selective triggering mechanism based on diagnostic priority. Dynamically call the large model multi-agent to perform diagnosis according to the degree of state change in different traffic areas, data reliability, and analysis needs, and obtain traffic state diagnosis results. Step 8: Generate a structured traffic condition diagnosis report based on the traffic condition diagnosis results, and apply the diagnosis results to traffic operation monitoring, condition assessment, and intelligent traffic management systems.

2. The masked perception large model multi-agent traffic state diagnosis method according to claim 1, characterized in that, Step 2 includes: For traffic analysis zones and discrete time steps, zone-level traffic state data is organized into a zone traffic state matrix wherein, denotes the number of traffic analysis time steps, denotes the number of traffic analysis zones, denotes the traffic state value corresponding to the time step, the traffic analysis zone. According to the regional traffic status matrix To determine whether each location has valid motion sensing observation information, construct the corresponding observation mask matrix. ,in, Indicates the first The time step, the first Each region has effective mobile sensing observation data. This indicates that there were no effective observations in the area during the corresponding time period; The first is calculated based on the observation mask matrix. Overall regional observation coverage within each time step , used to indicate the first The effective observation ratio of all traffic analysis areas within a time step is used as a time coverage feature. For any region Based on historical time windows Calculate the historical observation coverage of the region , used to represent regions At the target time The proportion of effective observations within the previous historical window serves as a characteristic of regional historical coverage. The length of the missing time within the historical window is calculated based on the observation mask matrix. , used to represent regions At the target time The number of time steps lacking effective observations within the previous historical window is used as a feature of continuous missing data. Construct a regional spatial correlation matrix based on the spatial relationships between transportation areas. , of which elements The value can be: 。 3. The mask-sensing large-scale multi-agent traffic state diagnosis method according to claim 2, characterized in that, Step 3 includes: Set target area In time The previous history input window length is The corresponding historical traffic state input sequence is represented as: ; The corresponding historical observation mask sequence is represented as follows: ; The two types of information mentioned above are fused to form a mask-aware input. ,in, Used to indicate the target area at the prediction time Previous comprehensive historical status input; Input the mask sensing data Input the traffic state estimation model to obtain the predicted traffic state of the target area at the next time step. ,in, Indicates the first Time step region Predicted traffic conditions This represents a parametric traffic state estimation model. Indicates model parameters.

4. The mask-sensing large-scale multi-agent traffic state diagnosis method according to claim 3, characterized in that, When the mask sensing input is given Before inputting the traffic state estimation model and obtaining the predicted traffic state of the target area at the next time step, the traffic state estimation model is trained. The training objective is: ; Among them, when When, it indicates that the current location has valid real observations and can participate in supervised model training; when When the time is specified, it indicates that the current location lacks real observations and is not included in the model error calculation.

5. The mask-sensing large-scale multi-agent traffic state diagnosis method according to claim 4, characterized in that, Step 4 includes: For any region In the predicted time The traffic conditions are defined as the basic evaluation feature set. ,in, This indicates the magnitude of the change in the predicted state relative to the most recent valid observation state; Construct the initial state scoring function, expressed as: ; in, Indicates the area In time Initial state score, Represents the normalized mapping function for the corresponding feature; Indicates the corresponding feature weights; Thresholds are divided according to preset states. Generate the initial traffic state category, represented as: ; in, Indicates a stable traffic situation. This indicates a state that requires further analysis. This indicates a region where a significant change in state exists. Build data reliability metrics ,in, This indicates the reliability of the data used for traffic condition analysis in the target area. This represents a data reliability mapping function constructed based on historical observation coverage and the degree of continuous missing data.

6. The mask-sensing large-scale multi-agent traffic state diagnosis method according to claim 5, characterized in that, Step 5 includes: For any region and diagnosis time Define the corresponding structured diagnostic evidence. ,in, Indicates the traffic status characteristics of the target area; Represents information about spatial relationships within a neighborhood; The traffic state characteristics Includes: predicting traffic conditions Recent effective observation status Traffic condition change range And historical traffic condition change sequences; The neighborhood spatial association information is represented as follows: ,in, Indicates the neighboring area that has a spatial relationship with the target area; Organize the data according to a preset format to form standardized traffic condition diagnosis cases. ,in, Indicates the target traffic area. Indicates the time of diagnosis.

7. The mask-sensing large-scale multi-agent traffic state diagnosis method according to claim 6, characterized in that, In step 6, the data quality assessment agent is used to analyze the reliability of the data in the current traffic condition diagnostic cases and determine whether the current traffic condition has sufficient observational support; the input of the data quality assessment agent is reliability evidence. The output is ,in, Indicates the data reliability level. Indicates the mode of missing observations. This describes the uncertainty caused by insufficient observation; The regional traffic condition diagnostic agent is used to analyze the current traffic operation status of the target area; the input of the regional traffic condition diagnostic agent is state evidence. The output is ,in, Indicates the category of regional traffic status. Indicates the confidence level of the diagnosis. Explanation of the basis for state judgment; The spatial correlation analysis agent is used to analyze the state correlation between the target region and its neighboring regions; the input of the spatial correlation analysis agent is... The output is ,in, Indicates the degree of spatial correlation. Indicates the related areas that may be affected. This indicates the basis for spatial correlation analysis; The final diagnostic result obtained by fusing the outputs of multiple agents is represented as follows: ; in, This represents the fusion function of results from multiple agents; The final structured traffic condition diagnosis result is represented as follows: ; in, This indicates diagnostic interpretation information.

8. The mask-sensing large-scale multi-agent traffic state diagnosis method according to claim 6, characterized in that, Step 7 includes: For any region and time The diagnostic priority score is defined as follows: ; in, Indicates the area In time Diagnostic priority scoring, Indicates the weight of the corresponding factor; Based on the calculated diagnostic priority scores, all regions to be analyzed are sorted, as follows: ; in, This function represents the sorting of a set of regions according to their diagnostic priority scores from highest to lowest. Represents the sorted set of regions; Based on the preset large model call quantity parameter Select the first sorted items Each region enters the large-scale model multi-agent diagnostic process, represented as: ; in, This represents the set of regions for which a large model multi-agent analysis needs to be invoked. This indicates the number of regions that can be analyzed using the large model within each time period; For belonging to The region is input into a large model multi-agent for traffic state diagnosis, generating structured diagnostic results that include state categories, reliability evaluations, and explanatory information; For those who have not entered The region is represented by a low-complexity state evaluation method as follows: ; Employ a multi-level triggering strategy: when When the target area is identified as a high-priority area, a large-scale multi-agent model is invoked for in-depth analysis; when At that time, the decision to call a model is made based on whether the current number of large model calls has reached the preset call limit; when In this case, only the rule-based state evaluation results are used; among them, Indicates the high-priority trigger threshold; This indicates a low-priority trigger threshold; Alternatively, a non-learning-based ranking strategy can be adopted, which prioritizes data based on initial state scores, the magnitude of changes in traffic conditions, the degree of continuity of missing data, and the extent of observation coverage.

9. The mask-sensing large-scale multi-agent traffic state diagnosis method according to claim 8, characterized in that, Step 8 includes: Using data reliability evaluation results as an important component of traffic condition diagnosis results, when data reliability indicators... Greater than or equal to the preset data reliability threshold When the data reliability index is [value missing], the reliability level of the diagnostic data for the target area is determined to be reliable; when the data reliability index is [value missing], the reliability level of the diagnostic data for the target area is determined to be reliable. Less than the preset data reliability threshold At that time, the reliability level of the diagnostic data in the target area is determined to be low reliability; Based on the output of the spatial correlation analysis agent, record the state influence relationship between the target region and the associated region; Diagnostic reports are stored using a standardized data structure, and the structured report is defined as follows: ; in, Indicates the diagnostic region number. Indicates the time of diagnosis; Indicates the target region in time Structured traffic condition diagnostic results, including at least traffic condition categories. Diagnostic confidence Data reliability level Spatial association information and diagnostic interpretation information .