A traffic condition prediction method and prediction system based on a trusted federated distillation-dynamic road network graph learning

Abstract

The present application relates to the technical field of intelligent transportation, mobile crowd sensing, traffic data mining and artificial intelligence prediction, and particularly relates to a road condition prediction method and system based on trusted federated distillation-dynamic road network graph learning. The method uses mobile phone terminals to collect positioning and motion sensor data, calculates participation credibility and generates local soft labels, and uploads them in a privacy-preserving manner. The server performs trusted federated distillation, constructs a dynamic road network graph containing a participation sensing credible edge, and predicts future road conditions through a spatio-temporal graph learning model. Based on the prediction residual feedback, the credibility, edge weight and model parameters are updated to improve the accuracy of road condition prediction and privacy protection.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation, mobile crowd sensing, traffic data mining, and artificial intelligence prediction technology, and in particular to a traffic condition prediction method and system based on trusted federated distillation-dynamic road network graph learning. Background Technology

[0002] With the continuous growth of urban motor vehicle ownership, problems such as traffic congestion, uncertain travel times, and uneven utilization of road resources are becoming increasingly prominent. Real-time, accurate, and continuous acquisition of urban road traffic conditions, and prediction of short-term future road conditions, are crucial foundations for traffic guidance, congestion warnings, travel route optimization, and traffic management decisions. Existing intelligent transportation systems typically consist of traffic information collection, information processing and analysis, and information dissemination, with traffic information collection forming the data foundation of the entire system. Traditional traffic information collection methods mainly rely on fixed detection equipment such as geomagnetic coils, microwave radar, video cameras, and infrared detectors. While these methods offer high detection accuracy, they suffer from high construction costs, heavy maintenance workloads, limited coverage, and significant limitations imposed by installation location, making it difficult to achieve continuous coverage of urban secondary roads, suburban roads, temporarily congested sections, and non-key monitoring areas. The attached project also points out that real-time traffic condition acquisition and prediction are key to the Internet of Things (IoT) for transportation, but existing traffic condition collection methods suffer from high costs, small coverage, and insufficient scalability of multi-source heterogeneous prediction models.

[0003] To reduce the cost of traffic data collection, participatory traffic data collection methods based on mobile smart terminals or smartphones are gaining increasing attention. For example, Chinese patent CN102737510B discloses a real-time traffic condition collection method based on mobile smart terminals. This method utilizes mobile smart terminals to identify the traffic environment based on sensor information, combines road databases to track driving routes in real time, and uploads travel time and real-time traffic information to a server. The server analyzes, deduplicates, and stores the massive amount of real-time traffic information received, and uses compressed sensing technology to estimate real-time traffic conditions for road sections lacking uploaded data, thereby achieving real-time traffic condition collection for urban traffic networks. This solution can expand the range of traffic condition collection by utilizing a large number of mobile smart terminals and reduce reliance on fixed detection facilities. However, these methods primarily focus on the aggregation, deduplication, and missing segment estimation of data collected by mobile terminals. They lack a reliable control mechanism that spans edge sampling, server aggregation, and road network prediction models to address data quality differences caused by different mobile phone brands, sensor quality, placement orientations, road matching errors, and the credibility of data from different participants. Furthermore, these methods may still require uploading information related to the travel route or road segment journey, leaving room for further improvement in privacy protection, continuous learning of edge models, and generalization of road segment-level predictions.

[0004] For example, Chinese patent CN104408917A discloses a real-time traffic condition estimation method based on a smartphone's accelerometer sensor. This method collects acceleration information using the smartphone's built-in accelerometer sensor, processes the measured data using the phone's memory, estimates traffic condition information using an offline-trained model, and then uploads the traffic condition estimation results to a server to reduce data transmission burden and server computational pressure. This solution moves some computation to the mobile device, offering advantages in low power consumption and lightweight design. However, it relies heavily on offline-trained models for edge estimation, making it difficult to adapt to the constantly changing dynamic environment of urban road networks, including road structure, traffic times, driving behavior, and the participating terminal groups. Furthermore, mobile sensor data is significantly affected by factors such as the phone's placement angle, vehicle vibration, road slope, differences in terminal hardware, and user operation status. If the edge estimation results are used only as ordinary uploaded data, low-reliability samples can easily accumulate on the server side, affecting the stability of subsequent traffic condition prediction models.

[0005] In traffic prediction models, time series models, Kalman filter models, parametric regression models, neural network models, and graph neural network models have been used to predict traffic flow, speed, or congestion. In recent years, graph learning methods based on road network topology have been able to express the spatial relationships between roads, while deep learning methods based on time series encoding can express the temporal evolution of congestion. Combining these two approaches can improve short-term traffic prediction capabilities. However, existing graph learning-based traffic prediction methods primarily use fixed detectors, historical traffic records, or centralized traffic data as training objects. For data generated by mobile phone-based sensing, they typically do not adequately consider the attitude errors of participating terminals, road matching errors, consistency among multiple terminals on the same road segment, historical prediction contributions, and privacy-preserving upload constraints. In other words, even if existing methods can utilize road network topology and historical traffic patterns for prediction, they struggle to address the comprehensive problems in mobile phone-based sensing scenarios, such as "data from a large number of heterogeneous terminals, data quality changing over time, sparse samples in some road segments, and the unsuitability of uploading original trajectories."

[0006] Furthermore, privacy-preserving machine learning techniques such as federated learning and knowledge distillation have been used for collaborative modeling of multi-party data. Their basic idea is to complete model training or transfer through intermediate information such as model parameters, gradients, embeddings, or soft labels without centrally uploading the original data. However, in scenarios involving mobile phones in traffic prediction, simply using federated learning still has shortcomings: on the one hand, the quality of model information uploaded by different mobile terminals varies significantly. Without a credibility assessment related to traffic semantics, low-quality terminals may produce negative transfer during the aggregation stage. On the other hand, the server-side road network prediction model not only needs to determine the current state of each road segment but also needs to identify the congestion propagation relationship between adjacent road segments. Traditional federated learning typically does not directly participate in the generation of dynamic graph edge weights and edge lifecycle control. Therefore, how to coordinate the credibility of mobile data with federated distillation, dynamic graph construction of road segments, and short-term traffic condition prediction while protecting the original trajectory and original sensor time series remains a problem that existing technologies need to solve.

[0007] In summary, existing technologies have disclosed real-time traffic data collection based on mobile smart terminals, end-side traffic estimation based on mobile phone accelerometers, and traffic state prediction methods based on machine learning or graph learning. However, a reliable prediction mechanism for mobile phone-based traffic data is still lacking. This mechanism should be able to quantify the data quality of different participating terminals without uploading the original trajectory and original motion sensor time series, and simultaneously apply this reliability to end-side sample admission, federated distillation and aggregation, dynamic road network graph edge weight generation, and prediction residual feedback update. This would improve the privacy protection, adaptability to low-coverage road sections, and continuous learning capability of traffic prediction. Summary of the Invention

[0008] The purpose of this invention is to provide a reliable federated distillation-dynamic road network graph learning traffic condition prediction method and system for mobile phone-participatory traffic data. This method is used to perform reliability assessment, privacy-preserving aggregation, and dynamic graph traffic condition prediction on participant-sensing data from different mobile terminals without uploading the original trajectory and original motion sensor time series. It solves the problems of terminal heterogeneity, unstable attitude, large road matching error, insufficient low-coverage road segment samples, and difficulty in continuously adaptively updating the prediction model in existing mobile phone-participatory traffic systems, thereby improving the accuracy, coverage, and stability of short-term traffic condition prediction for urban roads.

[0009] Firstly, in order to achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0010] A road condition prediction method based on trusted federated distillation-dynamic road network graph learning, comprising the following steps:

[0011] S1. Multiple mobile phones participate as sensing terminals to collect positioning data and motion sensor data during vehicle operation, and trigger traffic sensing tasks based on vehicle motion state, mobile phone posture stability and remaining power consumption; perform vehicle coordinate system correction, fixed time window segmentation and road segment matching on the motion sensor data to obtain road segment-level end-side traffic features.

[0012] S2. For each mobile phone participating in the perception terminal within each time window, generate participation credibility based on the mobile phone attitude correction residual, road segment matching residual, terminal sensor noise level, consistency of soft tags of multiple terminals within the same road segment, and prediction residual feedback item. and according to Control whether the corresponding sample participates in local training, edge inference, and federated distillation upload;

[0013] S3. The mobile phone participating in the perception terminal generates local soft tags based on the local model, including congestion level probability, speed range probability, and abnormal perception probability, and only uploads the local soft tags, the dimensionality-reduced edge model intermediate embedding, the road segment anonymous identifier, the time window anonymous identifier, and the participation credibility. ;

[0014] S4, server according to Based on the scarcity of road segment coverage and the historical contribution of terminals, multiple local soft tags within the same road segment and the same time window are subjected to trusted federated distillation aggregation to obtain road segment-level distillation tags.

[0015] S5. The server constructs a dynamic road network graph including road topology edges, traffic pattern similarity edges, and trusted edges for participation perception. The trusted edges for participation perception are generated, enhanced, attenuated, or deleted based on the congestion level transfer delay, aggregated participation credibility, and prediction residual.

[0016] S6. Input the road segment-level distillation labels, dynamic road network map, historical congestion level sequence, road attribute features and time context features into the spatiotemporal graph learning model, and output the road condition prediction results for multiple future prediction time windows.

[0017] As a further improvement, in step S1, the positioning data includes at least one of GNSS positioning data, base station positioning data, and wireless network positioning data;

[0018] The motion sensor data includes at least one of acceleration data, angular velocity data, magnetometer data, barometer data, and terminal system motion status data.

[0019] The vehicle coordinate system correction includes: estimating the rotational relationship between the phone attitude coordinate system and the vehicle driving coordinate system based on the direction of gravity; converting the acceleration components in the phone coordinate system into vehicle longitudinal acceleration components, vehicle lateral acceleration components, and vehicle vertical disturbance components; and determining the phone attitude stability based on the change amplitude of the rotational relationship within adjacent time windows.

[0020] And / or, in step S2, when the mobile phone pose correction residual or road segment matching residual exceeds the corresponding threshold, the sample in that time window is prohibited from participating in local training; when the mobile phone pose correction residual and road segment matching residual do not exceed the corresponding threshold, but participate in the confidence level... When the sample size is below the confidence threshold, only samples within that time window are allowed to participate in edge inference but not in federated distillation uploads.

[0021] And / or, in step S2, participate in credibility Determine as follows:

[0022] In the formula, The credibility of the participation of the u-th mobile phone as a sensing terminal within time window t; Rate the phone's posture stability; Assess the quality of road segment matching; A consistency score for soft labels across multiple terminals within the same road segment; The terminal's historical contribution score; This is a penalty term for energy consumption constraints; This is the penalty term for the forecast residuals from the previous forecast period; , , , , , These are the weights of the corresponding terms; σ is the normalized mapping function;

[0023] Preferably, the road segment matching quality score The distance from the location point to the candidate road segment, the angle between the vehicle's direction of travel and the direction of the candidate road segment, the connectivity of road segments in adjacent time windows, and the deviation between the vehicle speed and the road speed limit range are all determined together.

[0024] When the distance from the location point to the candidate road segment is greater than a distance threshold, or the angle between the vehicle's driving direction and the candidate road segment's direction is greater than an angle threshold, the road segment matching quality score is reduced. And mark the samples in this time window as low-confidence road matching samples;

[0025] Preferably, the consistency score of multiple terminal soft tags within the same road segment The probability distribution of congestion level, speed range, and abnormal perception probability generated by multiple mobile phone sensing terminals within the same road segment and time window is determined based on the distribution distance between these probabilities.

[0026] When the distribution distance is less than the consistency threshold, the consistency score of the soft label of multiple terminals within the same road segment is improved. When the distribution distance is greater than the anomaly threshold and the road segment matching quality score of the corresponding terminal is... When the value is below the matching threshold, the corresponding local soft tag will be excluded from trusted federated distillation aggregation.

[0027] As a further improvement, in step S4, the road segment-level distillation label is determined as follows:

[0028] In the formula, For road segment i, distillation labels are generated at the road segment level within time window t; The set of mobile phone sensing terminals that match road segment i within time window t; For the u-th mobile phone participating in the perception terminal, a local soft tag is generated for road segment i; Let the distillation and aggregation weights of the u-th mobile phone participating in the sensing terminal for road segment i be denoted as ;

[0029] The distillation polymerization weight Based on the credibility of participation The scarcity of road segment coverage and the historical contribution of the terminal are used to determine the contribution of road segment samples with high coverage scarcity and high participation credibility in distillation polymerization.

[0030] As a further improvement, in step S5, the road topology edge is determined by the upstream and downstream connection relationship, turning relationship and lane travel direction of the road;

[0031] The traffic pattern similarity edges are determined by the similarity of historical congestion level sequences, the similarity of road functional attributes, and the similarity of time periods;

[0032] The participating trusted edge is generated only when the congestion level transfer delay between two road segments falls within the allowable delay range determined by the road travel time, and the aggregate participation credibility of the corresponding participating terminal is not lower than the edge generation threshold; when the prediction residual corresponding to the participating trusted edge increases continuously and the consistency of the multi-terminal soft label is lower than the consistency threshold, the participating trusted edge is attenuated or deleted.

[0033] Preferably, the trusted edge for participation perception is jointly determined by the congestion level transfer relationship of the participation perception data in adjacent time windows, the allowable passage delay between road segments, the aggregated participation credibility, and the prediction residual; wherein, the trusted edge for participation perception is used to transmit congestion level information between road segments that lack fixed detectors or have sparse historical samples.

[0034] Prior to this, the edge weights of the participating trusted edges are updated as follows:

[0035] ;

[0036] In the formula, The trusted edge weights for participation perception of road segment i and road segment j in time window t+1; Let λ be the perceived credible edge weight for road segment i and road segment j within time window t; λ is the edge weight memory coefficient. The cross-segment credibility is obtained by aggregating the participation credibility of participating terminals on road segment i and road segment j; The congestion state transition correlation between road segment i and road segment j; This refers to the edge weight correction term determined based on historical prediction residuals;

[0037] when Below the edge retention threshold, or within multiple consecutive time windows When the residual is less than the residual correction threshold, delete the trusted edge between road segment i and road segment j.

[0038] As a further improvement, in step S6, the spatiotemporal graph learning model includes a graph coding layer, a time series coding layer, and a multi-task prediction layer;

[0039] The graph coding layer performs message passing for road topology edges, traffic pattern similarity edges, and trusted edges involved in perception, and sets different attention weights according to the edge type.

[0040] The time-series coding layer encodes historical congestion level sequences, road segment-level distillation label sequences, time context features, and road attribute features;

[0041] The multi-task prediction layer synchronously outputs the probability of congestion level, the probability of speed range, and the prediction uncertainty within multiple future prediction time windows.

[0042] Specifically, when the prediction uncertainty exceeds the uncertainty threshold, the influence weight of the prediction result of that road segment in the subsequent confidence feedback update is reduced.

[0043] As a further improvement, the method also includes step S7: adjusting the participation confidence level based on the prediction residual between the actual returned participation perception data and the road condition prediction result. The system participates in the perception of trusted edge weights and global edge model parameters for feedback updates; when the prediction residual increases and the consistency of soft labels across multiple terminals within the same road segment falls below the consistency threshold, the participation credibility of the corresponding mobile phone participating in the perception is reduced. When the prediction residual increases but the consistency of soft labels across multiple terminals within the same road segment is not lower than the consistency threshold, maintain the credibility of the participation of the corresponding mobile phone-based sensing terminal. And prioritize adjusting the perceived trust edge weights of participants to ensure that the trustworthiness of the same participant is... It is also used for end-side sample admission, federated distillation and polymerization, participation in the lifecycle control of the perceived trustworthy edge, and prediction residual feedback update;

[0044] Preferably, in step S7, the predicted residual feedback update includes the following steps:

[0045] Calculate the prediction residual of road segment i within time window t. ;

[0046] When predicting residuals When the residual value exceeds the residual threshold and the consistency of soft tags across multiple terminals within the same road segment is below the consistency threshold, it is determined that the corresponding prediction residual mainly originates from low-confidence sensing data, thus reducing the confidence of the corresponding mobile sensing terminal's participation. ;

[0047] When predicting residuals When the residual value is greater than the residual threshold and the consistency of soft labels across multiple terminals within the same road segment is not lower than the consistency threshold, it is determined that the corresponding prediction residual mainly originates from sudden changes in the actual traffic state, thus maintaining the credibility of the participation of the corresponding mobile phone-based sensing terminals. This will increase the frequency of participation-aware trusted edge weight updates associated with the road segment.

[0048] Secondly, the present invention also provides a traffic prediction system based on trusted federated distillation-dynamic road network map learning, including a mobile phone participating in the perception terminal, a trusted federated distillation server, a dynamic road network map learning module and a traffic prediction publishing module;

[0049] The mobile phone participating in the sensing terminal is used to collect positioning data and motion sensor data, and to perform vehicle coordinate system correction, fixed time window segmentation, road segment matching, local soft tag generation, and participation credibility assessment. calculate;

[0050] The trusted federated distillation server is used to receive, without receiving the original trajectory and original motion sensor time series, local soft tags uploaded by the mobile phone participating sensing terminal, the dimensionality-reduced edge model intermediate embedding, road segment anonymized identifiers, time window anonymized identifiers, and participation trustworthiness. And based on the credibility of participation Trusted federated distillation aggregation of road segment coverage scarcity and terminal historical contribution;

[0051] The dynamic road network graph learning module is used to construct a dynamic road network graph including road topology edges, traffic pattern similarity edges, and trusted edges for participation perception. Based on congestion level transfer delay, aggregation participation credibility, multi-terminal soft label consistency, and prediction residual, the trusted edges for participation perception are generated, attenuated, deleted, or enhanced.

[0052] The traffic prediction and publishing module is used to input road segment-level distillation labels, dynamic road network maps, historical congestion level sequences, road attribute features and time context features into the spatiotemporal graph learning model, and output and publish the traffic prediction results of each road segment within multiple future prediction time windows.

[0053] The system is configured to perform the method described.

[0054] As a further improvement, the mobile phone participating in the perception terminal includes a task triggering unit, a coordinate correction unit, a road matching unit, an end-side model inference unit, a credibility calculation unit, and a privacy upload unit;

[0055] The task triggering unit is used to trigger traffic perception tasks based on the vehicle's motion state, the phone's posture stability, and the remaining energy consumption.

[0056] The coordinate correction unit is used to convert motion sensor data in the mobile phone coordinate system into motion features in the vehicle driving coordinate system.

[0057] The road matching unit is used to match positioning data and motion features to road segments;

[0058] The edge model inference unit is used to generate congestion level probability, speed range probability, and anomaly detection probability.

[0059] The credibility calculation unit is used to generate a credibility score based on mobile phone attitude correction residual, road segment matching residual, terminal sensor noise level, multi-terminal soft tag consistency, and prediction residual feedback. ;

[0060] The privacy upload unit is used to block the upload of the original trajectory and the original motion sensor time series, and only uploads local soft tags, the dimensionality-reduced edge model intermediate embedding, road segment anonymized identifiers, time window anonymized identifiers, and participation credibility. .

[0061] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method described above.

[0062] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0063] This invention collects positioning and motion sensor data on the mobile phone participating in the perception terminal, and triggers traffic perception tasks by combining vehicle motion status, mobile phone posture stability, and remaining power consumption status. This reduces invalid sampling and low-quality data uploads, lowers terminal power consumption, and improves the effectiveness of the perception data. Furthermore, by introducing a participation reliability determined by mobile phone posture correction residuals, road segment matching residuals, terminal sensor noise levels, multi-terminal soft tag consistency, and prediction residual feedback terms, this invention further enhances the perception effectiveness. It can quantitatively distinguish the data quality of different mobile terminals, different time windows, and different road segments, avoiding the direct use of low-confidence samples for model training and server aggregation; by uploading only local soft labels, edge model intermediate embedding, road segment anonymization labels, time window anonymization labels, and participation confidence. This approach can reduce the risk of leakage of original trajectories and original sensor sequences while maintaining the road segment-level road condition learning capability. Through trusted federated distillation aggregation, it gives higher weight to high-trust terminals, scarce road segment samples, and terminals with high historical contributions in distillation label generation, thereby improving the prediction availability of low-coverage and sparsely data-sparse road segments. By constructing a dynamic road network graph containing road topology edges, traffic pattern similarity edges, and trusted edges of participating perception, it can simultaneously express road physical connectivity, historical traffic pattern similarity, and real-time congestion propagation relationships reflected by participating perception data. Furthermore, it assesses participation trustworthiness based on prediction residuals and multi-terminal soft label consistency. By participating in the feedback update of perceived credible edge weights, it can distinguish between sudden changes in real traffic conditions and noise from low-credibility terminals, reducing the situation of falsely penalizing reliable terminals or incorrectly reinforcing abnormal edge weights. This enables the end-side sample admission, federated distillation and aggregation, dynamic graph edge weight update, and short-term traffic prediction to form a closed-loop adaptive optimization, thereby improving the accuracy, robustness, privacy protection, and continuous learning ability of traffic prediction. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the overall structure of a trusted federated distillation-dynamic road network map learning traffic condition prediction system for mobile phone-based traffic data perception, according to the present invention.

[0065] Figure 2 This is a schematic diagram of the structure of the mobile phone participating in the sensing terminal in this invention.

[0066] Figure 3 This is a schematic diagram of the method flow of the present invention.

[0067] Figure 4 This is a schematic diagram illustrating the credibility calculation and sample admission control in this invention.

[0068] Figure 5 This is a schematic diagram of the reliable federated distillation polymerization process in this invention.

[0069] Figure 6 This is a schematic diagram of dynamic road network graph construction and edge weight update in this invention.

[0070] Figure 7 This is a schematic diagram of the prediction residual feedback closed-loop update in this invention.

[0071] Figure 8 This is a comparison chart of prediction accuracy and speed error for different road types.

[0072] Figure 9 This is a schematic diagram illustrating the changes in the weights of trusted edges involved in perception in a dynamic road network graph.

[0073] Figure 10 A comparison chart of the prediction results for the two-branch residual feedback.

[0074] Figure 11 This is a comparison chart of the accuracy, recall, and real-world congestion identification rate of different schemes in the ablation experiment.

[0075] Figure 12 This chart compares communication volume and prediction accuracy under different upload methods. Detailed Implementation

[0076] The present invention will be further described below with reference to the accompanying drawings. It should be understood that the following embodiments are used to explain the technical solutions of the present invention and do not limit the scope of protection of the present invention. Those skilled in the art can make equivalent substitutions or adaptive adjustments to the data sampling frequency, time window length, number of model structure layers, threshold range, road segment division method, etc., based on the disclosure of this specification, as long as the technical concepts of participation confidence-through end-side sample admission, confidence-federated distillation and aggregation, dynamic road network graph edge weight update, and prediction residual feedback closed loop are still adopted, all of which fall within the scope of the present invention.

[0077] I. Terminology Explanation

[0078] In this embodiment, "mobile phone participating in sensing terminal" refers to a mobile intelligent terminal installed with traffic sensing software, which can be a smartphone, in-vehicle intelligent terminal, tablet terminal, or mobile device with positioning and motion sensor data collection capabilities. Figure 2 As shown, the mobile phone participating in the perception terminal includes at least a task triggering unit, a data acquisition unit, a coordinate correction unit, a road matching unit, an edge model inference unit, a credibility calculation unit, and a privacy upload unit.

[0079] A "road segment" refers to the smallest unit for predicting traffic conditions in an urban road network, defined by intersections, ramps, road class change points, speed limit change points, lane direction change points, or a predetermined length. Each road segment has attributes such as road number, road direction, road class, number of lanes, speed limit, length, upstream and downstream connections, and permissible turning relationships.

[0080] "Local soft labels" refer to the probabilistic results output by the mobile terminal participating in the perception based on the edge model of the current road segment's traffic status, rather than a single hard classification label. Local soft labels can include congestion level probability, speed range probability, and anomaly perception probability. Congestion levels can be divided into smooth traffic, mostly smooth traffic, slow traffic, congested traffic, and severe congestion; speed ranges can be set according to road level; anomaly perception probability is used to characterize the likelihood of sudden braking, abnormal stopping, road disturbance, accidents, or sensor malfunctions.

[0081] "Participation credibility" refers to a dynamic indicator used to quantify the reliability of the participation sensing data generated by the u-th mobile phone participating terminal within a time window t, denoted as This metric is jointly determined by mobile phone attitude stability, road segment matching quality, multi-terminal soft label consistency, terminal historical contribution, energy consumption constraints, and prediction residual feedback. Participation credibility is not merely used as a score for uploaded data, but is also used for edge-side sample admission, trusted federated distillation and aggregation, participation-aware trusted edge lifecycle control, and prediction residual feedback updates.

[0082] "Trusted federated distillation" refers to a process where, without receiving the original trajectory and original motion sensor time series, the server receives local soft tags uploaded by multiple mobile sensing terminals, intermediate embeddings of the edge model after dimensionality reduction, anonymous road segment identifiers, anonymous time window identifiers, and participation credibility. Based on the participation credibility, road segment coverage scarcity, and terminal historical contribution, the server performs weighted distillation aggregation to obtain road segment-level distilled tags, while simultaneously updating the global edge model parameters.

[0083] A "dynamic road network map" refers to a dynamic graph structure built using road segments as nodes and the traffic relationships between road segments as edges. For example... Figure 6 As shown, this dynamic road network map includes road topology edges, traffic pattern similarity edges, and participant-perceived trusted edges. Road topology edges describe physically traversable connections; traffic pattern similarity edges describe relationships between road segments with similar historical congestion patterns that are not necessarily directly adjacent; and participant-perceived trusted edges describe congestion propagation relationships revealed by real-time participant-perceived data.

[0084] "Participation-aware trusted edge" refers to a dynamic edge that meets specific generation conditions. These conditions include: the congestion level transfer delay between two road segments falls within an allowable delay range determined by road travel time, and the aggregate participation trustworthiness of the corresponding participating terminal is not lower than the edge generation threshold. This edge can be strengthened, weakened, or deleted based on real-time data and prediction residual feedback, such as... Figure 9 As shown.

[0085] "Prediction residual feedback" refers to comparing the predicted results with the actual feedback data from the sensing system after the model outputs the prediction results for future road segments, obtaining the prediction residual, and combining this with the consistency of multi-terminal soft labels to determine the source of the residual. When the prediction residual is large and the consistency among multiple terminals is low, the system's judgment error mainly comes from low-confidence noise; when the prediction residual is large but the consistency among multiple terminals is high, the system's judgment error mainly comes from sudden changes in the actual traffic state, and the system prioritizes updating the dynamic graph edge weights without penalizing reliable terminals.

[0086] II. System Structure

[0087] like Figure 1 As shown, the system of the present invention includes a mobile phone-participating sensing terminal, a trusted federated distillation server, a dynamic road network map learning module, a traffic condition prediction and publishing module, and a data management module.

[0088] Mobile phones participate in the sensing terminal to complete edge data collection, task triggering, coordinate correction, road segment matching, local soft tag generation, participation in credibility calculation, and privacy-preserving uploading. For example... Figure 2 As shown, the mobile phone participating in the perception terminal acquires multi-source data during vehicle operation through GNSS positioning, accelerometers, angular velocity sensors, and optional magnetometers, barometers, and system motion state interfaces. The task triggering unit determines whether to initiate the traffic perception task based on the vehicle's motion state, the phone's attitude stability, and remaining power consumption. The coordinate correction unit converts the motion sensor data in the phone's coordinate system into motion features in the vehicle's driving coordinate system. The road matching unit matches the positioning data and motion features to road segments. The edge-side model inference unit generates local soft tags. The credibility calculation unit calculates the participation credibility. The privacy upload unit is used to block the upload of the original trajectory and the original motion sensor time series, and only uploads local soft tags, dimensionality reduction embeddings, anonymous road segment identifiers, anonymous time window identifiers, and participation credibility.

[0089] The trusted federated distillation server receives privacy-preserving intermediate information uploaded by multiple mobile sensing terminals and performs distillation aggregation based on participation trust, road segment coverage scarcity, and terminal historical contribution. For example... Figure 5As shown, the server generates road segment-level distillation labels for multiple terminal soft labels within the same road segment and the same time window. At the same time, it updates the global end-side model parameters based on the distillation results and sends the updated model or model increment to the mobile phone participating in the perception terminal.

[0090] The dynamic road network map learning module is used to build and maintain dynamic map structures. For example... Figure 6 As shown, the dynamic road network graph includes three types of edges: road topology edges, traffic pattern similarity edges, and trusted edges for participation perception. The dynamic road network graph learning module updates the weights of trusted edges for participation perception based on congestion state transition correlation, cross-segment participation credibility, and historical prediction residual correction terms. When an edge weight falls below the retention threshold or the edge weight correction term fails to meet the conditions for multiple consecutive time windows, the corresponding trusted edge for participation perception is deleted.

[0091] The traffic forecasting and publishing module takes road segment-level distillation labels, dynamic road network maps, historical congestion level sequences, road attribute features, and temporal context features as inputs to a spatiotemporal graph learning model, and outputs traffic forecast results for each road segment within multiple future prediction time windows. These results can be used in navigation applications, traffic management department backends, public travel service interfaces, and congestion warning systems.

[0092] The data management module stores basic road data, road segment attribute data, anonymous terminal data, time window indexes, distillation labels, dynamic graph edge weights, model versions, prediction results, and prediction residual records. The data management module can create multi-level indexes based on city regions, road segment numbers, and time windows to support real-time queries, model training, and historical playback.

[0093] III. Technical Route for Implementing the Method of the Invention

[0094] like Figure 3 As shown, the method of the present invention includes the following steps: S1 End-side data acquisition and task triggering; S2 Participation in credibility calculation and sample admission; S3 Local soft tag generation and privacy-preserving upload; S4 Trusted federated distillation and aggregation; S5 Dynamic road network map construction; S6 Spatiotemporal map learning and prediction; S7 Prediction residual feedback update.

[0095] The technical approach of this invention is not simply adding mobile phone data collection, federated learning, and graph neural networks together, but rather incorporating the credibility of the participants. Serving as a unified control variable across the edge, cloud, and graph learning processes, this participation confidence level determines whether a sample participates in training and uploading at the edge, determines distillation and aggregation weights at the server, decides whether to generate, enhance, decay, or delete participation-aware confidence edges during dynamic graph learning, and adaptively updates based on prediction residuals and multi-terminal consistency during prediction feedback. Through this closed loop, the system can handle common problems in mobile phone participation perception scenarios, such as unstable mobile phone poses, road matching errors, terminal hardware differences, sparse samples in low-coverage road sections, and the propagation of sudden congestion.

[0096] (I) S1, End-side data acquisition, task triggering and road segment-level feature generation

[0097] The mobile phone, as a sensing terminal, first collects location data and motion sensor data. Location data includes at least one of GNSS positioning data, base station positioning data, and wireless network positioning data. Motion sensor data includes at least acceleration and angular velocity data, and may also include magnetometer data, barometer data, and system motion state data. System motion state data can be provided by the mobile phone's operating system or the vehicle's in-vehicle system, used to initially determine whether the terminal is walking, stationary, cycling, or in motion.

[0098] To reduce energy consumption and invalid uploads, the terminal does not perform continuous high-intensity sampling. Instead, it determines whether to initiate a traffic perception task through a task triggering unit. Triggering conditions may include: the terminal's speed or location change trend meets vehicle driving characteristics; the phone's posture remains relatively stable over several consecutive short time windows; the terminal's remaining battery power is higher than a preset threshold or it is charging; the current road segment is a low-coverage road segment or a road area where the system needs to supplement samples. By filtering according to the above conditions, invalid traffic samples can be avoided when the user is walking, the phone is idle, or the phone is frequently operated.

[0099] After collecting motion sensor data, coordinate calibration is required. Since the phone may be placed in a holder, cup holder, center console, seat, or user's pocket, the phone's coordinate system is usually inconsistent with the vehicle's driving coordinate system. The coordinate calibration unit first estimates the phone's vertical orientation based on the direction of gravity, and then estimates the rotational relationship between the phone's coordinate system and the vehicle's driving coordinate system based on the principal direction of longitudinal acceleration during vehicle acceleration and deceleration. Let the acceleration vector of the k-th sampling point in the phone's coordinate system be... The acceleration vector in the vehicle coordinate system is The rotation matrix from the mobile phone coordinate system to the vehicle coordinate system is: ,but:

[0100] ;

[0101] In the formula, Let be the acceleration vector of the kth sampling point in the vehicle coordinate system; Let be the acceleration vector of the k-th sampling point in the mobile phone coordinate system; This is the rotation matrix used to transform the coordinate system from the mobile phone coordinate system to the vehicle coordinate system within the time window t.

[0102] After correction, the system obtains the vehicle's longitudinal acceleration, lateral acceleration, and vertical disturbance components. The longitudinal acceleration reflects acceleration, deceleration, and following distance; the lateral acceleration reflects turning, lane changing, and ramp driving; and the vertical disturbance component identifies road bumps, bridge expansion joints, or abnormal mobile phone movement. If the change in the rotation matrix exceeds the attitude stability threshold within adjacent time windows, it indicates that the mobile phone may have moved or rotated, and the reliability of the samples in the corresponding time window should be reduced.

[0103] The road matching unit matches positioning data and vehicle motion characteristics to road segments. Road matching is based not only on the distance from the positioning point to the candidate road segment, but also on factors such as the angle between the vehicle's direction of travel and the direction of the candidate road segment, the connectivity between the road segment in the previous time window and the current candidate road segment, the deviation between the vehicle speed and the road speed limit, and the degree of matching between road grade and speed characteristics. In scenarios such as elevated roads running parallel to ground-level roads, main roads running parallel to auxiliary roads, and ramps adjacent to main roads, the system prioritizes using vehicle speed, road direction, vertical change trends, and historical road segment connectivity for differentiation, thereby reducing road matching errors.

[0104] After the above processing, each effective time window generates road segment-level end-side traffic features. These features include average speed, rate of change of speed, mean longitudinal acceleration, variance of longitudinal acceleration, number of rapid decelerations, number of rapid accelerations, peak lateral acceleration, proportion of stopping time, proportion of low-speed driving, road matching quality score, mobile phone posture stability score, and sensor noise level. These features are input into the end-side model to generate local soft labels and participate in credibility calculations.

[0105] (ii) S2, Participation in credibility calculation and sample admission control

[0106] Participating in credibility calculation is a key technical aspect of this invention, such as... Figure 4 As shown. For the sensing results of the u-th mobile phone participating in the sensing terminal within the time window t, the system calculates the participation confidence level. In one specific implementation, the participation credibility is determined according to the following formula:

[0107] ;

[0108] In the formula, The credibility of the participation of the u-th mobile phone as a sensing terminal within time window t; Rate the phone's posture stability; Assess the quality of road segment matching; A consistency score for soft labels across multiple terminals within the same road segment; The terminal's historical contribution score; This is a penalty term for energy consumption constraints; This is the penalty term for the forecast residuals from the previous forecast period; , , , , , These are the weights of the corresponding items; For the normalization mapping function, either the Sigmoid function or a piecewise normalization function can be used.

[0109] Mobile phone posture stability score The score is determined based on the magnitude of changes in the rotation matrix, the magnitude of changes in the gravity direction, the peak angular velocity, and the duration of attitude disturbances within adjacent time windows. The score decreases when the user picks up the phone, switches between portrait and landscape modes, puts it in their pocket, or performs frequent operations. Road segment matching quality score. The score is determined based on the distance from the location point to the candidate road segment, the angle between the driving direction and the candidate road segment, the connectivity of the candidate road segment, and the degree of matching with the speed range. The score is reduced when the distance from the location point to the candidate road segment exceeds a distance threshold, or when the angle between the vehicle's driving direction and the candidate road segment's direction exceeds an angle threshold.

[0110] Multi-terminal soft tag consistency scoring The probability distribution of congestion level, speed range, and abnormal perception generated by multiple mobile phone sensing terminals within the same road segment and time window is determined based on the distribution distance between them. This distribution distance can be calculated using Jensen-Shannon distance, cosine distance, symmetric KL divergence, or Wasserstein distance. When the soft labels given by most terminals for the same road segment are similar, High; when the output of a terminal is significantly inconsistent with that of most terminals and its road matching quality score is low, the sample of that terminal is marked as a low confidence sample.

[0111] Terminal historical contribution score The historical contribution score is determined based on the contribution of soft tags uploaded by the terminal over several past prediction periods to the final prediction result. If a terminal's previously uploaded data is consistent with subsequent actual feedback results and it repeatedly supplements low-coverage road segment samples, its historical contribution score is increased. If a terminal repeatedly produces large residuals that are inconsistent with most terminals, its historical contribution score is decreased.

[0112] Energy consumption constraint penalty item This is used to reduce high-frequency sampling and uploading when the phone's battery is below a preset threshold and it is not charging. The system reduces the sampling frequency or only allows edge-side inference without uploading intermediate embeddings. (Prediction residual penalty term) Feedback results from the previous prediction period. If a terminal's soft tag causes a large prediction residual and is not supported by other terminals on the same road segment, then... Increase; if the prediction residual is large but the consistency among multiple terminals is high, then the residual is not directly used to penalize the terminal, but serves as possible evidence of abrupt changes in the actual traffic state.

[0113] Sample admission control is implemented according to the following rules: when the residual of mobile phone pose correction or road segment matching exceeds the corresponding threshold, samples in that time window are prohibited from participating in local training; when the residual does not exceed the threshold but participates in confidence... When the sample size is below the confidence threshold, only samples within that time window are allowed to participate in edge-side inference; they are not allowed to be uploaded to the server for federated distillation. When the trust threshold is reached, the sample is allowed to generate a local soft tag and be uploaded in a privacy-protected manner.

[0114] The contribution of this step is that participation credibility is not an independent quality assessment, but rather forms a common basis for both edge-side sample admission and cloud-side distillation and aggregation. By first removing pose anomalies, road matching anomalies, and low-confidence samples at the edge, the accumulation of low-quality data on the server side can be reduced, thereby lowering communication volume and the risk of model drift.

[0115] (III) S3, Local Soft Tag Generation and Privacy Upload

[0116] The edge-side model inference unit generates local soft labels based on road segment-level edge-side traffic features. Local soft labels can be represented as:

[0117] ;

[0118] In the formula, For the uth mobile phone participating in the perception terminal, a local soft tag is generated for road segment i within time window t; The probability distribution of congestion levels; This represents the probability distribution of the velocity range. This represents the probability distribution for anomaly detection.

[0119] Regarding the edge model structure, lightweight one-dimensional convolutional networks, gated recurrent networks, lightweight Transformer encoders, or multilayer perceptrons can be used. To adapt to the computing power of ordinary mobile phones, this implementation prefers a lightweight model structure. The input consists of statistical features and short sequence features within a fixed time window, and the output consists of congestion level probability, speed interval probability, and anomaly detection probability. The model's hidden dimension can be set to 32 to 128, and the number of network layers can be set to 2 to 4. The model can be distributed to the terminal through quantization, pruning, or incremental updates.

[0120] The privacy upload unit does not upload the original trajectory, original acceleration time series, original angular velocity time series, or plaintext terminal identifiers that can directly reconstruct continuous travel paths. Uploaded content includes local soft tags. Embedded in the middle of the dimension-reduced end-side model Road segment anonymization, time window anonymization, and participation credibility Among them, the middle embedding It can be output from the penultimate layer of the end-side model and compressed through random projection, principal component dimensionality reduction, or quantization encoding.

[0121] Road segment anonymization identifiers can be generated from road segment numbers and daily random salt value hashes, while time window anonymization identifiers can be generated from minute-level time window indexes and regional random salt value hashes. Through these methods, the server can perform distillation aggregation at the road segment and time window levels, but it cannot directly obtain the user's continuous trajectory and raw sensor sequences.

[0122] (iv) S4, Trusted Federated Distillation Polymerization and Global End-Side Model Update

[0123] like Figure 5 As shown, the server performs trusted federated distillation aggregation on local soft tags uploaded by multiple terminals within the same road segment i and the same time window t. Road segment-level distillation tags are determined as follows:

[0124] ;

[0125] In the formula, For road segment i, distillation labels are generated at the road segment level within time window t; The set of mobile phone sensing terminals that match road segment i within time window t; For the u-th mobile phone participating in the perception terminal, a local soft tag is generated for road segment i; Let be the distillation aggregation weight for road segment i by the u-th mobile phone participating in the sensing terminal.

[0126] The weighting of distillation polymerization can be determined as follows:

[0127] ;

[0128] In the formula, Weights for distillation polymerization; To enhance credibility; The coverage scarcity of road segment i within time window t; The historical contribution of the u-th terminal.

[0129] Among them, coverage scarcity This is used to improve the contribution of low-coverage road segment samples in distillation aggregation. When the number of valid terminals for a branch road, suburban road, or ramp segment is small within a single prediction period, the soft tags uploaded by that terminal will receive higher weight in distillation aggregation if the participating terminals have high credibility. Historical contribution. The determination is based on the prediction consistency, contribution frequency of low-coverage road sections, and residual feedback results over multiple past cycles of the terminal.

[0130] Trusted federated distillation is also used to update the global edge model parameters. The server uses road segment-level distillation labels as teacher signals to train the global edge model. The training loss can include distillation loss, classification consistency loss, and trust regularization term:

[0131] ;

[0132] In the formula, The training loss for the global edge model; The distillation loss between the road segment-level distillation label and the model output; The classification loss is for labeled samples; A trustworthy regular expression set based on the trustworthiness of participants; and This is the loss weighting coefficient.

[0133] After training, the server can distribute global edge model parameters or model increments to the mobile sensing terminals. For terminals with weaker computing power, a lightweight model can be distributed; for high-performance vehicle terminals, an enhanced model can be distributed. This step enables the system to continuously absorb sensing knowledge from different roads, time periods, and vehicle operating states without uploading the original data.

[0134] (V) S5, Dynamic Road Network Map Construction and Lifecycle Control of Participating Sensing Trusted Edges

[0135] like Figure 6 As shown, the server constructs a dynamic road network graph using road segments as nodes. The dynamic road network graph includes road topology edges, traffic pattern similarity edges, and trusted participation perception edges.

[0136] Road topology edges are determined by the upstream and downstream connections, turning relationships, and lane travel directions. If the exit of road segment i can lead to the entrance of road segment j, then a directed road topology edge is established from i to j. This type of edge is mainly determined by the road base database and is updated with events such as road construction, closure, and one-way adjustments.

[0137] Traffic pattern similarity edges are determined by the similarity of historical congestion level sequences, road functional attributes, and time period similarity. Even if two road segments are not directly connected, as long as they have similar morning and evening peak congestion patterns, similar road levels, and similar commuting attributes, traffic pattern similarity edges can be established. Similarity can be determined through dynamic time-warped distance, correlation coefficient, Wasserstein distance, or clustering results.

[0138] The key edge type in this invention is the perceived trustworthy edge. The system analyzes the congestion level transition relationship between road segment i and road segment j within adjacent time windows. If road segment i experiences an increase in congestion level during time window t, and road segment j experiences an increase in congestion level during time window t... The corresponding congestion level has increased, and If the distance falls within the allowable delay range determined by the road travel time, and the aggregation credibility of participating terminals on the two road segments is not lower than the edge generation threshold, then a participating perception credible edge is generated.

[0139] The edge weights involved in sensing trusted edges are updated according to the following formula:

[0140] ;

[0141] In the formula, The trusted edge weights for participation perception of road segment i and road segment j in time window t+1; Assess the credible edge weights for the participation of road segment i and road segment j within time window t; The edge weight memory coefficient; The cross-segment credibility is obtained by aggregating the participation credibility of participating terminals on road segment i and road segment j; The congestion state transition correlation between road segment i and road segment j; This is the edge weight correction term determined based on historical prediction residuals.

[0142] when Below the edge retention threshold, or within multiple consecutive time windows When the residual is less than the residual correction threshold, delete the trusted edge between road segment i and road segment j. Figure 9The process of edge generation, enhancement, decay, and deletion involving perceived trustworthy edges is further illustrated. Edges are generated when the passage delay meets the requirements and the aggregation trustworthiness is high; edges are enhanced when the prediction residual decreases and the trustworthiness continues to increase; edges are decayed when the prediction residual continuously increases or the consistency decreases; and edges are deleted when the edge weight is below the retention threshold.

[0143] The contribution of this step is that the dynamic road network map not only expresses the physical topology of roads, but also the congestion propagation relationships revealed by real-time participant sensing data, thereby improving the ability to predict sudden congestion, temporary traffic control, construction disturbances, and low-coverage road segments.

[0144] (vi) S6, Spatiotemporal Graph Learning Prediction

[0145] The spatiotemporal graph learning model is used to predict road conditions for road segments within multiple future time windows based on dynamic road network maps and historical sequences. The model input includes road segment-level distillation labels, dynamic road network maps, historical congestion level sequences, road attribute features, and temporal context features.

[0146] The input features of road segment i within time window t can be represented as:

[0147] ;

[0148] In the formula, Input features for the model of road segment i within time window t; For road segment distillation labeling; Characteristics of historical congestion level sequence; Road attribute features; This is a temporal context feature.

[0149] The model comprises a graph coding layer, a time-series coding layer, and a multi-task prediction layer. The graph coding layer performs message passing for road topology edges, traffic pattern similarity edges, and trusted edges involved in perception, and assigns weights to different edge types. For road segment i, its graph coding result can be represented as:

[0150] ;

[0151] In the formula, The graph encoding result for road segment i within time window t; It is a set of edge types, including road topology edges, traffic pattern similarity edges, and trusted edges for perception participation; Let i be the set of neighboring nodes connected to road segment i via edge of type r; Let J be the attention weight of road segment j on the edge of class r of road segment i. Let r be the feature transformation matrix corresponding to the r-th type edge; The input features of road segment j within time window t; It is a non-linear activation function.

[0152] The time-series coding layer captures the patterns of traffic state changes over time and can employ gated recurrent units, long short-term memory networks, temporal convolutional networks, or lightweight Transformer encoders. The multi-task prediction layer simultaneously outputs the congestion level probability, speed range probability, and prediction uncertainty. When the prediction uncertainty exceeds an uncertainty threshold, the system reduces the weight of that prediction in subsequent confidence feedback updates.

[0153] The model training loss can be expressed as:

[0154] ;

[0155] In the formula, The total loss of the spatiotemporal graph learning model; The prediction loss is the difference between the predicted result and the actual or returned road condition labels. The distillation loss between the model output and the road segment-level distillation label; The loss is due to edge weight constraints in a dynamic graph. For model regularization loss; , , This is the loss weighting coefficient.

[0156] Model training is divided into three stages: offline pre-training, online federated distillation update, and dynamic graph joint fine-tuning. The offline pre-training stage uses historical road traffic data, fixed detector data, floating car speed data, and historical participant perception data; the online federated distillation update stage uses local soft labels and dimensionality reduction embeddings uploaded from mobile devices; and the dynamic graph joint fine-tuning stage updates the participant perception trusted edge weights based on prediction residuals and multi-terminal consistency.

[0157] Regarding parameter selection, the edge-side time window length can be set to 5 to 60 seconds, preferably 10 to 30 seconds; the server prediction granularity can be set to 1 to 5 minutes; the historical input window can be set to 6 to 24 time windows; the prediction output window can be set to 1 to 12 future time windows; the participation confidence threshold can be set to 0.5 to 0.8; and the edge weight memory coefficient... The value can be set to 0.5 to 0.95, preferably 0.7 to 0.9; the number of graph encoding layers can be set to 1 to 4; the hidden dimension can be set to 32 to 256.

[0158] In terms of dataset usage, training data includes a road infrastructure database, historical congestion level data, speed or traffic flow data from fixed detectors, floating car speed data, privacy-preserving soft tags and dimensionality-reduced embeddings uploaded by mobile sensing terminals, weather data, holiday data, and road event data. The data can be divided into training, validation, and test sets by time, for example, the first 70% of the time period as the training set, the middle 15% as the validation set, and the last 15% as the test set; it can also be divided by region to verify the model's generalization ability on low-coverage road segments.

[0159] (vii) S7, Predictive residual feedback closed-loop update

[0160] like Figure 7 As shown, after each prediction cycle, the system compares the model prediction results with the subsequent actual feedback sensing data to obtain the prediction residual. The prediction residual of road segment i in time window t can be expressed as:

[0161] ;

[0162] In the formula, Let i be the prediction residual of road segment i within time window t; This represents the model's previous prediction of road segment i within time window t; This serves as a reference label based on the actual feedback of the sensing data. This is either a distribution distance function or a weighted error function.

[0163] When predicting residuals When the error exceeds the residual threshold and the consistency of soft tags across multiple terminals within the same road segment is below the consistency threshold, the system determines that the error mainly originates from low-confidence sensing data or abnormal terminal uploads. In this case, the participation confidence of the corresponding mobile phone sensing terminal is reduced. And reduce its subsequent distillation polymerization weight. When predicting residuals If the error exceeds the residual threshold, but the consistency of soft tags across multiple terminals within the same road segment is not lower than the consistency threshold, the system determines that the error mainly originates from sudden changes in the actual traffic state. In this case, the system maintains the credibility of the participation of the corresponding mobile phone-based sensing terminal. And prioritize adjusting the perceived trustworthy edge weights associated with the road segment.

[0164] Figure 10 The effect of predictive residual feedback bi-branch judgment is demonstrated. Compared with the method of penalizing terminals only based on residuals, the present invention can reduce the proportion of falsely penalized reliable terminals and the proportion of abnormal edge error enhancement, thereby improving the accuracy of identifying real sudden congestion.

[0165] IV. Examples, Comparative Examples, and Experimental Data

[0166] (I) Experimental Platform and Evaluation Methods

[0167] The experimental area selected included the core urban area, expressway connecting areas, and suburban branch road areas of a city. The road network was divided into 1268 road segments, including 462 arterial road segments, 381 secondary arterial road segments, 291 branch road segments, and 134 expressway and ramp segments. The experiment lasted for 14 consecutive days, with the data from the first 10 days used for training and parameter calibration, and the data from the last 4 days used for independent testing. The participating terminals included 53 smartphones of different models. The sensor sampling frequency was set to 20Hz, the positioning sampling interval was 1s, the terminal time window length was 15s, and the server prediction time granularity was 5 minutes, predicting road conditions for the next 5, 10, and 15 minutes.

[0168] Evaluation metrics include congestion level accuracy, macro average F1, speed MAE, speed RMSE, low-coverage road segment recall rate, accuracy of identifying real-time congestion spikes, proportion of reliable terminals falsely penalized, proportion of abnormal edge error enhancement, amount of uploaded data, and additional terminal power consumption. Speed ​​MAE is calculated using the following formula:

[0169] ;

[0170] In the formula, This represents the average absolute velocity error. This represents the number of test samples; The prediction speed for the i-th test sample; This is the reference speed for the i-th test sample.

[0171] The RMSE (Recovery Rate) is calculated using the following formula:

[0172] ;

[0173] In the formula, This is the root mean square error; This represents the number of test samples; The prediction speed for the i-th test sample; This is the reference speed for the i-th test sample.

[0174] (II) Example 1: Participation in Credibility Access Control Verification

[0175] This embodiment corresponds to Figure 2 and Figure 4 The system filters samples with abnormal phone pose, abnormal road matching, and low consistency across multiple terminals. The confidence threshold is set to 0.65. If the pose correction residual or road segment matching residual exceeds the threshold, the corresponding time window sample is not included in local training; if the residual does not exceed the threshold but... If the value is less than 0.65, it only participates in edge-side inference and is not uploaded; if If the value is not less than 0.65, then upload local soft tags, dimensionality reduction embedding, anonymous road segment identifiers, anonymous time window identifiers, and participation credibility.

[0176] Table 1. End-side sample admission control effect

[0177]

[0178] As shown in Table 1, after adopting this invention, the accuracy of the edge congestion level soft tag increased from 84.7% to 90.8%, the false alarm rate of anomaly detection decreased from 11.3% to 6.1%, and the average daily upload volume per terminal decreased from 42.6MB to 18.4MB, indicating that... Figure 4 The participation credibility calculation and sample admission control shown improve sample quality and reduce communication and energy consumption burden.

[0179] (III) Example 2: Prediction effect of reliable federated distillation under different road types

[0180] This embodiment corresponds to Figure 5 and Figure 8 The server determined the distillation polymerization weights based on participation credibility, road segment coverage scarcity, and terminal historical contribution. The test results are shown in Table 2.

[0181] Table 2 Prediction Results for Different Road Types

[0182]

[0183] like Figure 8 As shown in Table 2, this invention improves the accuracy of congestion level assessment and reduces speed errors on arterial roads, secondary arterial roads, local roads, expressways, and ramps. The improvement is most significant on local roads, indicating the coverage scarcity. It can improve the utilization of low-coverage road segment samples.

[0184] (iv) Example 3: Dynamic road network map and verification of trusted edges involved in perception

[0185] This embodiment corresponds to Figure 6 and Figure 9 Thirty-eight congestion propagation events during the morning rush hour were selected, including queue overflow at main road intersections, congestion spillover at expressway entrances, short-term congestion around schools, and temporary congestion at construction sites. The dynamics of only road topology edges, topology edges plus traffic pattern similarity edges, and the three types of edges described in this invention were compared. Figure 3 The results of the two models are shown in Table 3.

[0186] Table 3. Impact of Dynamic Road Network Map on Congestion Propagation Prediction

[0187]

[0188] As shown in Table 3, the present invention achieves higher accuracy in the 5-minute, 10-minute, and 15-minute prediction tasks. The 15-minute prediction accuracy increased from 76.8% to 84.9%, the early identification rate of congestion propagation increased from 62.5% to 81.6%, and the proportion of false propagation edges decreased from 14.7% to 6.8%. Figure 9 To further explain, the participating trusted edges can be generated, enhanced, attenuated, and deleted based on trustworthiness and residual feedback, thus preventing erroneous edges from being retained for a long time.

[0189] (V) Example 4: Verification of the two-branch judgment of prediction residual feedback

[0190] This embodiment corresponds to Figure 7 and Figure 10 120 high residual prediction samples were selected for verification, of which 52 were real sudden congestion samples and 68 were low-confidence terminal noise or road matching error samples. The results of comparing the method of penalizing terminals based solely on residuals with the residual plus consistency dual-branch feedback method of this invention are shown in Table 4.

[0191] Table 4. Effect of the Predicted Residual Feedback Mechanism

[0192]

[0193] From Table 4 and Figure 10 As can be seen, this invention can avoid misclassifying real sudden congestion as terminal noise. The accuracy of real sudden congestion identification is improved from 73.1% to 90.4%, the proportion of falsely penalized reliable terminals is reduced from 18.6% to 6.2%, and the proportion of abnormal edge error enhancement is reduced from 13.9% to 5.7%.

[0194] (vi) Comparative Example 1: No participation credibility is used, only normal upload and average aggregation are used.

[0195] Comparative Example 1 does not calculate participation credibility. The server performs ordinary averaging aggregation on the uploaded results without considering attitude residuals, road matching residuals, multi-terminal consistency, and prediction residual feedback for sample admission.

[0196] Table 5 Comparison of the overall effects of Comparative Example 1 and the present invention

[0197]

[0198] As shown in Table 5, compared with the ordinary aggregation scheme without credibility, the present invention improves the overall congestion level accuracy by 6.5 percentage points, the recall rate of low-coverage road sections by 14.4 percentage points, and reduces the average daily upload volume per terminal by about 56.6%.

[0199] (vii) Comparative Example 2: Centralized Graph Neural Network Prediction but without Trusted Federated Distillation

[0200] Comparative Example 2 employs a centralized graph neural network for traffic prediction. The server trains the model using historical traffic data and results uploaded from ordinary mobile terminals, but does not use trusted federated distillation, nor does it use local soft labels and edge-side model embedding for privacy-preserving collaborative training. The dynamic graph structure only includes road topology edges and traffic pattern similarity edges, excluding trusted edges involved in perception.

[0201] Table 6 compares the effects of Comparative Example 2 and the present invention on road sections with different coverage levels.

[0202]

[0203] As shown in Table 6, the advantages of this invention are more obvious on low-coverage road sections, with the congestion level accuracy increasing from 78.6% to 86.9% and the recall rate on low-coverage road sections increasing from 70.1% to 84.3%.

[0204] (viii) Comparative Example 3: Federated learning is adopted, but dynamic road network graph edge weight lifecycle control is not adopted.

[0205] Comparative Example 3 uses federated learning to train the edge model and uses a static road topology graph for prediction, but does not construct trusted edges for perception, nor does it generate, enhance, attenuate, or delete edge weights based on the prediction residuals.

[0206] Table 7 compares the impact of Example 3 and the present invention on congestion propagation prediction.

[0207]

[0208] As shown in Table 7, this invention improves the ability to identify congestion propagation in advance and the model recovery ability after sudden congestion by participating in the perception of trusted edges and the prediction residual feedback mechanism.

[0209] (ix) Ablation test

[0210] To verify the synergistic effect of the key technical features, the following schemes were set up: Scheme A removed end-side sample admission control and retained only trusted federated distillation; Scheme B retained end-side sample admission control, but distillation aggregation did not use road segment coverage scarcity; Scheme C retained trusted federated distillation, but did not construct trusted edges for perception; Scheme D retained trusted edges for perception, but did not use prediction residual bi-branch judgment; Scheme E adopted the complete scheme of this invention. Experimental results are shown in Table 8 and can be correlated with... Figure 11 illustrate.

[0211] Table 8 Ablation Experiment Results

[0212]

[0213] From Table 8 and Figure 11It can be seen that the complete solution achieves better results in terms of overall accuracy, recall rate of low-coverage road sections, and accuracy in identifying real sudden congestion. Removing any key mechanism will cause varying degrees of decline, indicating that there is a synergistic effect among edge sample admission, trusted federated distillation, trusted edges involved in perception, and prediction residual feedback.

[0214] (x) Privacy-protected upload and communication overhead verification

[0215] This embodiment corresponds to Figure 12 A comparison is made between the original data upload scheme and the privacy-enhancing upload scheme of this invention. The original data upload scheme uploads continuous trajectory points, original acceleration sequences, and original angular velocity sequences; this invention only uploads local soft tags, dimensionality reduction embeddings, anonymous road segment identifiers, anonymous time window identifiers, and participation credibility.

[0216] Table 9 Comparison of communication volume and prediction accuracy under different upload methods.

[0217]

[0218] From Table 9 and Figure 12 It can be seen that, without uploading the original trajectory and the original motion sensor time series, the average data upload volume per time window of this invention is reduced from 18.6KB to 4.2KB, and the traffic prediction accuracy is increased from 90.1% to 91.4%. This shows that this invention can maintain and improve traffic prediction performance while reducing the risk of privacy leakage and communication overhead.

[0219] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

[0220] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0221] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0222] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0223] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0224] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0225] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0226] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

Claims

1. A road condition prediction method based on trusted federated distillation-dynamic road network graph learning, characterized in that, The method includes the following steps: S1. Multiple mobile phones participate as sensing terminals to collect positioning data and motion sensor data during vehicle operation, and trigger traffic sensing tasks based on vehicle motion state, mobile phone posture stability and remaining power consumption; perform vehicle coordinate system correction, fixed time window segmentation and road segment matching on the motion sensor data to obtain road segment-level end-side traffic features. S2. For each mobile phone participating in the perception terminal within each time window, generate participation credibility based on the mobile phone attitude correction residual, road segment matching residual, terminal sensor noise level, consistency of soft tags of multiple terminals within the same road segment, and prediction residual feedback item. and according to Control whether the corresponding sample participates in local training, edge inference, and federated distillation upload; S3. The mobile phone participating in the perception terminal generates local soft tags based on the local model, including congestion level probability, speed range probability, and abnormal perception probability, and only uploads the local soft tags, the dimensionality-reduced edge model intermediate embedding, the road segment anonymous identifier, the time window anonymous identifier, and the participation credibility. ; S4, server according to Based on the scarcity of road segment coverage and the historical contribution of terminals, multiple local soft tags within the same road segment and the same time window are subjected to trusted federated distillation aggregation to obtain road segment-level distillation tags. S5. The server constructs a dynamic road network graph including road topology edges, traffic pattern similarity edges, and trusted edges for participation perception. The trusted edges for participation perception are generated, enhanced, attenuated, or deleted based on the congestion level transfer delay, aggregated participation credibility, and prediction residual. S6. Input the road segment-level distillation labels, dynamic road network map, historical congestion level sequence, road attribute features and time context features into the spatiotemporal graph learning model, and output the road condition prediction results for multiple future prediction time windows.

2. The method according to claim 1, characterized in that, In step S1, the positioning data includes at least one of GNSS positioning data, base station positioning data, and wireless network positioning data; The motion sensor data includes at least one of acceleration data, angular velocity data, magnetometer data, barometer data, and terminal system motion status data. The vehicle coordinate system correction includes: estimating the rotational relationship between the phone attitude coordinate system and the vehicle driving coordinate system based on the direction of gravity; converting the acceleration components in the phone coordinate system into vehicle longitudinal acceleration components, vehicle lateral acceleration components, and vehicle vertical disturbance components; and determining the phone attitude stability based on the change amplitude of the rotational relationship within adjacent time windows. And / or, in step S2, when the mobile phone pose correction residual or road segment matching residual exceeds the corresponding threshold, the sample in that time window is prohibited from participating in local training; when the mobile phone pose correction residual and road segment matching residual do not exceed the corresponding threshold, but participate in the confidence level... When the sample size is below the confidence threshold, only samples within that time window are allowed to participate in edge inference but not in federated distillation uploads. And / or, in step S2, participate in credibility Determine as follows: In the formula, The credibility of the participation of the u-th mobile phone as a sensing terminal in time window t; Rate the phone's posture stability; Assess the quality of road segment matching; For consistency scoring of soft labels across multiple terminals within the same road segment; The terminal's historical contribution score; This is a penalty term for energy consumption constraints; This is the penalty term for the forecast residuals from the previous forecast period; , , , , , These are the weights of the corresponding terms; σ is the normalized mapping function; Preferably, the road segment matching quality score The distance from the location point to the candidate road segment, the angle between the vehicle's driving direction and the candidate road segment's direction, the connectivity of road segments in adjacent time windows, and the deviation between the vehicle's speed and the road speed limit range are all determined together. When the distance from the location point to the candidate road segment is greater than a distance threshold, or the angle between the vehicle's driving direction and the candidate road segment's direction is greater than an angle threshold, the road segment matching quality score is reduced. And mark the samples in this time window as low-confidence road matching samples; Preferably, the consistency score of multiple terminal soft tags within the same road segment The probability distribution of congestion level, speed range, and abnormal perception probability generated by multiple mobile phone sensing terminals within the same road segment and time window is determined based on the distribution distance between these probabilities. When the distribution distance is less than the consistency threshold, the consistency score of the soft label of multiple terminals within the same road segment is improved. ; When the distribution distance is greater than the anomaly threshold and the corresponding terminal's road segment matching quality score is... When the value is below the matching threshold, the corresponding local soft tag will be excluded from trusted federated distillation aggregation.

3. The method according to claim 1, characterized in that, In step S4, the road segment-level distillation label is determined as follows: In the formula, For road segment i, the road segment level distillation label is defined within time window t; The set of mobile phone sensing terminals that match road segment i within time window t; The local soft tag generated by the u-th mobile phone participating in the sensing terminal for road segment i; For the u-th mobile phone participating in the perception terminal, the distillation and aggregation weights are for road segment i. The distillation polymerization weight Based on the credibility of participation The scarcity of road segment coverage and the historical contribution of the terminal are used to determine the contribution of road segment samples with high coverage scarcity and high participation credibility in distillation polymerization.

4. The method according to claim 1, characterized in that, In step S5, the road topology edge is determined by the upstream and downstream connection relationship, turning relationship and lane travel direction; The traffic pattern similarity edges are determined by the similarity of historical congestion level sequences, the similarity of road functional attributes, and the similarity of time periods; The participating trusted edge is generated only when the congestion level transfer delay between two road segments falls within the allowable delay range determined by the road travel time, and the aggregate participation credibility of the corresponding participating terminal is not lower than the edge generation threshold; when the prediction residual corresponding to the participating trusted edge increases continuously and the consistency of the multi-terminal soft label is lower than the consistency threshold, the participating trusted edge is attenuated or deleted. Preferably, the trusted edge for participation perception is jointly determined by the congestion level transfer relationship of the participation perception data in adjacent time windows, the allowable passage delay between road segments, the aggregated participation credibility, and the prediction residual; wherein, the trusted edge for participation perception is used to transmit congestion level information between road segments that lack fixed detectors or have sparse historical samples. Prior to this, the edge weights of the participating trusted edges are updated as follows: ； In the formula, The trusted edge weights for participation perception of road segment i and road segment j in time window t+1; Let λ be the perceived credible edge weight for road segment i and road segment j within time window t; λ is the edge weight memory coefficient. The cross-segment credibility is obtained by aggregating the participation credibility of participating terminals on road segment i and road segment j; The congestion state transition correlation between road segment i and road segment j; This refers to the edge weight correction term determined based on historical prediction residuals; when Below the edge retention threshold, or within multiple consecutive time windows When the residual is less than the residual correction threshold, delete the trusted edge between road segment i and road segment j.

5. The method according to claim 1, characterized in that, In step S6, the spatiotemporal graph learning model includes a graph coding layer, a time series coding layer, and a multi-task prediction layer; The graph coding layer performs message passing for road topology edges, traffic pattern similarity edges, and trusted edges involved in perception, and sets different attention weights according to the edge type. The time-series coding layer encodes historical congestion level sequences, road segment-level distillation label sequences, time context features, and road attribute features; The multi-task prediction layer synchronously outputs the probability of congestion level, the probability of speed range, and the prediction uncertainty within multiple future prediction time windows. Specifically, when the prediction uncertainty exceeds the uncertainty threshold, the influence weight of the prediction result of that road segment in the subsequent confidence feedback update is reduced.

6. The method according to claim 1, characterized in that, The method further includes step S7: assessing the participation confidence level based on the prediction residual between the actual returned participation sensing data and the road condition prediction result. The system participates in the perception of trusted edge weights and global edge model parameters for feedback updates; when the prediction residual increases and the consistency of soft labels across multiple terminals within the same road segment falls below the consistency threshold, the participation credibility of the corresponding mobile phone participating in the perception is reduced. When the prediction residual increases but the consistency of soft labels across multiple terminals within the same road segment is not lower than the consistency threshold, maintain the credibility of the participation of the corresponding mobile phone-based sensing terminal. And prioritize adjusting the perceived trust edge weights of participants to ensure that the trustworthiness of the same participant is... It is also used for end-side sample admission, federated distillation and polymerization, participation in the lifecycle control of the perceived trustworthy edge, and prediction residual feedback update; Preferably, in step S7, the predicted residual feedback update includes the following steps: Calculate the prediction residual of road segment i in time window t. ; When predicting residuals When the residual value exceeds the residual threshold and the consistency of soft tags across multiple terminals within the same road segment is below the consistency threshold, it is determined that the corresponding prediction residual mainly originates from low-confidence sensing data, thus reducing the confidence of the corresponding mobile sensing terminal's participation. ; When predicting residuals When the residual value is greater than the residual threshold and the consistency of soft labels across multiple terminals within the same road segment is not lower than the consistency threshold, it is determined that the corresponding prediction residual mainly originates from sudden changes in the actual traffic state, thus maintaining the credibility of the participation of the corresponding mobile phone-based sensing terminals. This will increase the frequency of participation-aware trusted edge weight updates associated with the road segment.

7. A road condition prediction system based on trusted federated distillation-dynamic road network graph learning, characterized in that, This includes a mobile phone-based sensing terminal, a trusted federated distillation server, a dynamic road network map learning module, and a traffic condition prediction and publishing module; The mobile phone participating in the sensing terminal is used to collect positioning data and motion sensor data, and to perform vehicle coordinate system correction, fixed time window segmentation, road segment matching, local soft tag generation, and participation credibility assessment. calculate; The trusted federated distillation server is used to receive, without receiving the original trajectory and original motion sensor time series, local soft tags uploaded by the mobile phone participating sensing terminal, the dimensionality-reduced edge model intermediate embedding, road segment anonymized identifiers, time window anonymized identifiers, and participation trustworthiness. And based on the credibility of participation Trusted federated distillation aggregation of road segment coverage scarcity and terminal historical contribution; The dynamic road network graph learning module is used to construct a dynamic road network graph including road topology edges, traffic pattern similarity edges, and trusted edges for participation perception. Based on congestion level transfer delay, aggregation participation credibility, multi-terminal soft label consistency, and prediction residual, the trusted edges for participation perception are generated, attenuated, deleted, or enhanced. The traffic prediction and publishing module is used to input road segment-level distillation labels, dynamic road network maps, historical congestion level sequences, road attribute features and time context features into the spatiotemporal graph learning model, and output and publish the traffic prediction results of each road segment within multiple future prediction time windows. The system is configured to perform the method according to any one of claims 1 to 6.

8. The system according to claim 7, characterized in that, The mobile phone participating in the perception terminal includes a task triggering unit, a coordinate correction unit, a road matching unit, an end-side model inference unit, a credibility calculation unit, and a privacy upload unit; The task triggering unit is used to trigger traffic perception tasks based on the vehicle's motion state, the phone's posture stability, and the remaining energy consumption. The coordinate correction unit is used to convert motion sensor data in the mobile phone coordinate system into motion features in the vehicle driving coordinate system. The road matching unit is used to match positioning data and motion features to road segments; The edge model inference unit is used to generate congestion level probability, speed range probability, and anomaly detection probability. The credibility calculation unit is used to generate a credibility score based on mobile phone attitude correction residual, road segment matching residual, terminal sensor noise level, multi-terminal soft tag consistency, and prediction residual feedback. ; The privacy upload unit is used to block the upload of the original trajectory and the original motion sensor time series, and only uploads local soft tags, the dimensionality-reduced edge model intermediate embedding, road segment anonymized identifiers, time window anonymized identifiers, and participation credibility. .

9. An electronic device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 7.

Images