Traffic travel service quality evaluation system based on multi-source data fusion

Abstract

The invention relates to the technical field of traffic management and information, and particularly discloses a traffic travel service quality evaluation system based on multi-source data fusion. The system comprises a high-precision space-time synchronization fusion unit, a service quality index calculation engine, a dynamic weight adaptive module, an evaluation result credibility quantification unit and a visual decision support interface. Through online time stamp calibration, dynamic adjustment of data source fusion weight and quantification of uncertainty of an evaluation result, the system can generate high-precision and high-credibility multi-dimensional service quality indexes, and provides reliable support for traffic management decision. According to the system, a high-precision time-space synchronization fusion unit is constructed, a dynamic clock skew model is innovatively established and maintained for each heterogeneous data acquisition device, online and accurate calibration of an original timestamp is realized, and the problem of time asynchronization caused by hardware clock drift is eliminated from the root of data fusion.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of traffic management and information technology, and particularly relates to a traffic travel service quality evaluation system based on multi-source data fusion. BACKGROUND

[0002] In the field of intelligent transportation and urban computing, accurate evaluation of traffic travel service quality is a core link for optimizing traffic resource allocation and improving public travel experience. Service quality evaluation relies on comprehensive analysis of multi-dimensional information such as complete travel trajectory, time consumption and environmental state of travelers, and its accuracy directly determines the effectiveness of management decisions.

[0003] Traffic travel service quality evaluation based on multi-source data fusion is an important technical direction. This technology aims to integrate heterogeneous data from different sensors, mobile terminals and infrastructure to build a more comprehensive and detailed travel profile, thereby realizing quantitative evaluation of various service modes such as public transportation and shared travel.

[0004] Existing technologies usually rely on GPS, base station positioning, Bluetooth beacons and vehicle-mounted sensors to collect data and perform fusion processing to restore travel trajectories. However, these collection devices have inherent drift in their hardware clocks, resulting in system deviations of microseconds or even milliseconds in the timestamps generated by each device. In the data fusion process, this asynchrony of timestamps can cause serious trajectory matching errors, such as incorrectly associating location points of different times and different travelers, or splitting continuous location points of the same trip.

[0005] This underlying data fusion distortion directly affects the upper-layer evaluation model, significantly increasing the error rate of the key evaluation indicator "travel time", and thus affecting the overall judgment of service punctuality, efficiency and reliability, reducing the credibility and practical value of the evaluation results. Therefore, how to realize multi-source traffic data fusion under high-precision time synchronization to support reliable service quality evaluation has become a technical problem to be solved. SUMMARY

[0006] The purpose of the present application is to provide a traffic travel service quality evaluation system based on multi-source data fusion to solve the technical contradiction in the prior art that trajectory fusion distortion is caused by asynchronous timestamps of multi-source data collection devices, which seriously affects the accuracy of travel service quality evaluation.

[0007] The technical solution of the present application is a traffic travel service quality evaluation system based on multi-source data fusion, which comprises: The high-precision spatiotemporal synchronization fusion unit is used to receive raw traffic data packets from multiple heterogeneous data sources, and to perform online calibration of the timestamp and unified mapping of the spatial location of each data packet to generate fused trajectory data with a unified spatiotemporal reference. The service quality index calculation engine is used to extract individual travel chains based on the fused trajectory data output by the high-precision spatiotemporal synchronous fusion unit, and calculate multi-dimensional service quality indicators including travel time, punctuality rate, travel reliability and comfort according to the preset evaluation model. The dynamic weight adaptive module is used to analyze the data quality characteristics of different data sources in a specific spatiotemporal environment in real time, and dynamically adjust the contribution weight of each data source in the spatiotemporal synchronization fusion process based on the quality characteristics. The evaluation result credibility quantification unit is used to perform uncertainty propagation calculation on each indicator calculated by the service quality indicator calculation engine, combined with the confidence information in the data fusion process and the weight information output by the dynamic weight adaptive module, and to add a quantified credibility score to each evaluation result. A visual decision support interface is used to render and overlay the evaluation results and their corresponding credibility scores, which are organized in a spatiotemporal grid.

[0008] As one embodiment of the present invention, the high-precision spatiotemporal synchronization fusion unit includes a data packet parsing subunit, a clock deviation estimation subunit, a timestamp calibration subunit, and a spatial coordinate transformation subunit.

[0009] The packet parsing subunit is used to extract the device identifier, original timestamp, original location coordinates, and system reference time of the packet reception from the received raw data packet.

[0010] The clock skew estimation subunit is used to maintain a dynamic clock skew model for each active data acquisition device. This model is built and updated through the following process: continuously acquiring data packets sent by the same device at multiple times, recording the original timestamp and system reference time for each data packet; calculating the time skew for each data packet based on the system reference time; and fitting the time skew sequence using an exponentially weighted sliding window regression algorithm to obtain the current clock skew value and drift rate of the device relative to the system reference time.

[0011] The timestamp calibration subunit is used to input the raw timestamp extracted by the data packet parsing subunit into the clock deviation model of the corresponding device, and use the current deviation value and drift rate output by the model to perform compensation calculations to generate a calibrated accurate timestamp.

[0012] The spatial coordinate transformation subunit is used to transform the original location coordinates extracted by the data packet parsing subunit, regardless of whether they come from the Global Navigation Satellite System, base station positioning, or Bluetooth beacon, into coordinates in the standard plane coordinate system used by the urban geographic information system through a preset coordinate transformation parameter matrix.

[0013] Furthermore, the service quality indicator calculation engine includes a travel chain reconstruction subunit and an indicator calculation subunit.

[0014] The travel chain reconstruction subunit is used to sort the fused trajectory data points in time sequence based on the calibrated accurate timestamps, and apply a density-based spatial clustering algorithm to aggregate the sequence of location points belonging to the same traveler and with continuous time into a travel trajectory. Then, combined with the preset traffic network topology data, the trajectory point sequence is matched to specific road segments or public transportation lines through a map matching algorithm to identify the origin, destination, transfer points and mode of transportation used for each trip.

[0015] The indicator calculation subunit incorporates multiple parallel calculation logics. The trip time calculation logic calculates the actual trip time based on the precise timestamps of the origin and destination points recorded in the trip chain. The punctuality calculation logic compares the actual trip time with the preset planned trip time or historical average trip time for the corresponding route to determine if the trip is on time. The trip reliability calculation logic calculates the coefficient of variation of trip time between a specific route or a pair of origin and destination points within a statistical period. The comfort calculation logic analyzes derived parameters representing motion states in the trajectory data, including the rate of change of acceleration (jerkness) and the distribution characteristics of stop points, and outputs a comfort level based on preset thresholds.

[0016] Furthermore, the dynamic weight adaptive module includes a data quality feature extraction subunit and a fusion weight decision subunit.

[0017] The data quality feature extraction subunit is used to monitor multiple quality dimensions for each data source in real time. These dimensions include the stability of data reporting frequency, the accuracy dilution factor of location coordinates, signal strength values, and the fitting residuals of historical clock skew models. This subunit periodically calculates a normalized quality score for each quality dimension of each data source.

[0018] The fusion weight decision subunit receives the multi-dimensional quality scores output by the data quality feature extraction subunit. This subunit maintains a multi-input, single-output adaptive neural network model. The input layer nodes of this model correspond to the scores of each quality dimension, and the output layer nodes represent the fusion weight values ​​between 0 and 1. This neural network model is trained using historical labeled data to learn the optimal fusion weights for different combinations of quality features. For a real-time input quality score vector, this subunit calculates and directly outputs the dynamic fusion weights of the data source at the current moment through forward propagation.

[0019] Furthermore, the high-precision spatiotemporal synchronous fusion unit simultaneously records the data source combination and its corresponding dynamic fusion weight for each fused trajectory data point when generating it. The reliability quantification unit uses this information to perform uncertainty propagation calculations. This calculation process, based on the error propagation law, progressively transmits the calibration uncertainty of the underlying data points in time and space, as well as the uncertainty introduced by weight allocation during multi-source data fusion, to the final service quality index through the functional relationship of the mathematical model used in the index calculation engine. The unit ultimately outputs a value between 0 and 1 as a reliability score, which is inversely proportional to the uncertainty of the index calculation result.

[0020] In one embodiment of the present invention, a visual decision support interface divides the urban geographic space into a regular square grid. This interface receives the mean values ​​and credibility scores of various service quality indicators, with the grid as the statistical unit, output by the credibility quantification unit of the evaluation results. The interface employs a dual-layer overlay rendering technique. The bottom layer uses a gradient color scheme from red to green to visually display the levels of the indicator values, while the top layer uses a semi-transparent grayscale mask to represent the credibility score; the higher the score, the more transparent the mask, and the lower the score, the deeper the mask. Decision-makers can adjust the layer transparency to simultaneously observe the distribution pattern of service quality and the credibility of the evaluation results.

[0021] Furthermore, the system runs on a distributed stream processing framework. High-precision spatiotemporal synchronization fusion units, service quality index calculation engines, dynamic weight adaptive modules, and evaluation result reliability quantification units are deployed as independent processing topologies. Data flows sequentially between topologies via message queues, enabling real-time processing throughout the entire process, from raw data access to evaluation result generation.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention innovatively establishes and maintains a dynamic clock deviation model for each heterogeneous data acquisition device by constructing a high-precision spatiotemporal synchronization fusion unit. This enables online and accurate calibration of the original timestamps, eliminating the time asynchrony problem caused by hardware clock drift at the root of data fusion. This ensures accurate alignment of multi-source trajectory data under a unified spatiotemporal reference, fundamentally avoiding trajectory matching errors and journey fragmentation. It provides highly reliable input data for upper-layer service quality assessment and significantly improves the accuracy of core assessment indicators such as journey time.

[0023] 2. This invention introduces a dynamic weight adaptive module, which can perceive data quality fluctuations from different data sources in real time under changing environments and intelligently decides the fusion weights based on multi-dimensional quality characteristics using an adaptive neural network. This design makes the data fusion process no longer static or homogeneous, but rather context-aware and capable of selecting the best option. When the quality of a specific data source declines, the system automatically reduces its contribution, thereby effectively suppressing the contamination of the overall fusion result by low-quality noisy data, further enhancing the robustness and adaptability of the system output.

[0024] 3. This invention creatively integrates a unit for quantifying the reliability of assessment results, upgrading the single deterministic index output by traditional assessment systems into a binary tuple of "index value reliability score." This unit, through rigorous uncertainty propagation calculations, quantitatively transmits the inherent uncertainties during the underlying data calibration and fusion process to the final assessment result, presenting it in the form of an intuitive reliability score. This mechanism enables decision-makers to clearly identify the reliability of assessment conclusions, avoiding misjudgments based on indicators derived from low-reliability data, and greatly enhancing the reference value and application safety of assessment results in actual traffic management decisions. Combined with the dual-layer rendering of the visual decision support interface, it achieves effective communication and in-depth utilization of data reliability information. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall technical solution architecture of the transportation service quality assessment system based on multi-source data fusion proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of high-precision spatiotemporal synchronization fusion and dynamic weight adaptive fusion in this invention; Figure 3 This is a flowchart of the main stages of the service quality index calculation engine in this invention. Figure 4 This is a schematic diagram illustrating the propagation of uncertainty in the credible quantification of data from multiple sources to the evaluation results in this invention. Figure 5 This is a schematic diagram illustrating the principle of dual-layer overlay rendering of the evaluation results and credibility scores in the visual decision support interface of this invention. Detailed Implementation

[0026] This invention provides a transportation service quality assessment system based on multi-source data fusion. Please refer to the appendix. Figure 1 To be continued Figure 5 The system's overall architecture consists of a high-precision spatiotemporal synchronization fusion unit, a service quality index calculation engine, a dynamic weight adaptive module, a reliable quantification unit for evaluation results, and a visual decision support interface. These modules are deployed on a distributed stream processing framework, forming a complete processing pipeline from real-time access of raw data to the visualization of evaluation results. Data flows sequentially between processing topologies through high-throughput message queues, ensuring the system's real-time processing capability for massive amounts of multi-source traffic data.

[0027] The high-precision spatiotemporal synchronization fusion unit is the core of this system's data preprocessing and fusion. This unit continuously receives raw traffic data packets from various heterogeneous data acquisition devices deployed in the urban transportation network. These data sources include, but are not limited to, vehicle-mounted GPS terminals installed in buses and taxis, base station positioning data reported by passengers' smartphones via mobile communication networks, and short-range sensing data collected by Bluetooth beacons deployed in key transportation hubs such as subway stations and bus stops.

[0028] Each raw data packet contains a unique device identifier, a raw timestamp generated by the device's own clock, and raw location coordinates obtained based on a specific positioning technology. The task of the high-precision spatiotemporal synchronization fusion unit is to perform spatiotemporal alignment on these asynchronous and heterogeneous raw data to generate a fused trajectory data stream with a unified spatiotemporal reference.

[0029] Please refer to the attached document. Figure 2 The high-precision spatiotemporal synchronization fusion unit contains a data packet parsing subunit, a clock deviation estimation subunit, a timestamp calibration subunit, and a spatial coordinate transformation subunit, and each subunit works in concert.

[0030] The packet parsing subunit, acting as the data entry point, is responsible for decapsulating and extracting fields from each incoming raw data packet. This subunit listens to the incoming message queue, and upon capturing a data packet, immediately executes the parsing protocol. The parsing process first verifies the data packet's format integrity and protocol compliance, discarding invalid data packets that fail verification.

[0031] For valid data packets, the parsing subunit precisely extracts four key fields: device identifier, which is a string or numeric code used to uniquely distinguish the data source; raw timestamp, which is the time recorded by the data acquisition device according to its local hardware clock when it senses location information, usually expressed in milliseconds or microseconds since a certain epoch; and raw location coordinates, the format of which depends on the positioning technology. For example, GPS data provides latitude, longitude, and altitude, base station positioning data provides cell identifier and signal strength triangulation results, and Bluetooth beacon data provides beacon identifier and received signal strength indication value.

[0032] The system reference time at the moment of data packet reception is provided by a high-precision network time protocol clock maintained by a server cluster deploying high-precision spatiotemporal synchronization fusion units. This serves as the sole time reference for the entire system, with an accuracy down to the microsecond level. The parsing subunit assembles these four extracted fields into a structured intermediate data object and outputs it to the subsequent processing subunit.

[0033] The core function of the clock skew estimation subunit is to establish and maintain a dynamic mathematical model for each active data acquisition device, reflecting its clock skew and drift. Because different manufacturers and models of devices have inherent errors in their internal clock crystals, and these errors are also affected by factors such as temperature and voltage, a constantly changing deviation exists between the original timestamp reported by the device and the actual system reference time. This subunit assigns an independent estimator instance to each device identifier. The estimator continuously receives a series of intermediate data objects belonging to the same device from the packet parsing subunit.

[0034] For each object, the estimator records its original timestamp and system reference time. The clock skew estimation subunit employs an exponentially weighted sliding window regression algorithm to fit the device clock skew model. This algorithm maintains a fixed-length sliding window, which stores a sequence of system reference times paired with original timestamps for all data packets from the device within a recent time period. For newly arriving data points, older data points are assigned exponentially decaying weights based on their time proximity, with more recent data points receiving higher weights.

[0035] The algorithm performs linear regression on the weighted data points to fit a straight line that best represents the trend of device clock skew over time. The intercept of this line is the current average deviation of the device clock relative to the system reference time, and its slope is the clock drift rate, representing the rate at which the deviation changes over time. The model is periodically updated as new data is continuously received, typically with a refit triggered every 10 new data points or every 30 seconds to adapt to the dynamic changes in clock drift. This sub-unit outputs a state object for each device containing the current deviation value, drift rate, and model fit confidence parameters.

[0036] The timestamp calibration subunit uses the dynamic model output by the clock skew estimation subunit to accurately correct the original timestamp. When an intermediate data object is received, the timestamp calibration subunit retrieves the corresponding clock skew model based on its device identifier. The calibration calculation follows a defined compensation formula. This subunit first calculates the model time elapsed from the device clock's reference epoch to the original timestamp of the current data packet. Then, it multiplies the model time by the current drift rate correction factor obtained from the skew model, and adds the current skew value to obtain the total compensation time.

[0037] Finally, this compensation time is added to the original timestamp to obtain the calibrated accurate timestamp. This accurate timestamp is theoretically aligned with the system reference time base, eliminating asynchronous errors caused by inaccurate hardware clocks. The calibrated timestamp is recorded along with the original timestamp and the system reference time.

[0038] The spatial coordinate transformation subunit is responsible for mapping raw location coordinates in various formats to a unified geospatial reference system. This subunit contains a pre-set or dynamically loaded database of high-precision coordinate transformation parameters. For latitude, longitude, and elevation coordinates provided by the Global Navigation Satellite System (GNSS), the transformation subunit calls the appropriate map projection transformation algorithm, such as the Universal Transverse Mercator projection or the Gauss-Krüger projection, and combines it with the local geoid model to convert it into Cartesian coordinates, typically in meters, used in urban geographic information systems.

[0039] For base station positioning data, this sub-unit first queries the base station location database based on the cell identifier to obtain the antenna center coordinates of the serving base station. Then, combining received signal strength values, time difference of arrival, and other multi-dimensional information, it estimates the location of the mobile device using a specific positioning engine algorithm, such as a weighted centroid algorithm or fingerprint matching algorithm, and converts it to a standard plane coordinate system. For Bluetooth beacon data, it queries the precise installation location coordinates based on the beacon identifier, and estimates the distance between the device and the beacon by combining the received signal strength indication value with a preset beacon signal attenuation model, thereby determining the device's location. All conversion processes use the same coordinate transformation parameter matrix to ensure that location data from different sources can be accurately superimposed in space. The conversion sub-unit outputs the unified plane coordinates.

[0040] After completing timestamp calibration and spatial coordinate transformation, the high-precision spatiotemporal synchronization fusion unit generates fused trajectory data points. These data points include device identifiers, calibrated precise timestamps, unified planar coordinates, data source type identifiers, and raw data characteristics transmitted from the data packet parsing subunit. These data points are arranged in chronological order, forming a pre-cleaned and synchronized trajectory stream, ready to be delivered to downstream units.

[0041] The dynamic weight adaptive module works closely with the high-precision spatiotemporal synchronization fusion unit. Its goal is to evaluate the quality of each data source in real time and intelligently adjust their influence during the fusion process. Please refer to the appendix for further details. Figure 2 This module includes a data quality feature extraction subunit and a fusion weight decision subunit.

[0042] The data quality feature extraction subunit performs multi-dimensional quality monitoring on each incoming raw data packet or intermediate data object. The monitoring dimensions are preset and include data reporting frequency stability, location accuracy dilution factor, signal strength, and clock deviation model fitting residual. For reporting frequency stability, this subunit maintains a sequence of data packet arrival time intervals for each device over a recent period and calculates the standard deviation or coefficient of variation of the sequence. The more stable the interval, the higher the score.

[0043] The position accuracy dilution factor is directly derived from a field in the Global Navigation Satellite System (GNSS) data. A smaller value indicates a better positioning geometry and higher accuracy. For non-satellite positioning data, this dimension is simulated by estimating the area or radius of the error ellipse within the positioning algorithm. Signal strength is crucial for base station and Bluetooth positioning; higher strength generally means more accurate distance estimation. The clock skew model fitting residual is obtained from the clock skew estimation subunit and reflects the goodness of fit of the current clock skew model. A small residual indicates a reliable model and high reliability of the timestamp calibration.

[0044] This sub-unit periodically, for example every 60 seconds, calculates the raw score for each quality dimension for each active data source, and then maps the raw scores of each dimension to a normalized quality score between 0 and 1, where 1 represents the best quality, using a max-min normalization or quantile normalization method.

[0045] The fusion weight decision subunit receives the multidimensional normalized quality score vector output by the data quality feature extraction subunit. The core of this subunit is an offline-trained multilayer perceptron adaptive neural network model. The number of input layer nodes in this model equals the number of quality dimensions; for example, four dimensions would have four input nodes. Hidden layers can contain one or more non-linear activation function layers. The output layer consists of nodes, using the sigmoid activation function to restrict their output to between 0 and 1, directly serving as the dynamic fusion weights for the data source. The training data for the neural network model comes from historical annotations, i.e., annotations by experts or other high-precision reference systems, indicating the optimal fusion weights to be assigned to the data source under different combinations of quality features.

[0046] The model learns a complex nonlinear mapping from quality features to optimal weights through backpropagation. During online operation, the fusion weight decision subunit performs forward propagation computation of the neural network on the quality score vector of each data source that comes in real time. The computation process does not involve iterative optimization, but is only a single-step function mapping, resulting in extremely low latency. The calculated dynamic fusion weights are fed back to the high-precision spatiotemporal synchronization fusion unit in real time. When generating the final fusion trajectory data point, the high-precision spatiotemporal synchronization fusion unit records one or more data source identifiers and their corresponding real-time dynamic fusion weights as metadata supplementary information for that data point.

[0047] The service quality metric calculation engine is responsible for extracting business-meaning evaluation metrics from the fused trajectory data. Please refer to the appendix. Figure 3 The engine mainly includes a travel chain reconstruction subunit and an indicator calculation subunit.

[0048] The travel chain reconstruction subunit processes the spatiotemporally synchronized fused trajectory data stream output by the high-precision spatiotemporal synchronization fusion unit. This subunit first groups the data points according to device identifiers. Then, within each device group, it strictly sorts all data points in ascending order based on calibrated precise timestamps, forming a time-ordered sequence of location points. Next, it applies density-based spatial clustering algorithms, such as noisy density-based clustering methods, to analyze the location point sequence.

[0049] This algorithm can identify clusters of densely connected points and remove noise points. In travel scenarios, continuous movement generates a series of spatially and temporally densely connected location points, which are clustered into the same cluster, thus separating different travel segments. Each cluster represents a potential travel trajectory. Subsequently, the travel chain reconstruction sub-unit loads pre-set high-precision urban traffic network topology data, which includes vector geometric information and topological connections of all road segments and public transportation routes. Using map matching algorithms, such as Hidden Markov Model matching or geometric topology matching algorithms, each clustered sequence of location points is matched to the most probable road segment or public transportation route. The map matching process considers not only the nearest distance between points but also the connectivity between consecutive points, the consistency between the direction of movement and the road direction, and historical matching probabilities. After successful matching, the algorithm can identify the origin, destination, key path points along the route, and transfer points of a trip.

[0050] At the same time, by combining the speed characteristics of the trajectory points, the matched road type, and the known public transportation vehicle identifier information, this sub-unit can infer the main mode of transportation used for the trip, such as subway, bus, bicycle, or walking.

[0051] Ultimately, the trip chain reconstruction sub-unit generates a structured trip chain record for each complete trip. This record includes the traveler's anonymous identifier, the precise timestamps of the trip's start and end, the coordinates of the origin and destination points, the sequence of road segments or routes, the list of transfer events, and the inferred mode of transportation.

[0052] The indicator calculation subunit runs multiple independent calculation logics in parallel, batch processing trip chain records to generate multi-dimensional service quality indicators. The trip time calculation logic is the most straightforward; it reads the trip start and end timestamps from the trip chain records, calculates the difference between them, and obtains the actual trip time in seconds. The on-time rate calculation logic requires a benchmark. This logic maintains either a planned timetable database or a historical trip time statistics database.

[0053] For public transportation trips, the system queries the scheduled travel time for a particular trip based on the route identifier, direction, and trip start time matched by the travel chain. For trips with non-fixed schedules, it uses the historical average travel time for the same origin-destination pair and time period as a benchmark. If the difference between the actual travel time and the benchmark time falls within a preset tolerance threshold, the trip is considered on time; otherwise, it is considered off-time. Within a statistical period, such as a day or an hour, the system cumulatively calculates the ratio of on-time trips to the total number of trips on a specific route or in a specific area to obtain the on-time rate index.

[0054] The trip reliability calculation logic focuses on the volatility of trip time. Within a specified statistical period and spatial range—for example, for all trips on a bus route within a week—it collects the actual trip times and calculates the ratio of the standard deviation to the mean of this time series, i.e., the coefficient of variation, as a quantitative indicator of reliability. The smaller the coefficient of variation, the more stable the trip time and the higher the reliability. The comfort calculation logic, on the other hand, performs a more in-depth analysis of the motion state. It first calculates the instantaneous velocity and acceleration of each point in the trajectory point sequence within the trip chain.

[0055] Next, the rate of change of acceleration, or jerkiness, is calculated. Jerkiness is a key physical quantity for measuring ride comfort; an excessively high absolute value of jerkiness indicates rapid acceleration or deceleration, which reduces comfort. This logic counts the number and proportion of times jerkiness exceeds a preset comfort threshold during a single trip. Simultaneously, it analyzes unnecessary stops along the route, such as prolonged low-speed wandering due to traffic congestion. By analyzing the spatiotemporal clustering of points with speeds below a certain threshold, abnormal stops are identified. Finally, the comfort calculation logic combines the proportion of jerkiness exceeding the standard with the proportion of abnormal stop durations, and outputs discrete comfort levels, such as excellent, good, average, and poor, through a preset weighted scoring model or fuzzy logic judgment model.

[0056] Specifically, a weighted scoring model is used, and the mathematical expression is: ; The overall comfort score ranges from [0,1]. The urgency weight is preset to 0.6 and can be dynamically adjusted according to the mode of transportation. The weight for abnormal stays is preset to 0.4 and can be dynamically adjusted according to the mode of transportation. ; The percentage of individuals exhibiting excessive agitation. This represents the total duration of abnormal stays. This represents the total travel time.

[0057] Comfort Level Classification Standards Excellent Good, For the middle, It is poor.

[0058] The reliability quantification unit of the evaluation results introduces an important dimension for measuring uncertainty in this system. Please refer to the appendix. Figure 4 This unit receives preliminary index calculation results from the service quality index calculation engine, as well as metadata from the high-precision spatiotemporal synchronous fusion unit, which includes data source combinations and dynamic fusion weights associated with each underlying fusion trajectory data point. The core task of this unit is to perform uncertainty propagation calculations, attaching a quantified credibility score to each calculated service quality index.

[0059] The computational basis for uncertainty propagation is the law of error propagation. The credibility quantification unit for evaluation results first quantifies the uncertainty of the underlying data. The uncertainty of timestamps originates from the fitting residuals and prediction errors of the clock skew estimation sub-unit model, which can be quantified as the time standard deviation. The uncertainty of spatial coordinates stems from inherent errors in positioning technology, such as the user equivalent distance error of the Global Navigation Satellite System and the error radius of base station positioning. This information can be obtained from the original data or during coordinate transformation and quantified as a covariance matrix in planar coordinates. The multi-source data fusion process itself also introduces uncertainty because dynamic weights are assigned based on real-time quality assessment, and the decision-making process for weights involves confidence levels, which can be converted into the variance of the weight values ​​themselves.

[0060] This unit propagates the aforementioned uncertainties upwards level by level according to the functional relationships of the mathematical model used in the service quality indicator calculation engine. For example, the trip time is a function of the difference between the end timestamp and the start timestamp in the indicator calculation sub-unit. Let the uncertainty of the start timestamp be... The uncertainty of the end timestamp is And since the two are independent, the uncertainty of the travel time is... It can be calculated using the error propagation formula: ; Indicates the uncertainty of travel time. Indicates the uncertainty of the start timestamp. Indicates the uncertainty of the end timestamp; For more complex metrics, such as punctuality rate, the calculation involves comparing the actual travel time with a baseline time. The uncertainty of the actual travel time has been derived from the above formula, and the baseline time, if derived from historical statistics, also has its own statistical variance. The punctuality rate judgment is a threshold comparison function, and the uncertainty of its output Boolean value can be derived by calculating the probability that the actual travel time falls within the threshold range.

[0061] The calculation of the coefficient of variation for travel reliability involves the mean and standard deviation, and the propagation of its uncertainty requires the use of the error propagation formula for multivariate functions. The jerkiness involved in comfort calculation is obtained by acceleration difference, and acceleration is obtained by velocity difference. This is a multi-level difference process, and the error will accumulate and amplify, so it needs to be propagated through the chain rule.

[0062] The specific implementation method is as follows: speed Calculation: The central difference method is used to calculate the instantaneous velocity based on the spatial coordinates of the trajectory points and the calibrated timestamps. The formula is as follows: ; For the first The instantaneous velocity of each trajectory point , The first , The spatial coordinates of the trajectory points , These are the calibrated timestamps for the corresponding points; the first and last trajectory points are supplemented by forward and backward difference methods.

[0063] acceleration Calculation: The central difference method is also used to calculate the acceleration of the velocity sequence. The formula is: ; , Don't be the first , The velocity of each trajectory point The initial and final accelerations are supplemented using the corresponding difference method, and are calculated using the same velocity definition.

[0064] Chain rule error propagation: Let the measurement error of the trajectory point coordinates be... The timestamp calibration error is ,but: speed error ; Acceleration error ; Acceleration error ; , Don't be the first , The acceleration of each trajectory point; By using the above recursive formula, the underlying measurement error is propagated step by step to the jerk, thereby achieving quantitative control of error accumulation.

[0065] The confidence quantification unit for the evaluation results incorporates the partial derivatives or sensitivity coefficients of these index calculation functions. It traces the complete computational chain from the original fused data points to the final index value, synthesizing the uncertainty inputs and weighted uncertainty inputs at each layer through the corresponding propagation model. Finally, it calculates the total uncertainty estimate for each index, typically expressed as the standard deviation or the half-width of the confidence interval.

[0066] The specific calculation method is as follows: Standard deviation calculate: Based on the indicator sample series within the statistical period ( For sample size, preset Using the unbiased standard deviation formula: ; Indicator Sample Mean , For the first The index values ​​of each sample.

[0067] Confidence interval half-width ( )calculate: use Distribution (small sample) ) or normal distribution (large sample) Calculate the half-width of the confidence interval at a 95% confidence level (preset confidence level, adjustable to 90% or 99%): Large sample cases ( ): ; (The standard normal distribution quantile corresponding to the 95% confidence level); Small sample cases ( ): , For degrees of freedom corresponding Distribution quantiles (e.g.) hour, ).

[0068] Parameter description: Sample size This refers to the number of valid travel records within the same evaluation unit (such as the same grid or the same route) within a statistical period (e.g., 1 hour or 1 day). Valid records must meet a credibility score. (Remove data with low credibility).

[0069] The unit then maps this uncertainty estimate to a confidence score between 0 and 1. The mapping function is monotonically decreasing, for example, using an exponentially decaying function: , This is an adjustment parameter. The greater the uncertainty, the closer the confidence score is to 0; the smaller the uncertainty, the closer the confidence score is to 1. This confidence score is tied to the indicator value, forming a "indicator value confidence score" tuple output.

[0070] The visual decision support interface serves as the terminal for interaction between the system and decision-makers. Please refer to the appendix. Figure 5 This interface receives the spatially aggregated evaluation results output by the evaluation result credibility quantification unit. The aggregation method involves dividing the urban geographical area into regular square grids, such as a 500m x 500m grid. For each grid, the interface receives the average or median of various service quality indicators statistically obtained within that grid, as well as the average credibility score corresponding to these indicators.

[0071] The interface employs advanced dual-layer overlay rendering technology to achieve intuitive information presentation. The bottom layer is a basic indicator layer. This layer uses a grid as the basic rendering unit, and fills the grid with a color selected from a preset continuous gradient based on the specific value of a certain indicator. For example, for the travel time indicator, a gradient from green to red is used, with green representing short travel time and smooth traffic, and red representing long travel time and severe congestion. For punctuality rate, a gradient from blue to yellow is used, with blue representing high punctuality rate and yellow representing low punctuality rate. The color of each grid intuitively reflects the service quality of that area.

[0072] The upper layer is a credibility mask layer. This layer is also rendered in grid units, but the rendered content is a semi-transparent grayscale mask. The shade of gray is determined by the average credibility score of the index within that grid. The mapping relationship is as follows: when the credibility score is 1, the mask is completely transparent, i.e., no gray is displayed; when the credibility score is 0, the mask is a completely opaque dark gray. Intermediate states use linear or non-linear interpolation.

[0073] Therefore, in high-confidence areas, the upper mask is almost transparent, and the colors of the underlying indicators are clearly visible; in low-confidence areas, the upper layer is covered by a dark gray semi-transparent mask, making the colors of the underlying indicators appear dull and blurry.

[0074] When using this interface, decision-makers can freely switch between different service quality indicator layers. More importantly, they can dynamically control the display intensity of the upper credibility mask layer using the transparency slider. When the slider is at its lowest setting, the mask is completely hidden, and the interface only displays the spatial distribution of service quality; when the slider is at its highest setting, the mask is fully visible, clearly showing which areas have low credibility and require careful consideration. This dual-layer overlay rendering mechanism allows decision-makers to simultaneously and adjustably perceive both "service quality" and "reliability of this judgment" on a single image, greatly enhancing the practical value and security of the assessment results in decision-making scenarios such as traffic planning, operation management, and public information services.

[0075] The entire system operates within a resilient distributed stream processing framework. High-precision spatiotemporal synchronization fusion units, dynamic weight adaptive modules, service quality index calculation engines, and evaluation result reliability quantification units are all encapsulated as independent stream processing topologies or jobs. They are connected via a highly reliable and scalable message queue service. Raw data is collected from edge devices, ingested into the system via the message queue, and flows sequentially through each of the aforementioned topologies. Each topology performs specific transformations, calculations, or enrichments on the data stream, ultimately writing the gridded evaluation results with reliability scores into a spatial database or cache for real-time querying and rendering by a visual decision support interface. This architecture ensures the system can scale horizontally to handle ever-increasing data volumes and guarantees low-latency output of evaluation results.

Claims

1. A transportation service quality assessment system based on multi-source data fusion, characterized in that, include: The high-precision spatiotemporal synchronization fusion unit is used to receive raw traffic data packets from multiple heterogeneous data sources, and to perform online calibration of the timestamp and unified mapping of the spatial location of each data packet to generate fused trajectory data with a unified spatiotemporal reference. The service quality index calculation engine is used to extract individual travel chains based on the fused trajectory data output by the high-precision spatiotemporal synchronous fusion unit, and calculate multi-dimensional service quality indicators including travel time, punctuality rate, travel reliability and comfort according to the preset evaluation model. The dynamic weight adaptive module is used to analyze the data quality characteristics of different data sources in a specific spatiotemporal environment in real time, and dynamically adjust the contribution weight of each data source in the spatiotemporal synchronization fusion process based on the quality characteristics. The evaluation result credibility quantification unit is used to perform uncertainty propagation calculation on each indicator calculated by the service quality indicator calculation engine, combined with the confidence information in the data fusion process and the weight information output by the dynamic weight adaptive module, and to add a quantified credibility score to each evaluation result. A visual decision support interface is used to render and overlay the evaluation results and their corresponding credibility scores, which are organized in a spatiotemporal grid.

2. The transportation service quality assessment system based on multi-source data fusion according to claim 1, characterized in that, The high-precision spatiotemporal synchronization fusion unit includes a data packet parsing subunit, a clock deviation estimation subunit, a timestamp calibration subunit, and a spatial coordinate transformation subunit. The data packet parsing subunit is used to extract the device identifier, original timestamp, original location coordinates, and system reference time of the data packet reception from the received raw data packet. The clock skew estimation subunit is used to maintain a dynamic clock skew model for each active data acquisition device. The process of establishing and updating the model is as follows: continuously acquire data packets sent by the same device at multiple times, and record the original timestamp and system reference time of each data packet. Calculate the time deviation of each data packet based on the system reference time; An exponentially weighted sliding window regression algorithm is used to fit the time deviation sequence to obtain the current deviation value and drift rate of the device clock relative to the system reference time. The timestamp calibration subunit is used to input the original timestamp extracted by the data packet parsing subunit into the clock deviation model of the corresponding device, and use the current deviation value and drift rate output by the model to perform compensation calculations to generate a calibrated accurate timestamp. The spatial coordinate transformation subunit is used to uniformly transform the original location coordinates extracted by the data packet parsing subunit to coordinates in the standard plane coordinate system adopted by the urban geographic information system through a preset coordinate transformation parameter matrix.

3. The transportation service quality assessment system based on multi-source data fusion according to claim 2, characterized in that, The service quality index calculation engine includes a travel chain reconstruction subunit and an index calculation subunit; The travel chain reconstruction subunit is used to sort the fused trajectory data points in time sequence according to the calibrated accurate timestamps, and apply a density-based spatial clustering algorithm to aggregate the sequence of location points that belong to the same traveler and are continuous in time into a travel trajectory. Furthermore, by combining the preset traffic network topology data, the trajectory point sequence is matched to specific road segments or public transportation routes through map matching algorithms, and the origin, destination, transfer points and mode of transportation used for each trip are identified. The indicator calculation subunit has multiple parallel calculation logics built in, including travel time calculation logic, punctuality rate calculation logic, travel reliability calculation logic, and comfort calculation logic.

4. The transportation service quality assessment system based on multi-source data fusion according to claim 3, characterized in that, The trip time calculation logic is used to calculate the actual trip time based on the precise timestamps of the origin and destination points recorded in the trip chain. The punctuality calculation logic is used to compare the actual travel time with the preset planned travel time or historical average travel time of the corresponding route to determine whether the trip is on time. The travel reliability calculation logic is used to calculate the coefficient of variation of travel time between a certain route or a certain origin-destination pair within a statistical period. The comfort calculation logic is used to analyze the derived parameters representing the motion state in the trajectory data, including the rate of change of acceleration (i.e., jerkiness) and the distribution characteristics of rest points, and to determine the model output comfort level through a preset threshold.

5. The transportation service quality assessment system based on multi-source data fusion according to claim 1, characterized in that, The dynamic weight adaptive module includes a data quality feature extraction subunit and a fusion weight decision subunit; The data quality feature extraction subunit is used to monitor multiple quality dimensions of each data source in real time. The quality dimensions include the stability of data reporting frequency, the accuracy dilution factor value of location coordinates, signal strength value, and the fitting residual of historical clock deviation model. It also periodically calculates a normalized quality score for each quality dimension of each data source. The fusion weight decision subunit is used to receive the multi-dimensional quality score output by the data quality feature extraction subunit. This subunit maintains a multi-input single-output adaptive neural network model. The input layer nodes of this model correspond to the scores of each quality dimension, and the output layer nodes are fusion weight values ​​between 0 and 1. For a real-time input quality score vector, this sub-unit calculates the dynamic fusion weight of the data source at the current moment through forward propagation.

6. The transportation service quality assessment system based on multi-source data fusion according to claim 5, characterized in that, When generating each fusion trajectory data point, the high-precision spatiotemporal synchronous fusion unit synchronously records the combination of data sources for that point and its corresponding dynamic fusion weight. The credibility quantification unit of the evaluation result performs uncertainty propagation calculation based on this information. This calculation process is based on the error propagation law, which transmits the calibration uncertainty of the underlying data points in time and space, as well as the uncertainty introduced by weight allocation when fusion of multi-source data, to the final service quality index through the functional relationship of the mathematical model used in the index calculation engine, and outputs a value between 0 and 1 as a credibility score.

7. A transportation service quality assessment system based on multi-source data fusion according to claim 6, characterized in that, In the uncertainty propagation calculation, for the travel time index, its uncertainty is obtained by calculating the square root of the sum of the squares of the uncertainty of the start timestamp and the squares of the uncertainty of the end timestamp.

8. The transportation service quality assessment system based on multi-source data fusion according to claim 1, characterized in that, The visual decision support interface divides the urban geospatial space into regular square grids and receives the average values ​​and credibility scores of various service quality indicators based on the grids as statistical units. The interface uses a dual-layer overlay rendering technology. The bottom layer uses a gradient color scheme from red to green to intuitively display the level of the indicator value, while the top layer uses a semi-transparent grayscale mask to represent the credibility score. The higher the score, the more transparent the mask, and the lower the score, the deeper the mask.

9. A transportation service quality assessment system based on multi-source data fusion according to claim 1, characterized in that, The system runs on a distributed stream processing framework; The high-precision spatiotemporal synchronization fusion unit, service quality index calculation engine, dynamic weight adaptive module, and evaluation result reliable quantification unit are deployed as independent processing topologies, and the data stream flows sequentially between the topologies through message queues.

10. A transportation service quality assessment system based on multi-source data fusion according to claim 1, characterized in that, The heterogeneous data sources include vehicle-mounted global satellite navigation system terminals, smartphone base station positioning data, and Bluetooth beacon data collected from transportation hubs.

Images