A bus route operation state cooperative estimation method for sparse data

Abstract

This invention discloses a collaborative estimation method for the operational status of bus routes using sparse data, relating to the field of intelligent transportation technology. Addressing the problems of sparsity, latency, drift, and modal misjudgment in bus GPS data, this invention projects GPS onto the route arc-length coordinate system to construct vehicle state vectors. Based on single-vehicle prediction and fleet chain topology, collaborative estimation is achieved through neighbor vehicle state alignment and consistency compensation. A hysteresis-based three-modal state machine with multi-source evidence is used to switch dynamic parameters and physical constraints under different operational scenarios. By using normalized innovation statistics and modal physical corridors, real disturbances and pseudo-jumps are distinguished and the state is corrected, outputting a continuous, stable, and interpretable bus operational status. This invention requires no additional sensors, has strong engineering adaptability, and can support applications such as bus scheduling and operation monitoring.

Description

Technical Field

[0001] This invention belongs to the technical fields of intelligent transportation, public transport operation monitoring, traffic cyber-physical system application and state estimation. Specifically, it relates to a method for coordinating the estimation of the continuous operating state of buses under conditions of sparse positioning data, delay, out-of-order and drift, based on the consistency compensation of the front and rear topology relationships of the bus fleet on the same route and in the same direction of operation, and combined with operation mode discrimination and anomaly interpretation. Background Technology

[0002] Multiple buses on urban bus routes operate continuously along the same path and in the same direction, naturally exhibiting an ordered queue relationship. The platform's inputs can be categorized into five types: spatial and facility inputs such as route centerline, station locations, and stop line locations; operational observation inputs such as GPS data; operational event inputs such as card swipe times; traffic control inputs such as signal phases; and time reference inputs used for sorting, alignment, and backtracking. Among these, the key difference in implementation lies in whether the time reference input uses a single or dual time reference. Most existing bus operation status perception methods still focus on individual vehicles, primarily relying on raw GPS points for trajectory interpolation, speed differentiation, or simple rule recognition, making it difficult to fully utilize the stable topological relationships between vehicles on the same route.

[0003] In real-world operating environments, bus GPS typically has sampling intervals of about ten seconds or even longer, and is often accompanied by problems such as delayed arrival, out-of-order uploads, short-term missing data, and drift / jumping points. Directly using such observations for speed estimation or state recognition can easily lead to phenomena such as trajectory breaks, sudden speed changes, reversed vehicle order, and misjudgments near intersections, thus affecting station dwell recognition, intersection queue recognition, vehicle merging diagnosis, and dispatching decisions.

[0004] Meanwhile, buses constantly switch between various operational scenarios during operation, such as driving, stopping at stations, and queuing at intersections. If the same set of fixed dynamic parameters, observation tolerances, and physical constraints are continuously used, it is easy to misinterpret positioning drift near the station as slow vehicle movement, or to misinterpret queuing creep as stationary movement, or to experience frequent switching at modal boundaries, leading to unstable state estimation.

[0005] Furthermore, traditional anomaly handling typically involves directly discarding anomalous observations when the observation residuals are too large, or simply compressing state changes after adding constraints. This approach fails to effectively distinguish between two fundamentally different situations: one is a genuine disturbance that actually occurred on-site, such as sudden braking, temporary yielding, or rapid queue compression; the other is a pseudo-abrupt change caused by GPS drift or jump points. If these two are not differentiated, the system will either excessively suppress genuine changes, leading to estimation lag, or it will follow pseudo-abrupt changes and output results that do not conform to physical laws.

[0006] From the perspective of transportation cyber-physical systems, bus operation monitoring and dispatch support is essentially a lightweight CPS (Common Power Systems) problem oriented towards route operation management: the physical side corresponds to the actual movement of buses on the route, including driving, entering stations, stopping, queuing, and being constrained by signals; the information side corresponds to the platform's digital reconstruction results of continuous vehicle position, route speed, operating mode, and abnormal states. While existing methods can process some GPS data, they often fail to integrate heterogeneous elements such as vehicles, routes, stations, stop lines, card swipe events, signal phases, and time references into a unified update chain, making it difficult to form a closed-loop processing mechanism from physical event acquisition to information-side state reconstruction and subsequent estimation feedback.

[0007] The shortcomings of existing methods are mainly threefold: first, they lack coordinated state correction based on the topological relationship between the preceding and following buses on the same route; second, they lack parameter switching and physical boundary switching driven by operational modes; and third, they lack anomaly interpretation mechanisms to distinguish between real disturbances and pseudo-sudden jumps. Therefore, based on the actual needs of the management platform, there is an urgent need for a method that can operate under the existing platform's available data conditions, simultaneously solving the problems of continuous state reconstruction of bus fleets, operational mode discrimination, and anomaly interpretation and diversion processing without relying on additional high-cost sensors, thus providing stable and interpretable state outputs for scheduling decisions and operational monitoring. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide a collaborative estimation method for the operational status of bus routes based on sparse data, in order to solve the problems of trajectory breakage, speed change, order reversal, and mode misjudgment caused by sparse, delayed, out-of-order, missing, and drifting points in existing bus status perception technologies. At the same time, it overcomes the shortcomings of traditional methods, such as not fully utilizing fleet topology association, unstable mode switching, and inability to distinguish between real disturbances and pseudo-changes in abnormal observations.

[0009] This invention maps bus fleets operating on the same route and in the same direction to a route arc-length coordinate system, constructing a chain-like ordered topology and implementing consistent collaborative compensation. Based on multi-source evidence and a state machine with hysteresis, it distinguishes between driving, station dwelling, and intersection queuing modes, switching dynamic parameters and physical constraints in real time according to the mode. Through modal physical corridors, it interpretably decentralizes abnormal observations, handling real disturbances and pseudo-jumps separately, ultimately outputting a continuous, stable, and physically consistent bus operation status. This invention requires no additional sensors, is compatible with single / dual time bases and late observation backtracking, has strong engineering adaptability, and can effectively support bus operation monitoring, intelligent scheduling, and traffic cyber-physical systems applications.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A collaborative estimation method for the operation status of bus routes based on sparse data includes the following steps:

[0012] S1. The platform side can determine the input of multiple buses on the same route and combine it with the vehicle number as the vehicle entity identifier;

[0013] S2. Project the GPS positioning data of each bus onto the center line of the route to obtain the corresponding arc length observation value of the route, and convert the two-dimensional latitude and longitude position of the vehicle into an arc length position that changes monotonically along the route;

[0014] S3. The vehicle's continuous state vector is constructed using the line arc length position and line speed. The vehicle state is initialized based on the first effective arc length observation, the difference between adjacent observations, or the default speed. At the same time, a buffer window and a unified internal update clock are established.

[0015] S4. At each internal time step, perform single-vehicle prior prediction based on the vehicle's posterior state at the previous time step to obtain the prior position and prior velocity at the current time step.

[0016] S5. Sort each vehicle in ascending order of its predicted arc length position at the current moment, construct an ordered chain topology of front vehicle-this vehicle-follower vehicle, and establish an adjacency matrix, degree matrix and Laplace matrix based on this adjacency relationship;

[0017] S6. When the vehicle has valid and new GPS positioning observations and meets the update gating conditions determined based on observation freshness and residual statistics, local correction is performed on the prior state using the vehicle's observations; otherwise, the prior state is retained and cooperative compensation is entered.

[0018] S7. Predict the most recent valid state of the adjacent vehicles to the current internal time based on the valid time, and obtain the aligned state of the adjacent vehicles. Perform consistency compensation on the state of the vehicle based on the front and rear topology relationship of the vehicle and the influence weight of the adjacent vehicles. When there are intersection signal constraints, gating or attenuating the adjacent edge weight that crosses the stop line.

[0019] S8. When the valid time corresponding to a late observation still falls within the buffer window, the system backtracks to the corresponding historical moment to perform observation correction, and then propagates forward along the complete estimation chain to the current moment to correct the current state output.

[0020] S9. Based on continuous state and environmental information, construct the station proximity degree, stop line proximity degree, front vehicle distance constraint, speed change, and observation residual evidence quantity, and use them as enhanced evidence when card swiping events or signal phases are available, establish a three-modal state machine with hysteresis, and determine whether the vehicle is currently in the driving mode, station dwell mode, or intersection queuing mode.

[0021] S10. Apply corresponding motion degree of freedom constraints based on the noise, observation noise and physical projection rules of the real-time switching process of the current operating mode;

[0022] S11. Calculate the current observation residual, residual covariance, and normalized innovation statistic; when the normalized innovation statistic exceeds the threshold corresponding to the current mode, trigger the anomaly interpretation process and construct a physical corridor around the current mode; based on whether the candidate update result falls within the physical corridor, determine the abnormal sample as a real disturbance or a pseudo-jump;

[0023] S12. For real disturbances, amplify process noise while maintaining the ability to track sudden changes; for pseudo-jumps, amplify observation noise, reduce observation update gain, and project the update results back to the physical feasible region allowed by the current mode.

[0024] S13. Output the vehicle's continuous line position, line speed, operating mode, anomaly explanation label, and related confidence information at the current moment, and write the results back to the cache as the basis for the next internal moment and subsequent late observation backtracking.

[0025] Furthermore, in step S1, the platform side can determine the following inputs: the center line, station location and stop line location in the space and facility inputs, the GPS positioning data in the operation observation inputs, the card swiping time in the business event inputs, the signal phase in the traffic control inputs, and the time field in the time reference inputs;

[0026] The time field adopts a single time reference mode or a dual time reference mode to complete the sorting, alignment and backtracking of observations. In the single time reference mode, the system configures a valid time field representing the time when the state belongs to each observation. In the dual time reference mode, the system saves the observation record time and the platform reception time at the same time. The observation record time is used to determine the internal time when the observation participates in sorting, prediction, update and backtracking, and the platform reception time is used to estimate link hysteresis, configure the buffer window and determine the backtracking priority.

[0027] Furthermore, in step S3, the vehicle continuous state vector is represented as:

[0028]

[0029] In the formula, Indicates the first A bus at discrete moments Continuous state; Indicates the first The car at any time The position of the arc length of the line; Indicates the first The car at any time The line speed.

[0030] Furthermore, in step S4, the single-vehicle prior prediction model is expressed as:

[0031]

[0032] In the formula, Let A be the prior state of the i-th vehicle at time k+1; A is the state transition matrix. For internal update step size; This is a process disturbance.

[0033] Furthermore, in step S5, the ordered chain topology adopts a non-negative edge weight design based on the headway distance and introduces a signal gating factor; if there is a stop line between the current vehicle and the neighboring vehicle and the signal phase is a no-entry constraint, the edge weight across the stop line is attenuated or set to zero.

[0034] Furthermore, in step S6, the local correction of the vehicle is achieved based on the following observation model and correction relationship:

[0035]

[0036]

[0037] In the formula, This indicates the vehicle status after partial correction; This is the observation value at the current location; To observe noise; The observation matrix; This is the result of partial correction for this vehicle; To observe the effective indication quantity; This is the Kalman gain, i.e., the correction weight.

[0038] Furthermore, in step S7, the consistency compensation term is represented as follows:

[0039]

[0040]

[0041] In the formula, For consistency gain; This is a group of vehicles adjacent to this vehicle; Let be the adjacent edge weight, representing the influence weight of neighboring vehicle j on vehicle i; This represents the state of the j-th adjacent vehicle at the current time k after alignment; Indicates the first The difference in steps between the most recent valid state of each neighboring vehicle and the current time; A is the state transition matrix; This indicates the most recent valid state of the neighboring vehicle j.

[0042] Furthermore, in step S11, the physical corridor Defined as the physically feasible region of position and velocity in the current mode:

[0043]

[0044] In the formula, Used to define the physical feasible region of position and velocity in the current mode; Let be the state vector, where Indicates the position of the arc length of the line. Indicates the line's operating speed; Indicates the lower bound of the position; Indicates the upper bound of the position; Indicates the lower bound of velocity; Indicates the upper bound of velocity.

[0045] Furthermore, in step S11, the innovation statistics are normalized. The degree of conflict between observation and prediction is characterized by the following formula:

[0046]

[0047]

[0048]

[0049] In the formula, This is a normalized information statistic used to characterize the degree of conflict between observation and prediction; Represents the observation residual; Represents residual covariance; Let be the prior state covariance matrix, representing the uncertainty of the predicted state; This represents the transpose of the observation matrix; To observe the noise covariance, the subscript... This indicates that the noise varies with mode.

[0050] Beneficial effects:

[0051] 1. This invention explicitly transforms the natural front-to-back queue relationship of buses on the same route and in the same direction into a chain-like topological cooperative constraint, upgrading the state estimation of a single vehicle to the cooperative estimation of the entire fleet. Even under conditions of GPS sparsity, short-term missing data, and delayed out-of-order conditions, structural correction can still be achieved through neighbor vehicle consistency compensation, which greatly improves the stability and robustness of continuous trajectory reconstruction.

[0052] 2. This invention is based on a three-modal state machine of driving, station dwelling, and intersection queuing, which realizes real-time adaptive switching of process noise, observation noise, and physical constraints, effectively suppressing problems such as platform drift, frequent jitter of modal boundaries, and misjudgment of queuing crossing, making the state output more consistent with the real bus operation scenario.

[0053] 3. This invention upgrades the anomaly handling from "directly eliminating excessive residuals" to a residual triggering + physical corridor interpretation + branch correction mechanism, which can accurately distinguish between real traffic disturbances and GPS pseudo-jumps. While maintaining the ability to track real events such as sudden braking and queue compression, it suppresses non-physical jumps caused by positioning drift, taking into account both estimation sensitivity and physical rationality.

[0054] 4. This invention is compatible with single / dual time bases, supports backtracking of late observations and forward repropagation, and can make full use of out-of-order and delayed arrival of effective data. It does not require additional high-precision sensors and can be deployed directly based on existing platform data. It has strong engineering adaptability and low implementation cost.

[0055] 5. This invention forms a complete closed loop of input organization - state collaborative estimation - modality discrimination - anomaly interpretation - result output, and unifies and integrates heterogeneous information such as routes, stations, GPS, card swiping, and signal phase. The output state is continuous, the modality is discriminable, and the anomaly is interpretable, which can directly support business applications such as intelligent bus scheduling, operation monitoring, and vehicle-road-cloud collaboration.

[0056] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the architecture of a collaborative estimation method for the operation status of bus routes based on sparse data according to the present invention.

[0058] Figure 2 This is a schematic diagram of the specific process under the three-tiered progressive overall architecture;

[0059] Figure 3 A schematic diagram of the continuous state collaborative estimation subprocess for bus routes;

[0060] Figure 4 This is a schematic diagram of the sub-process for modal discrimination and real-time parameter switching.

[0061] Figure 5 This is a schematic diagram of the subprocesses for anomaly interpretation and physical corridor constraints. Detailed Implementation

[0062] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of this application.

[0063] like Figure 1 As shown, the overall processing architecture of this invention includes a complete closed-loop process such as input organization, continuous state collaborative estimation, operational mode discrimination, anomaly interpretation and diversion, and result output write-back. This collaborative estimation method for the operational state of bus routes for sparse data specifically includes the following steps:

[0064] Step 1: Obtain the platform-side input for multiple buses on the same route, and combine it with the vehicle number as the vehicle entity identifier.

[0065] At the data organization level, the platform-side inputs include five categories, among which the identifiable inputs include at least: the centerline of the line, station location, and stop line location in the spatial and facility inputs; GPS positioning data in the operational observation inputs; and the time field in the time reference inputs. When relevant information is available on the platform side, it may further include card swiping time in the business event inputs and signal phase in the traffic control inputs. The time field adopts a single time reference mode or a dual time reference mode to complete the sorting, alignment, and backtracking processing of observations.

[0066] At the time organization level, the time reference input is divided into a single time reference mode and a dual time reference mode. In the single time reference mode, the system configures a valid time field for each observation to represent the time at which its state belongs. In the dual time reference mode, the system simultaneously saves the observation recording time and the platform receiving time. The observation recording time is used to determine which internal time period the observation should participate in for sorting, prediction, updating, and backtracking, while the platform receiving time is used to estimate link latency, configure the buffer window, and determine the backtracking priority.

[0067] For ease of standardized processing, the first... vehicle number The effective time for each observation is The internal update starting point is Fixed internal step size Therefore, the internal history index to which an observation belongs can be obtained from the effective time mapping; in the dual time reference mode, the link hysteresis can be further defined to describe the time difference between the arrival of an observation and the occurrence of the observation.

[0068] Single time base:

[0069] Dual time base:

[0070] Historical Index:

[0071] The system operates at a fixed step size. Perform prediction, update, modality determination, and anomaly routing. For late observations within the cache window, perform corrections according to their validity time back to the corresponding historical index before forward repropagation, so that historical corrections can be passed to the current output state.

[0072] Step 2: Project the bus GPS positioning data onto the center line of the route to obtain the observed value of the route arc length.

[0073] In this step, the GPS positioning data of each bus is projected onto the center line of the route to obtain the corresponding arc length observation value, so that the spatial position of the vehicle is converted from a two-dimensional latitude and longitude representation to an arc length position representation that changes monotonically along the route.

[0074] By first projecting GPS points onto the centerline of the route and then using arc length observations in the update process, lateral drift at locations such as bus stops, parallel auxiliary roads, and intersection guide lanes can be uniformly converted into observation errors along the route direction. This avoids the problem of "lateral drift being misinterpreted as forward or backward movement along the route" when two-dimensional coordinates are directly used in the estimation. This method is particularly suitable for areas near bus stops, parallel road sections, and adjacent areas between upper and lower levels of elevated roads.

[0075] Step 3: Construct the vehicle's continuous state vector and complete the state initialization.

[0076] like Figure 2 As shown, in the overall process, continuous state cooperative estimation occurs after arc lengthening and before modal discrimination; among which, as Figure 3 As shown, the continuous state collaborative estimation subprocess takes arc-length observations, vehicle sorting relationships, and a unified internal clock as inputs. At each internal time step, it continuously recovers the vehicle's line position and line speed through local vehicle observation updates, neighbor vehicle state alignment, and consistency compensation, and supports backtracking and forward repropagation of late observations.

[0077] In real-world public transportation scenarios, vehicles may be in various states, such as traveling within a section, stopping at stations, being blocked at intersections, experiencing short-term delays or data loss, or experiencing data loss due to delays. However, these states are continuous in time, constrained by the geometry of the route in space, and restricted by the order of preceding and following vehicles at the fleet level. Therefore, the following state definitions, prediction formulas, topology constructions, and compensation rules are not isolated theoretical derivations, but rather transform the laws governing public transportation operations into a continuously updated state mechanism that can be computed online.

[0078] In terms of the overall process chain, this sub-process is responsible for mapping the evolution of the physical side's "continuous movement of buses along the route" into a computable state chain on the information side, serving as the state foundation for subsequent modal discrimination and anomaly interpretation. It focuses not only on the geometric position of a single GPS point at a given moment, but also on the topological order of the multi-vehicle queue, the temporal relationships of observed arrivals, and the joint constraints between the spatial boundaries of the route.

[0079] For the A bus at discrete moments The continuous state is defined as:

[0080]

[0081] in, Indicates the first The car at any time The position of the arc length of the line, Indicates the first The car at any time The line speed.

[0082] Step 4: Perform single-vehicle prior prediction based on the posterior state of the previous time step.

[0083] In the basic implementation, the single-vehicle prior prediction model is written as follows:

[0084]

[0085] in, This represents the current prior prediction state. This represents a process perturbation. The model states that within an internal step, the vehicle's position advances by "current position plus velocity multiplied by time step," and the velocity continues from the previous state when no additional information is available. Although this model is low-dimensional, it exhibits good real-time performance and engineering feasibility in sparse GPS scenarios with a time interval of ten seconds.

[0086] In a typical implementation scenario, the platform-side GPS updates only once every ten seconds, while the internal state update clock can run continuously with shorter step sizes. In this case, the prediction model is used to fill the state gaps between two observations, enabling the system to continuously output the continuous arc length position and line speed along the route. The line speed is understood as the equivalent running speed along the centerline of the route, rather than the original two-dimensional planar speed, and is therefore more suitable for describing the longitudinal movement of a vehicle approaching a station, leaving a station, or approaching a stop line.

[0087] Step 5: Construct an ordered chain topology according to the predicted arc length position.

[0088] At each internal time step, the system sorts the vehicles in ascending order of their predicted arc length positions. For the middle vehicle, the vehicle in front of it and the vehicle behind it are defined as its neighbors; for the first and last vehicles, only one neighbor is retained. A weighted adjacency matrix is ​​constructed based on this adjacency relationship. Degree matrix And the Laplace matrix. To avoid inconsistencies in edge weight signs due to differences in the directions of adjacent vehicles, a non-negative design based on the absolute value of relative vehicle distance or equivalent time distance is adopted for adjacent edge weights. Furthermore, a signal gating factor can be introduced to attenuate the adjacency relationship across stop lines.

[0089]

[0090] To incorporate the semantics of public transport operations into the topological coupling, a boundary weight design based on the headway is adopted:

[0091]

[0092]

[0093] in, Indicates the first Vehicle and the The equivalent headway between two adjacent vehicles Used to prevent denominator degradation at low speeds or when the machine is stopped. To prevent small positive numbers from diverging in value, This is an optional signal gating factor. If there is a stop line between the current vehicle and the adjacent vehicle, and the signal phase indicates that the stop line still imposes a no-entry constraint on this vehicle, the gating factor can attenuate or set the weight of the vehicle crossing the stop line to zero, in order to avoid unreasonable cross-line coupling in red light queuing scenarios.

[0094] The aforementioned edge weight design can enhance the coupling effect between vehicles with short time intervals and suppress unreasonable traction of vehicles with long distances or vehicles crossing stop lines through time interval attenuation and signal gating.

[0095] Step 6: When the vehicle has available observations at the current moment, perform local corrections for the vehicle.

[0096] By projecting the GPS points onto the centerline of the route, the observed values ​​of the vehicle in the arc length coordinate system can be obtained. Its observation model is written as:

[0097]

[0098] in, To account for observation noise. The above model shows that, in the basic implementation, GPS only directly observes position and not velocity; therefore, velocity needs to be estimated indirectly through prediction, observation updates, and subsequent collaborative compensation.

[0099] When available observations are available for this vehicle at the current moment, perform local corrections for this vehicle:

[0100]

[0101] in, This is the result of partial correction for this vehicle. For observations that are valid indicators; when the current observation is valid, not expired, and meets the update gating conditions, ,otherwise Therefore, when observations are missing, expired, or do not meet the current update conditions, this invention allows the system to skip the vehicle-specific correction and maintain continuous state propagation.

[0102] In this invention, the observation gating conditions, process noise parameters, observation noise parameters, and physical projection rules for the current internal time are preferably initially selected based on the operating mode determined in the previous internal time. After completing the local correction of the vehicle, the arrival time alignment of neighboring vehicles, and the consistency compensation for the current time, the operating mode is refreshed by combining the evidence amount of the current time, and parameter correction, physical projection, and anomaly interpretation are completed accordingly. This avoids a circular dependency between operating mode discrimination, observation gating, and anomaly diversion.

[0103] Step 7: Align the states of neighboring vehicles upon arrival and perform consistency coordination compensation.

[0104] Since neighboring vehicle observations may not arrive at the same time as the current vehicle's observations, the system first predicts the most recent valid state of the neighboring vehicle to the current internal time, thus obtaining the aligned neighboring vehicle state:

[0105]

[0106] in, Indicates the first The step difference between the most recent valid state of each neighboring vehicle and the current time. After aligning the neighboring vehicles, perform consistency coordination compensation:

[0107]

[0108] in, This is a consistency gain. The compensation is not a simple averaging, but rather utilizes the stable preceding and following relationships between adjacent vehicles on the same route to structurally correct the vehicle's state. When the vehicle's GPS is reliable, the consistency term only makes a slight correction; when the vehicle's GPS is missing or drifting, the consistency term can bring the vehicle's state back to a region consistent with the preceding and following vehicles in terms of sequence and reasonable spacing.

[0109] For example, when a vehicle in the middle experiences a brief loss of speed while the vehicles in front and behind continue to report, the consistency compensation can use the convoy's front and rear order and reasonable spacing constraints to maintain the target vehicle within a reliable operating range. When the convoy's spacing expands or contracts due to signal release or passenger boarding and alighting, the compensation will not forcibly pull each vehicle to a fixed equal spacing, but will allow the real spacing changes to continue to be transmitted while maintaining a reasonable order.

[0110] Step 8: Perform backtracking correction and forward repropagation on data that arrives late or out of order.

[0111] For late observations, if their validity period corresponds to a historical index If the observation is still within the cache window, the system will perform observation correction at the historical time point:

[0112]

[0113] Subsequently, from the historical index The process begins with forward repropagation following the complete flow of "prior prediction—topology construction—neighbor vehicle alignment—consistency compensation—modal discrimination—anomaly diversion" to the current time step by step. Through this mechanism, late observations are no longer considered outdated and useless data, but can substantially correct the state output at the current time step.

[0114] This backtracking and repropagation mechanism is suitable for scenarios involving network congestion, terminal retransmission, or platform link jitter. For critical observations that arrive late but still fall within the buffer window, backtracking correction can prevent the state from remaining in the old value for a long time and pass the correction result to the current output.

[0115] In one embodiment, the aforementioned internal update step size is preferably smaller than the average reporting interval of the platform-side GPS, and the buffer window length is set according to the maximum allowable delay time of the observation; the entry threshold, exit threshold, entry radius, exit radius, and maximum queuing creep speed of the station dwell mode and the intersection queuing mode can be obtained offline based on the historical operation data of the target line, the geometric scale of the station, the accuracy of the stop line position, and the observation noise level of the platform side; wherein, the exit threshold is preferably wider than the entry threshold, so as to form hysteresis and suppress mode boundary jitter.

[0116] Step 9: Construct modal discrimination evidence and determine the operating mode through a state machine with hysteresis.

[0117] like Figure 4 As shown, this step constructs modal evidence based on spatial geometry, operational observations, business events, and traffic control information. It then drives the real-time switching of estimator parameters, observation gating thresholds, and physical projection rules through a three-modal state machine with hysteresis, enabling business semantics to directly participate in the continuous estimation process.

[0118] During bus operation, station dwell and intersection queuing can both manifest as low speed or even stationary states, making it difficult to distinguish based on instantaneous speed alone. Furthermore, under low-frequency GPS conditions, individual location points often drift near platforms or creep before red lights. Therefore, the following modal discrimination and parameter switching rules employ a combination of multi-evidence joint analysis, continuous counting, and hysteresis constraints to improve the distinguishability and switching stability between different business scenarios.

[0119] From the perspective of a multi-level closed loop, this sub-process is located between continuous state collaborative estimation and anomaly interpretation decentralization, serving as an intermediate link in the transmission of business semantics to parameter sets and physical constraint rules. Through this link, the system output is no longer just a geometric result of "where the car is," but further forms an executable discrimination result of "what kind of operating scenario the car is currently in, and what kind of estimation tolerance and boundary rules should be adopted."

[0120] In this invention, the operational modes are divided into driving mode, station dwell mode, and intersection queuing mode. To distinguish these three modes under existing platform data conditions, evidence quantities such as station proximity, stop line proximity, queue compression, speed variation, observation conflict, card swipe events, and signal phase constraints are constructed.

[0121]

[0122]

[0123]

[0124]

[0125]

[0126] in, and Inherit the continuous states output by the collaborative estimation layer; This represents the arc length coordinates of the current candidate site. Indicates the coordinates of the arc length of the stop line ahead. Indicates the position of the continuous arc length of the preceding vehicle. This represents the difference between the current observation and the prior prediction in the observation space. The aforementioned amounts of evidence respectively reflect the proximity of the platform, the proximity of the stop line, the degree of queue compression, the degree of speed change, and the degree of conflict between observation and prediction. For card swiping time and signal phase, this invention further organizes them into business event evidence and traffic control evidence to enhance the distinguishability of station dwell time and intersection queuing.

[0127]

[0128]

[0129]

[0130]

[0131] in, This represents the set of card-swiping times for vehicle i. The set of signal phases that indicate prohibition or the need to maintain a queue. , , as well as , , The weights are non-negative. Therefore, the card swipe time, as a business event input, participates in station dwell determination, and the signal phase, as a traffic control input, participates in intersection queuing determination, thus forming a multi-evidence basis for determination together with location, speed, and platoon relationship.

[0132] To avoid frequent jitter near platform boundaries or queue release boundaries, this invention introduces a continuous counting and hysteresis mechanism. For the station dwell mode, only when the vehicle satisfies the comprehensive dwell evidence score within a consecutive number of internal steps... The vehicle will only switch to the queuing mode if the vehicle's speed does not fall below the entry threshold and simultaneously meets the conditions that the distance to the station does not exceed the entry radius and the vehicle speed does not exceed the entry threshold. When exiting the queuing mode, a wider exit radius and a higher exit speed threshold are used. For the intersection queuing mode, the vehicle will only switch to the queuing mode if it meets the comprehensive queuing evidence score within a certain number of consecutive internal steps. The vehicle will only switch to queuing mode when it is not lower than the entry threshold and is simultaneously constrained by the stop line, the preceding vehicle queue, or the signal phase; when exiting the queuing mode, the exit threshold is different from the entry threshold.

[0133] Therefore, the current discrete mode is denoted as:

[0134]

[0135] By separating entry and exit thresholds, the system can effectively eliminate boundary jitter caused by low-frequency GPS conditions. or Frequent jumps ensure the stability of subsequent parameter switching and smooth updates.

[0136] Step 10: Switch parameters in real time according to the current mode and perform physical projection.

[0137] After obtaining the current mode, the system selects the process noise matrix and observation noise parameters corresponding to the mode for the vehicle. The driving mode, dwelling mode, and queuing mode correspond to the following process noises:

[0138]

[0139] Among them, the parameter settings satisfy < < 1, and < < 1 indicates that the stationary mode has the smallest degree of freedom of motion, followed by the queuing mode, and the driving mode has the largest degree of freedom.

[0140] The above parameter relationships correspond to the bus service scenarios: the noise is lowest in the stationary mode, followed by the queuing mode, and highest in the driving mode, to adapt to stable stopping, low-speed crawling, and normal operation, respectively.

[0141] Correspondingly, the observation noise is taken as:

[0142]

[0143] Among them, usually satisfying 1. > 1, and usually have > This means that in the stationary mode, the system tends to treat positioning drift near the platform as observation noise; in the queuing mode, the system still retains a certain low-speed tracking capability, but will not easily follow observations that cross the line.

[0144] Given the current parameter set, first calculate the covariance of the observed residuals:

[0145]

[0146] Then, observation gating is performed based on the residual magnitude.

[0147]

[0148] If the current observation is deemed suitable for local vehicle updates, it is considered acceptable; otherwise, it is treated as an abnormal or outdated observation and not directly entered into local correction. Therefore, this invention does not perform modality identification first and then output modality labels separately, but rather applies the modality judgment results to the estimator parameters and update rules in real time.

[0149] In this invention, after a local update, a physical projection consistent with the current mode is performed to obtain the final posterior state. For the stationary mode, the vehicle speed is projected to zero, and the position is restricted to within the effective station docking window. For the queuing mode, the speed is allowed to be between zero and the maximum queuing creep speed, but the vehicle position is not allowed to exceed the stop line safety margin or the minimum reasonable distance behind the preceding vehicle if the stop line constraint or the preceding vehicle constraint is not released. For the driving mode, the above-mentioned station or stop line projection is not applied, and only the normal constraints given by the track speed and track geometry are retained.

[0150] By using modal-related physical projection, the system can impose different degrees of freedom constraints on different business scenarios within the same continuous state space, thereby significantly reducing drift and forward movement in station parking conditions and reducing line-crossing estimation in queuing conditions.

[0151] The projection process can be understood as a final screening of the update results for business rationality. For example, a position in a station stop state should not cross the stop window without evidence of departure; a position in a red light queue state should not cross the stop line before the prohibition constraint is lifted; and in normal driving state, it should not be compressed into an excessively narrow area due to a single isolated abnormal observation.

[0152] Step 11: Calculate the residual statistics and construct the physical corridor corresponding to the current mode.

[0153] like Figure 5 As shown, when the residual statistic exceeds the threshold corresponding to the current mode, this step constructs a physical corridor and, based on whether the candidate update result falls into the corridor, divides the abnormal samples into two processing branches: real disturbance preservation or pseudo-jump suppression. The purpose is to suppress pseudo-jumps caused by positioning jumps, projection errors, or time mismatches while maintaining the ability to track real disturbances.

[0154] In bus operation scenarios, phenomena such as sudden braking, temporary yielding, and sudden compression of the queue ahead can cause significant deviations between observation and prediction. However, such deviations are part of real traffic events. Conversely, when GPS momentarily drifts to a parallel road, outside the platform, or on the other side of the stop line, it may also produce large residuals, but these should not be considered as real motion. Therefore, this subprocess adopts a two-stage mechanism of "residual triggering + physical corridor interpretation" to distinguish between real disturbances and pseudo-jumps.

[0155] In cyber-physical mapping, this sub-process can be understood as a stage where the information side provides a secondary interpretation of sudden events and measurement anomalies on the physical side. By unifying track geometry, platform windows, stop line no-entry boundaries, and preceding vehicle constraints into a unified physical corridor, this invention organizes multiple heterogeneous boundary conditions into the same judgment framework, thereby avoiding overly conservative or misguided approaches caused by relying solely on statistical thresholds for anomaly removal.

[0156] In this invention, anomaly handling does not involve directly deleting observations when residuals are too large; instead, it first triggers a physical interpretation based on residual statistics. For the first... A bus at the time The observations are defined as follows: residuals, residual covariance, and normalized innovation statistics:

[0157]

[0158]

[0159]

[0160] And define the trigger threshold corresponding to the current mode, denoted as . Typically, the threshold is set to the strictest for the stationary mode, followed by the queuing mode, and the most lenient for the driving mode. This setting is consistent with the bus operation scenario: the stationary state should be the most stable, so it is easier to trigger anomaly checks; the driving state allows for normal dynamic changes, so the threshold can be appropriately relaxed.

[0161] To determine whether a large residual corresponds to a genuine disturbance or a pseudo-jump, this invention constructs a physical corridor around the current position that is consistent with the current mode. First, based on the current base state and allowable acceleration / deceleration capabilities, the free motion boundary is given:

[0162]

[0163]

[0164] Based on this, if the current mode is stationary, the position boundary is constrained to intersect with the platform window, and the speed boundary is tightened to zero; if the current mode is queuing, the upper boundary position is further limited to the safety margin in front of the stop line or the minimum reasonable distance behind the preceding vehicle; if the current mode is driving, the free movement boundary is used as the main constraint.

[0165] For queuing mode, the upper bound of the queuing position can be written as:

[0166]

[0167] Therefore, the physical corridor at the current moment is uniformly defined as:

[0168]

[0169] The aforementioned physical corridor unifies the geometric boundaries of the line, the platform stopping boundaries, the stop line constraints, the preceding vehicle constraints, and the modal-dependent velocity upper bound into a single feasible domain expression, providing a basis for subsequent anomaly interpretation.

[0170] The physical corridor is not an abstract geometric region, but an engineering expression of "where a vehicle might go and what speed it might take in the current scenario." It unifies the centerline of the route, the platform stop window, the stop line no-entry boundary, and the minimum reasonable distance between vehicles in front into the same judgment framework, so that the interpretation of anomalies no longer depends solely on the magnitude of statistical quantities, but can be attributed to the actual operating conditions of public transportation.

[0171] Step 12: Perform branch corrections for real disturbances and pseudo-jumps respectively.

[0172] When the normalized innovation statistic exceeds the trigger threshold, candidate update states are first calculated, and then interpretation and routing are performed in conjunction with the physical corridor. After obtaining the prior state and the current observation, the system first performs a candidate local update:

[0173]

[0174]

[0175] Subsequently, combining residual statistics and physical corridors, the interpretation category of the current sample is defined:

[0176]

[0177] in, This indicates a normal sample. This represents a real disturbance. This indicates a false jump. The key to this judgment logic is that a large residual only indicates a conflict between the observation and the model; however, whether the candidate state is still within the physical corridor allowed by the current mode determines whether the conflict originates from changes in the actual field or measurement anomalies.

[0178] Therefore, when the candidate update result deviates significantly from the prior but still falls within the corridor, it usually indicates that a real event such as sudden deceleration, temporary yielding, or queue compression has indeed occurred on site, and it is more reasonable to continue tracking in this case; conversely, if the candidate result directly crosses the stop line, jumps out of the platform window, or intrudes into the unreasonable distance of the preceding vehicle, it is more consistent with the characteristics of measurement anomaly, time mismatch, or projection distortion.

[0179] When a sample is identified as a genuine disturbance, the system does not treat it as a bad pixel, but instead increases the process noise to enhance tracking capability. The corresponding amplification factor can be written as:

[0180]

[0181] This yields the process noise under the actual perturbation branch. And calculate the actual perturbation branch state. This processing allows the system to more quickly approximate the actual situation when real-world changes occur, such as sudden braking, temporary yielding, and rapid queue compression, avoiding significant estimation lag.

[0182] When a sample is identified as a spurious jump, the system increases the observation noise and decreases the observation update gain. The corresponding amplification factor can be written as:

[0183]

[0184] This yields the observation noise under the pseudo-jump branch. and received a temporary update status. Then, the system projects this temporary state back onto the physical corridor:

[0185]

[0186] in, This represents the projection operator over the physical corridor. Through a combination of "reducing the impact of observations + projecting back into the feasible region," this invention can effectively suppress GPS jump points that do not conform to the relationship between the line, platform, stop line, or preceding vehicle.

[0187] By employing the aforementioned dual-branch processing, the system avoids both excessive conservatism that smooths out all real disturbances and blindly following single observations that could produce non-physical trajectories. For public transport operation monitoring, this mechanism helps to simultaneously improve the stability of arrival judgment, queue identification, fleet compression early warning, and anomaly location diagnosis.

[0188] Finally, the system outputs the posterior state based on the interpretation category:

[0189]

[0190] It also outputs anomaly explanation labels, which are divided into three categories: normal, real disturbance maintenance, and pseudo-sudden jump suppression, for direct use by upper-level monitoring, scheduling, or analysis modules.

[0191] Step 13: Output the collaborative estimation results of the bus operation status of the route and write them back to the cache.

[0192] In actual operation, this invention is executed cyclically in the following order: First, GPS access, time alignment, map matching, and arc length conversion are completed; second, at the current internal time, single-vehicle prior prediction, local correction of the vehicle, arrival time alignment of neighboring vehicles, and consistency collaborative compensation are performed to obtain a continuous state; third, evidence quantity is constructed based on the continuous state and environmental information, and the current operating mode is determined by a three-modal state machine with hysteresis, while selecting the corresponding parameter set and physical projection rules; subsequently, residuals are triggered for detection based on the current mode, physical corridors are constructed, and abnormal samples are diverted into a true disturbance preservation branch or a pseudo-spurious jump suppression branch; finally, the continuous line position, line speed, operating mode, and abnormal interpretation label at the current time are output, and the final result is written to a cache for use in the next internal time and for backtracking of late observations.

[0193] In single-time-base mode, this invention uses a single valid time field to complete observation attribution, sorting, alignment, and backtracking, and achieves continuous state estimation and modality discrimination based on unified spatial and facility inputs, operational observation inputs, business event inputs, and traffic control inputs. In dual-time-base mode, it further distinguishes between observation recording time and platform reception time, making late observation processing, cache scheduling, and observation freshness assessment more refined. Thus, this invention maintains a stable input structure while balancing the simplicity of implementation in general platform deployment with estimation accuracy under refined time processing conditions.

[0194] In one optional output mode, in addition to outputting the vehicle's continuous arc length position, line speed, and current mode, the system can also simultaneously output an observation freshness indicator, anomaly explanation label, whether backtracking has been triggered, the type of the currently applicable parameter set, and confidence information. These outputs can be directly displayed on the dispatch monitoring platform or used as input for subsequent fleet balancing control, arrival analysis, and operational diagnostic modules.

[0195] In practical applications, the above outputs can be directly used for operation monitoring, scheduling assistance, service evaluation, arrival analysis, or subsequent upper-level vehicle-road-cloud collaborative services, thus forming a reusable state interface for the application plane. However, specific vehicle control execution, scheduling decision generation, or wide-area resource scheduling are not essential components for the existence of this invention. Therefore, this invention maintains a degree of decoupling from the execution control plane while reserving standardized extension interfaces for subsequent system evolution.

[0196] In summary, this invention no longer views bus operation status identification as a simple labeling and classification of individual GPS points. Instead, it incorporates three interconnected core improvements: First, under a unified input structure and single / dual time reference support, continuous state recovery is achieved through fleet chain topology collaborative estimation and late observation backpropagation. Second, business semantics are directly injected into the continuous estimation process through modality-driven parameter switching and physical projection. Third, interpretable diversion of real disturbances and pseudo-spurious jumps is achieved through residual triggering, physical corridor interpretation, and branch correction. Therefore, this invention can output continuous, stable, interpretable operation status results that conform to the business semantics of public transportation.

[0197] It is hereby declared that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A collaborative estimation method for the operational status of bus routes based on sparse data, characterized in that, Includes the following steps: S1. The platform side can determine the input of multiple buses on the same route and combine it with the vehicle number as the vehicle entity identifier; S2. Project the GPS positioning data of each bus onto the center line of the route to obtain the corresponding arc length observation value of the route, and convert the two-dimensional latitude and longitude position of the vehicle into an arc length position that changes monotonically along the route; S3. The vehicle's continuous state vector is constructed using the line arc length position and line speed. The vehicle state is initialized based on the first effective arc length observation, the difference between adjacent observations, or the default speed. At the same time, a buffer window and a unified internal update clock are established. S4. At each internal time step, perform single-vehicle prior prediction based on the vehicle's posterior state at the previous time step to obtain the prior position and prior speed at the current time step. S5. Sort each vehicle in ascending order of its predicted arc length position at the current moment, construct an ordered chain topology of front vehicle-this vehicle-follower vehicle, and establish an adjacency matrix, degree matrix and Laplace matrix based on this adjacency relationship; S6. When the vehicle has valid and new GPS positioning observations and meets the update gating conditions determined based on observation freshness and residual statistics, local correction is performed on the prior state using the vehicle's observations; otherwise, the prior state is retained and cooperative compensation is entered. S7. Predict the most recent valid state of the adjacent vehicles to the current internal time based on the valid time, and obtain the aligned state of the adjacent vehicles. Perform consistency compensation on the state of the vehicle based on the front and rear topology relationship of the vehicle and the influence weight of the adjacent vehicles. When there are intersection signal constraints, gating or attenuating the adjacent edge weight that crosses the stop line. S8. When the valid time corresponding to a late observation still falls within the buffer window, the system backtracks to the corresponding historical moment to perform observation correction, and then propagates forward along the complete estimation chain to the current moment to correct the current state output. S9. Based on continuous state and environmental information, construct the station proximity degree, stop line proximity degree, front vehicle distance constraint, speed change, and observation residual evidence quantity, and use them as enhanced evidence when card swiping events or signal phases are available, establish a three-modal state machine with hysteresis, and determine whether the vehicle is currently in the driving mode, station dwell mode, or intersection queuing mode. S10. Apply corresponding motion degree of freedom constraints based on the noise, observation noise and physical projection rules of the real-time switching process of the current operating mode; S11. Calculate the current observation residuals, residual covariance, and normalized innovation statistic; When the normalized innovation statistic exceeds the threshold corresponding to the current mode, the anomaly interpretation process is triggered, and a physical corridor is constructed around the current mode; based on whether the candidate update result falls within the physical corridor, the abnormal sample is determined to be a real perturbation or a pseudo-jump. S12. For real disturbances, amplify process noise while maintaining the ability to track sudden changes; for pseudo-jumps, amplify observation noise, reduce observation update gain, and project the update results back to the physical feasible region allowed by the current mode. S13. Output the vehicle's continuous line position, line speed, operating mode, anomaly explanation label, and related confidence information at the current moment, and write the results back to the cache as the basis for the next internal moment and subsequent late observation backtracking.

2. The collaborative estimation method for bus route operation status based on sparse data according to claim 1, characterized in that, In step S1, the platform side can determine the following inputs: the center line, station location and stop line location in the space and facility inputs, the GPS positioning data in the operation observation inputs, the card swiping time in the business event inputs, the signal phase in the traffic control inputs, and the time field in the time reference inputs. The time field adopts a single time reference mode or a dual time reference mode to complete the sorting, alignment and backtracking of observations. In the single time reference mode, the system configures a valid time field representing the time when the state belongs to each observation. In the dual time reference mode, the system saves the observation record time and the platform reception time at the same time. The observation record time is used to determine the internal time when the observation participates in sorting, prediction, update and backtracking, and the platform reception time is used to estimate link hysteresis, configure the buffer window and determine the backtracking priority.

3. The collaborative estimation method for bus route operation status based on sparse data according to claim 2, characterized in that, In step S3, the vehicle's continuous state vector is represented as follows: In the formula, Indicates the first A bus at discrete moments Continuous state; Indicates the first The car at any time The position of the arc length of the line; Indicates the first The car at any time The line speed.

4. The collaborative estimation method for bus route operation status based on sparse data according to claim 3, characterized in that, In step S4, the prior prediction model for a single vehicle is expressed as follows: In the formula, Let be the prior state of the i-th vehicle at time k+1; A is the state transition matrix; For internal update step size; This is a process disturbance.

5. The collaborative estimation method for the operation status of bus routes based on sparse data according to claim 4, characterized in that, In step S5, the ordered chain topology adopts a non-negative edge weight design based on the headway of the vehicle and introduces a signal gating factor; if there is a stop line between the current vehicle and the neighboring vehicle and the signal phase is a no-entry constraint, the edge weight across the stop line is attenuated or set to zero.

6. The collaborative estimation method for bus route operation status based on sparse data according to claim 5, characterized in that, In step S6, the local correction of the vehicle is achieved based on the following observation model and correction relationship: , In the formula, This indicates the vehicle status after partial correction; This is the observation value at the current location; To observe noise; The observation matrix; This is the result of partial correction for this vehicle; To observe the effective indication quantity; This is the Kalman gain, i.e., the correction weight.

7. The collaborative estimation method for bus route operation status based on sparse data according to claim 6, characterized in that, In step S7, the consistency compensation term is represented as follows: , In the formula, For consistency gain; This is a group of vehicles adjacent to this vehicle; Let be the adjacent edge weight, representing the influence weight of neighboring vehicle j on vehicle i; This represents the state of the j-th adjacent vehicle at the current time k after alignment; Indicates the first The difference in steps between the most recent valid state of each neighboring vehicle and the current time. A is the state transition matrix; This indicates the most recent valid state of the neighboring vehicle j.

8. The collaborative estimation method for the operation status of bus routes based on sparse data according to claim 7, characterized in that, In step S11, the physical corridor Defined as the physically feasible region of position and velocity in the current mode: In the formula, Used to define the physical feasible region of position and velocity in the current mode; Let be the state vector, where Indicates the position of the arc length of the line. Indicates the line's operating speed; Indicates the lower bound of the position; Indicates the upper bound of the position; Indicates the lower bound of velocity; Indicates the upper bound of velocity.

9. A collaborative estimation method for the operation status of bus routes based on sparse data, as described in claim 8, is characterized in that... In step S11, the normalized innovation statistics are used. The degree of conflict between observation and prediction is characterized by the following formula: , , In the formula, This is a normalized information statistic used to characterize the degree of conflict between observation and prediction; Represents the observation residual; Represents residual covariance; Let be the prior state covariance matrix, representing the uncertainty of the predicted state; This represents the transpose of the observation matrix; To observe the noise covariance, the subscript... This indicates that the noise varies with mode.

Images