Vehicle travel path dynamic detection system and method based on deep learning

Abstract

The invention provides a vehicle travel path dynamic detection system and method based on deep learning. The method comprises the steps of obtaining a station sequence of a scheduling task, a time window corresponding to a station and a task number; inputting the rule condition representation vector into a pre-constructed track generation network, and generating a probable track sequence meeting scheduling rule requirements; obtaining an actual vehicle track, and extracting offset features between the actual vehicle track and the affordable track sequence and a track abnormal point index set; obtaining a reference vehicle set of the same line of the current task, constructing a vehicle behavior graph in combination with a target vehicle, and performing group consistency analysis based on the vehicle behavior graph to obtain a collaborative migration score; and judging whether the vehicle behavior is abnormal or not in combination with the offset feature, the collaborative offset score and the group acceptable abnormal label set. According to the invention, the travel path monitoring accuracy and availability in a travel transportation scene are significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of vehicle movement dynamic detection, and particularly relates to a vehicle movement dynamic detection system and method based on deep learning. Background Technology

[0002] The operational rules of tourism transportation, chartered bus services, and passenger dispatch scenarios are highly constrained and structurally complex. Vehicles must adhere to predetermined station sequences, route directions, and time periods. However, objective factors such as sudden road conditions, temporary traffic control, and navigation detours can lead to reasonable deviations, which are superficially similar to subjective anomalies such as driver detours, missed stops, and arbitrary stops. Existing methods primarily rely on static rule matching and trajectory playback, depending on electronic fences or fixed threshold comparisons. These methods struggle to accommodate acceptable deviations during dynamic operations and fail to generate meaningful reference trajectories based on the fine-grained differences in dispatch rules. This results in rigid judgment standards and a susceptibility to false alarms and missed alarms. Furthermore, fleet operations are common in tourism transportation. A single-vehicle perspective cannot reveal the common behaviors of a group on the same route during the same time period, making it difficult to distinguish between widespread detours caused by road incidents and individual violations. Even with the introduction of learning methods, if dispatch rules are not explicitly integrated into the modeling, if quantitative indicators of deviation related to station reach and local deviation density are not constructed, and if consistency criteria are not provided at the group level, stable engineering implementation remains challenging.

[0003] Based on this, a systematic approach is needed that can structure scheduling rules into learnable conditions, generate expected trajectories that conform to task intent, characterize the differences in actual trajectories in the spatiotemporal dimension, verify the consistency of deviation behavior at the fleet level, and finally complete the anomaly judgment with a fusion strategy. This approach can simultaneously solve three key challenges from a mechanism perspective: compliance expression, deviation quantification, and group rationality identification. Summary of the Invention

[0004] The purpose of this invention is to propose a vehicle movement dynamic detection system and method based on deep learning to solve the above-mentioned problems.

[0005] To achieve the above objectives, a deep learning-based vehicle movement dynamic detection method is provided in a first aspect of the present invention, the method comprising: Obtain the station sequence, time window corresponding to the station, and task number of the scheduling task; map the station sequence to a station spatial projection vector; map the time window corresponding to the station to a station time vector; map the task number to a fixed-length vector through a lookup table; and combine the spatial projection vector, station time vector, and fixed-length vector rule conditions to represent the vector. The rule condition representation vector is input into a pre-constructed trajectory generation network to generate a desired trajectory sequence that meets the scheduling rule requirements; wherein, the trajectory generation network includes a rule vector embedding module and a trajectory sequence generation module; The actual vehicle trajectory is obtained, and the offset features between the actual vehicle trajectory and the expected trajectory sequence, as well as the trajectory anomaly point index set, are extracted. The offset features include average offset, maximum offset, station coverage loss, and burst offset density. Obtain a set of reference vehicles on the same route as the current task, construct a vehicle behavior map by combining it with the target vehicle, perform group consistency analysis based on the vehicle behavior map to obtain a cooperative offset score; determine whether the anomaly occurs within the time window of the group offset set based on the cooperative offset score and the trajectory anomaly point index set, so as to generate a set of acceptable anomaly labels for the group. By combining the aforementioned offset features, collaborative offset scores, and the set of acceptable abnormal labels for the group, it is determined whether the vehicle behavior is abnormal.

[0006] Furthermore, the location information of each station in the station sequence represents longitude and latitude, and each station is a station location specified for scheduling; the time window corresponding to the station includes start and end times.

[0007] Furthermore, the rule vector embedding module is an embedding network consisting of three layers of linear transformations, each layer of which contains a linear mapping and a hyperbolic tangent activation function, and the output is a rule control vector; The trajectory sequence generation module consists of two layers of gated loop units, each with a hidden state of 64 dimensions, and outputs the spatial coordinates of each trajectory point.

[0008] Furthermore, the trajectory generation network also includes a trajectory rule consistency regularization term; the trajectory rule consistency regularization term generates the time index of the first point in the trajectory that is closest to the scheduled station location based on the rule control vector, spatial coordinates, and the center time of the time window of the station location specified by the schedule. Based on the trajectory rule consistency regularization term, the spatial coordinates of each trajectory point are regularized to obtain the expected trajectory sequence that meets the scheduling rule requirements.

[0009] Furthermore, each trajectory of the actual vehicle trajectory is the two-dimensional geographic coordinates of the vehicle at each moment; Then, the average offset represents the average value of the two-dimensional spatial difference vector between the actual trajectory and the expected trajectory; the maximum offset represents the maximum value of the two-dimensional spatial difference vector between the actual trajectory and the expected trajectory; the station coverage loss is calculated based on the spatial distance between the vehicle and the station location specified by the dispatcher, and is used to represent the risk level of the vehicle failing to reasonably approach the station specified by the dispatcher in the task; the sudden offset density is calculated based on the two-dimensional spatial difference vector between the future position and the future expected position, and is used to represent the maximum degree of high offset of the vehicle in a continuous period of time.

[0010] Furthermore, the set of trajectory anomaly point indices is a set of points whose two-dimensional Euclidean distance between the two-dimensional spatial difference vectors between the actual position and the expected position is greater than the offset anomaly judgment threshold set by the system.

[0011] Furthermore, the edges of the vehicle behavior graph are the edge weights between vehicles, i.e., the behavior similarity. The results of the group consistency analysis are generated based on the cosine similarity of the feature vectors between vehicles, and are used to measure their directional consistency. The higher the score of the group consistency analysis, the greater the difference in the deviation behavior of the target vehicle from other vehicles, and the more likely it is to be an independent anomaly.

[0012] Furthermore, the set of acceptable abnormal labels for the group is as follows: If the abnormal time period of the target vehicle is in the set of trajectory abnormal point indexes and also appears in the abnormal time windows of most reference vehicles, then the abnormality is marked as acceptable to the group and a set of acceptable abnormality labels for the group is constructed; otherwise, it is suspected of being an individual violation.

[0013] Furthermore, the step of combining the offset features, collaborative offset score, and the set of acceptable anomaly labels for the group to determine whether the vehicle behavior is abnormal specifically includes: An anomaly score is calculated by weighted fusion based on the average offset, maximum offset, site coverage loss, burst offset density, and the proportion of time in anomaly time slices not exempted by the group; wherein, the proportion of time in anomaly time slices not exempted by the group is used to strengthen the penalty for individual anomalies. The final anomaly label is obtained by comparing the anomaly score with the system-defined threshold. If the abnormal score is greater than the system's set threshold, it is judged as abnormal; otherwise, it is judged as normal.

[0014] A second aspect of the invention provides a deep learning-based vehicle movement dynamic detection system, the system comprising: The rule condition construction module is used to obtain the station sequence, the time window corresponding to the station, and the task number of the scheduling task. The station sequence is mapped to a station spatial projection vector, the time window corresponding to the station is mapped to a station time vector, and the task number is mapped to a fixed-length vector through a lookup table. The spatial projection vector, the station time vector, and the fixed-length vector rule condition representation vector are combined. The expected trajectory generation module is used to input the rule condition representation vector into a pre-constructed trajectory generation network to generate an expected trajectory sequence that meets the scheduling rule requirements; wherein, the trajectory generation network includes a rule vector embedding module and a trajectory sequence generation module; The offset feature extraction module is used to obtain the actual trajectory of the vehicle, extract the offset features between the actual trajectory and the expected trajectory sequence, and the set of trajectory anomaly point indices. The offset features include average offset, maximum offset, station coverage loss, and burst offset density. The group behavior analysis module is used to obtain a set of reference vehicles on the same route as the current task, construct a vehicle behavior map in combination with the target vehicle, perform group consistency analysis based on the vehicle behavior map, and obtain a cooperative offset score; based on the cooperative offset score and the set of trajectory anomaly point indexes, it is determined whether the anomaly occurs within the time window of the group offset set, so as to generate a set of acceptable anomaly labels for the group. The anomaly detection module is used to determine whether vehicle behavior is abnormal by combining the offset features, collaborative offset scores, and the set of acceptable anomaly labels for the group.

[0015] The beneficial technical effects of the present invention are at least as follows: This invention focuses on the rule constraints and fleet coordination behavior of tourism transportation tasks, proposing a dynamic vehicle path detection system and method driven by rule conditions. It generates expected trajectories by mapping scheduling rules to unified condition vectors, constructs offset features oriented towards station reach and local offset density, and establishes a group consistency metric based on feature similarity within the same route and batch of tasks. Finally, it uses a fusion score to determine abnormal behavior, thus forming a closed-loop mechanism from rule reference, difference quantification, group verification to decision output. This scheme avoids the rigidity of pure threshold comparison at the expected trajectory level, introduces indicators oriented towards station compliance and short-term clustered offsets at the offset quantification level to improve interpretability, identifies general detours and individual anomalies at the group level using consistency scores, and introduces time-limited exemption information from the group at the decision level to reduce false alarms.

[0016] The entire method uses scheduling data and vehicle trajectories as the sole input source, has an achievable engineering path, and can robustly identify behaviors such as detours, missed stops, and illegal parking in actual operation. It also provides clear exemptions for reasonable deviations caused by road incidents, significantly improving the accuracy and usability of behavior monitoring in tourism transportation scenarios. Attached Figure Description

[0017] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0018] Figure 1 This is a flowchart of the vehicle movement dynamic detection method based on deep learning according to the present invention. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0020] like Figure 1 As shown in the embodiment of the present invention, a vehicle movement dynamic detection method based on deep learning is provided, the method comprising: S1. Obtain the station sequence, the time window corresponding to the station, and the task number of the scheduling task. Map the station sequence to a station spatial projection vector, map the time window corresponding to the station to a station time vector, and map the task number to a fixed-length vector through a lookup table. Combine the spatial projection vector, the station time vector, and the fixed-length vector rule condition representation vector.

[0021] Specifically, this step aims to transform the rule information relied upon by scheduling tasks in tourism transportation scenarios into mathematical vectors with fixed dimensions, serving as input conditions for subsequent deep learning models to generate expected trajectories. Since tourism passenger transport tasks often have clearly defined routes, stop order, time limits, and task numbers before departure, this step structures and maps this information, making it learnable and inputtable, thereby improving the trajectory generation model's ability to express the task context.

[0022] Furthermore, the input to this step is the task rule data provided by the scheduling system, including the following three categories: station sequence, denoted as... Each of them This represents the stations that the vehicle must pass through in sequence. The latitude and longitude coordinates are generated by the route configuration table of the operation system and are a two-dimensional real number vector; the time window corresponding to each station is denoted as... This indicates the allowed arrival time range for each station, in minutes; this data is automatically configured by the shift template system; the task number is denoted as... A unique identifier is assigned to each scheduling task to distinguish different tasks; the departure time is recorded as... , used to standardize the time window range.

[0023] To unify the rule structure across different tasks, this step first performs numerical mapping on the site location information. Each site... The mapping process for representing its longitude and latitude is as follows: ; in, , These are the affine transformation matrices, , For bias terms, This represents a piecewise linear activation function, and its output is... Let be the spatial projection vector of the station. To incorporate station order information, the relative position of the station in the entire sequence is defined as . ,in This represents the total number of stations. and The structure of each site is obtained after splicing. The dimension is three-dimensional, obtained by stitching together multiple sites. To ensure consistent output dimensions for each task, the system defines a maximum number of stations. If there are not enough sites Then, the vector is filled with all zeros, and the valid position information is recorded for subsequent masking operations.

[0024] Furthermore, the time window portion is processed using a time difference discretization method. Each time window... Compared to Calculate the time difference and record it as follows: and The unit is minutes. This difference is discretized at a fixed time granularity (e.g., 15 minutes), mapped to an integer number, and then transformed into a fixed-dimensional vector representation using a lookup table, denoted as . , indicating the site The time-constrained embedding vector. The time vectors for all sites are denoted as... Also, uniformly supplement to the length .

[0025] Task Number By looking up a table, it is mapped to a fixed-length vector, denoted as... This is used to enhance the uniqueness of task-level representation and avoid confusion between tasks.

[0026] Finally, the rule condition vector It is composed of a site structure vector, a time window structure vector, and a task number vector, and the calculation formula is as follows: ; in This is a vector representing the rule conditions, with a fixed structure, capable of expressing path information, timing constraints, and task identifiers in a scheduling task, with dimensions of [dimensionality missing]. .

[0027] S2. Input the rule condition representation vector into the pre-constructed trajectory generation network to generate a desired trajectory sequence that meets the scheduling rule requirements; wherein, the trajectory generation network includes a rule vector embedding module and a trajectory sequence generation module.

[0028] Specifically, the goal of this step is to represent the vector of rule conditions. Based on this, a "desired trajectory" that meets the requirements of the current tourism transportation task is generated. This serves as a reference benchmark for subsequent actual trajectory deviation detection. Unlike trajectory prediction problems in general scenarios, this patent targets the scenario of tourist charter buses or fixed-route passenger transport. This type of task has two significant characteristics: (1) the trajectory must be forced or approximately obey scheduling rules, such as station order and time period requirements, and has the nature of "accurate trajectory determination"; (2) in some scenarios, the constraint relationship between vehicle trajectory and station time window cannot be ignored, especially involving fixed station shuttle or batch pick-up and drop-off tasks. Therefore, it is insufficient to generate path points alone. It is necessary to design a generation method that combines spatial point sequence and time window constraints to generate the trajectory. It can serve as a standard trajectory for "how it should operate," providing a reasonable and quantifiable reference for subsequent deviation identification.

[0029] The input for this step is the rule condition representation vector generated in the previous step. The vector consists of three parts: site structure representation Time window embedded sequence Task number embedded These variables all come from the dispatching platform system and are information that is determined before the vehicle performs its transportation task, requiring no additional collection or prediction.

[0030] Furthermore, to simulate the desired driving path under real-world conditions to the greatest extent possible, this step employs a trajectory generation network with a rule-based control mechanism. The overall structure includes: a rule vector embedding module. A trajectory sequence generation module In addition to a trajectory-rule consistency regularization design, the innovation of the entire system is mainly reflected in the generation control method, which not only captures the evolution of trajectory space, but also explicitly models the rule conditions and applies deviation penalties.

[0031] First, the rule condition vector Input an embedding network (regular vector embedding module) consisting of three layers of linear transformations. Each layer contains a linear mapping and a hyperbolic tangent activation function, and the output is a regular control vector. This vector controls the initial state of the generative model and participates in the generation of trajectory points as an auxiliary input in each decoding step.

[0032] Trajectory sequence generation module It consists of two layers of gated recurrent units (GRUs), each with a 64-dimensional hidden state, and outputs the spatial coordinates of each trajectory point. The generation process is as follows: ; in For the trajectory points of the previous step, This is the previous hidden state. and These are the initial affine transformation parameters. In this structure... As a control input throughout the process, it ensures consistency between the generated trajectory and the task rules.

[0033] To enhance the strong binding between generated trajectories and rules, this step proposes an innovative trajectory rule consistency regularization term, constructed on the site set. Based on this, it is defined as: ; in, For generating the first trajectory One point, To schedule the specified station location, For the site The central moment of the time window The first station location specified by the schedule in the generated trajectory. The time index of the most recent point. To control the timing error penalty weight, this regularization term has two functions: (1) it forces the generated trajectory to be close to the specified station; (2) it constrains the time distribution of the trajectory to the station to not be far from the center of the preset time window, so as to ensure that the generated trajectory meets the task expectation in both spatial and temporal dimensions.

[0034] The aforementioned regularization term is used in conjunction with the basic trajectory fitting loss during the training phase to improve the scheduling fit of the trajectory generation model. During the deployment phase, the trajectory generation process no longer relies on historical trajectory data, but only on rule-based conditions. It features lightweight design and high controllability. The system sets the trajectory length based on the average task duration and sampling frequency. For example, sampling every 10 seconds for 20 minutes. Finally, the trajectory is generated. , as the output of the expected trajectory.

[0035] The output of this step is: This refers to the expected trajectory sequence generated under the scheduling task, which is used for subsequent deviation feature extraction between the actual trajectory and the expected trajectory.

[0036] S3. Obtain the actual trajectory of the vehicle, extract the offset features between the actual trajectory and the expected trajectory sequence, and the set of trajectory anomaly point indexes. The offset features include average offset, maximum offset, station coverage loss, and burst offset density.

[0037] Specifically, this step is responsible for the core function of quantifying the difference between the actual trajectory and the expected trajectory in the system. Its goal is to extract feature vectors describing the vehicle's deviation behavior for subsequent anomaly detection. Unlike traditional offset calculations based on spatial distance, this step incorporates the business characteristics of tourism transportation tasks, specifically introducing innovative features such as station arrival status and trajectory segment offset density, establishing a trajectory offset expression method that more closely reflects the operational scenario. This method not only considers the overall error of spatial location but also introduces modeling for two types of high-risk behaviors in the task: "stations must be reached" and "abnormal deviation within a certain time period," thereby improving the robustness and interpretability of anomaly detection.

[0038] Furthermore, the input for this step includes two parts: (1) the actual driving trajectory of the vehicle. Each of them For vehicles at any time The two-dimensional geographic coordinates are derived from the vehicle's onboard BeiDou / GPS device, with a sampling frequency of once every 10 seconds and a total sampling length of [missing information]. Determined by the duration of the task; (2) the expected trajectory generated in the second step Each This indicates the reference position that the vehicle should be in under the given conditions, and the corresponding time. Each trajectory corresponds one-to-one with the actual trajectory. The two trajectories are perfectly aligned on the time axis, making them comparable point-to-point.

[0039] In addition, the site set obtained in step one is also introduced. Each scheduling specifies the site location. The coordinates of the stations that must be reached are fixedly set by the scheduling system when configuring the task; and a station determination threshold is set. This indicates that a vehicle is considered to have arrived at the station when the distance between it and the station is less than this threshold, which is usually set to 50 meters.

[0040] Furthermore, trajectory difference analysis begins with the overall spatial error, defining the average offset distance between the actual trajectory and the expected trajectory. With maximum offset distance : ; in This represents the two-dimensional spatial difference vector between the actual position and the expected position. This represents the two-dimensional Euclidean distance. To measure the severity of the vehicle's overall deviation from its path. This reflects whether there has been a sudden and drastic shift. This measurement method is based solely on spatial differences and does not consider the "site reach" behavior unique to tourism transportation.

[0041] To address this, this step proposes an innovative "penalty for insufficient site coverage". This is used to indicate the degree of risk of a vehicle failing to reasonably approach the designated dispatch station during a task, and it is defined as follows: ; in Indicates the vehicle's time With the specified site location for scheduling Spatial distance, Let Gaussian similarity function be used. This is a distance attenuation control parameter, typically taken as half of the area the station is allowed to reach during the task. The formula calculates this for each scheduled station location. In the trajectory The system finds the closest point and converts the distance into an "arrival probability" using a Gaussian mapping. If a station is never approached, this value approaches 1, increasing the overall loss. This term can be added as structural loss during training or used for behavior scoring during inference.

[0042] Simultaneously considering the issue of local offset density in the trajectory, a density index for sudden offset segments in the trajectory is defined. This represents the maximum degree of high deviation that a vehicle experiences within a continuous time period, expressed as: ; in To adjust the sliding window size (e.g.) (corresponding to 1 minute), continuous calculation by sliding along the entire trajectory. The average offset within a point is calculated, and the maximum value is taken. This indicator is mainly used to detect behavioral segments that do not conform to the intended scheduling path, such as detours, U-turns, and illegal temporary stops.

[0043] In summary, this step outputs two core variables: (1) trajectory offset feature vector Its dimensions are fixed and include , , and (2) Track anomaly point index set ,in The offset anomaly judgment threshold set for the system (e.g., 100 meters) is used for downstream abnormal behavior labeling and alarms.

[0044] This solution fully considers the high sensitivity of tourism transportation operations to "whether it has reached the station" and "whether it requires a detour," and introduces two highly business-interpretive innovative quantities on the basis of traditional spatial error calculation: one is a differentiable station reach scoring function. First, it compensates for hard judgment errors; second, it uses sliding window offset density calculation. It is used to identify discontinuous but high-risk anomalous behaviors. All variables have clear sources and complete structures. The algorithm is practical and has a closed-loop logic, providing accurate input for subsequent system judgments.

[0045] S4. Obtain the reference vehicle set of the same route for the current task, construct a vehicle behavior map in combination with the target vehicle, perform group consistency analysis based on the vehicle behavior map to obtain a cooperative offset score; determine whether the anomaly occurs within the time window of the group offset set based on the cooperative offset score and the trajectory anomaly point index set, so as to generate a group acceptable anomaly label set.

[0046] Specifically, the core objective of this step is to utilize the group synchronous operation characteristics of vehicles in tourism transportation tasks to determine whether the current vehicle trajectory deviation behavior exhibits group consistency, thereby assisting in the deviation feature vector analysis in the previous step. To achieve more robust and interpretable anomaly identification, this method considers not only whether the vehicle itself deviates from its designated path, but also whether the deviation is a "general phenomenon" or an "individual anomaly." In tourism scheduling tasks, multiple vehicles often depart along the same route and within the same time period (such as a group of buses transporting tourists). These vehicles should typically have highly similar spatial paths and operating rhythms. Therefore, if one vehicle deviates from its designated path but other vehicles exhibit the same deviation characteristics, the system should avoid misjudging it as an anomaly. Conversely, if most vehicles have consistent trajectory characteristics, but the target vehicle deviates significantly, it can be determined as an individual anomaly. This step provides this crucial basis for collaborative judgment.

[0047] Furthermore, the input is the trajectory offset feature vector output from step three. and the set of trajectory anomaly point indices .in, It is a four-dimensional vector containing: average offset Maximum offset Losses due to site coverage and sudden shift density Its specific meaning and origin have been defined in detail in step three. This represents the set of time indices marked as offset anomalies in the trajectory. Additionally, the set of reference vehicles for the current task must also be entered. These vehicles operate on the same route as the target vehicle, with similar departure times. Their trajectories are all provided by the dispatching system, and the trajectory acquisition methods are identical. Each reference vehicle has a corresponding offset feature vector. .

[0048] Furthermore, to construct a collaborative behavior judgment mechanism, the system designed a graph-based offset feature similarity modeling process. First, a vehicle behavior graph is constructed. ,in Includes target vehicle With all reference vehicles, The edge weights represent the behavioral similarity between vehicles. To control the number of variables and enhance interpretability, the edge weights... Instead of relying solely on Euclidean distance, a distance structure based on feature weight recalibration is introduced: ; in Indicates the first 3D features (such as) ) in the The value on the vehicle Indicates the first 3D features (such as) The value of ) on the j-th car, These are manually set feature importance coefficients used to highlight dimensions that are more important in tourism transportation (such as station reach and sudden shifts), rather than average weighting. For example, you can set... , indicating a sudden shift and site miss items It has a higher weight in comparisons of group behavior.

[0049] After constructing the graph structure, the target vehicle is then designed. Outlier scoring function relative to overall group behavior The score is not a simple calculation of the difference from the mean, but rather introduces a "reverse consistency index" as an innovation scoring item: ; in Cosine similarity is used to represent the similarity between two feature vectors, measuring their directional consistency. Compared to Euclidean distance, this metric focuses more on "whether behavioral trends are similar" rather than absolute numerical similarity. The advantage of this is that even if the degree of offset differs, as long as the offset type (e.g., both are site omissions) is consistent, they can be identified as similar behaviors, thus avoiding false positives. Final score. The higher the value, the greater the difference in deviation behavior between the target vehicle and other vehicles, and the more likely it is to be an independent anomaly.

[0050] Furthermore, to further enhance the deployability and behavioral interpretability of this module, the system will integrate the collaborative scoring results with information on abnormal time periods. Combined, a "time period consistency criterion" is constructed, which determines whether the anomaly occurs within the concentrated time window of the group offset. If the target vehicle's anomaly occurs during a specific time period... If an anomaly occurs within the same time window as most reference vehicles, it is marked as "group acceptable"; otherwise, it is marked as "suspected individual violation". Only logical labels are output here; no formula is needed. The judgment strategy is controlled by the set time window overlap threshold.

[0051] This step ultimately outputs two variables: (1) Cooperative offset score (2) Set of acceptable anomaly labels for the group, for use in subsequent joint anomaly determination; ,in For reference vehicles The set of abnormal time periods is automatically calculated from their trajectories.

[0052] S5. Combining the aforementioned offset features, collaborative offset scores, and the set of acceptable abnormal labels for the group, determine whether the vehicle behavior is abnormal.

[0053] Specifically, this step is used to make the final determination of whether the vehicle exhibits abnormal trajectory behavior. Unlike traditional static threshold judgment, this step comprehensively considers trajectory deviation characteristics. Coherence score By constructing a fusion scoring function, vehicle trajectory deviation behavior is quantitatively scored, and the anomaly is determined based on the score. Additionally, a set of group exemption periods is introduced. and abnormal trajectory time period set This allows for the identification and exemption of acceptable deviations in group behavior, thereby enhancing the overall system's intelligent judgment capabilities and practical applicability.

[0054] Furthermore, this step constructs a weighted fusion scoring function that quantifies all input variables into anomaly scores. The formula is as follows: ; in, to These are the weighting coefficients for each feature term, which can be determined during model training or rule setting; The percentage of time in an aberration time slice that was not exempted from group exemption is used to strengthen the penalty for individual aberrations, and is defined as follows: ; The above This indicates the portion of the abnormal trajectory not covered by group behavior, reflecting whether the target vehicle "deviated independently." For example: If and ,but This indicates that two of the three anomalies are individual shifts.

[0055] After obtaining the score Then, set a threshold with the system. The comparison is performed to obtain the final anomaly label. : like If so, it is judged as abnormal. ; Otherwise, it is considered normal. .

[0056] A second aspect of the invention provides a deep learning-based vehicle movement dynamic detection system, the system comprising: The rule condition construction module is used to obtain the station sequence, the time window corresponding to the station, and the task number of the scheduling task. The station sequence is mapped to a station spatial projection vector, the time window corresponding to the station is mapped to a station time vector, and the task number is mapped to a fixed-length vector through a lookup table. The spatial projection vector, the station time vector, and the fixed-length vector rule condition representation vector are combined. The expected trajectory generation module is used to input the rule condition representation vector into a pre-constructed trajectory generation network to generate an expected trajectory sequence that meets the scheduling rule requirements; wherein, the trajectory generation network includes a rule vector embedding module and a trajectory sequence generation module; The offset feature extraction module is used to obtain the actual trajectory of the vehicle, extract the offset features between the actual trajectory and the expected trajectory sequence, and the set of trajectory anomaly point indices. The offset features include average offset, maximum offset, station coverage loss, and burst offset density. The group behavior analysis module is used to obtain a set of reference vehicles on the same route as the current task, construct a vehicle behavior map in combination with the target vehicle, perform group consistency analysis based on the vehicle behavior map, and obtain a cooperative offset score; based on the cooperative offset score and the set of trajectory anomaly point indexes, it is determined whether the anomaly occurs within the time window of the group offset set, so as to generate a set of acceptable anomaly labels for the group. The anomaly detection module is used to determine whether vehicle behavior is abnormal by combining the offset features, collaborative offset scores, and the set of acceptable anomaly labels for the group.

[0057] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0058] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or units may be electrical, mechanical, or other forms.

[0059] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0060] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for detecting vehicle behavior dynamics based on deep learning, characterized in that, The method comprises: obtaining a site sequence of a scheduling task, a time window corresponding to the site, and a task number, mapping the site sequence into a site space projection vector, mapping the time window corresponding to the site into a site time vector, mapping the task number into a fixed-length vector through a lookup table, and combining the space projection vector, the site time vector, and the fixed-length vector to represent a rule condition vector; inputting the rule condition vector into a pre-constructed trajectory generation network to generate a necessary trajectory sequence that meets the scheduling rule requirements; wherein the trajectory generation network comprises a rule vector embedding module and a trajectory sequence generation module; obtaining an actual trajectory of a vehicle, extracting an offset feature between the actual trajectory of the vehicle and the necessary trajectory sequence and an index set of trajectory abnormal points, the offset feature comprising an average offset, a maximum offset, a site non-coverage loss, and a sudden offset density; obtaining a reference vehicle set of the same line of the current task, combining a target vehicle to construct a vehicle behavior graph, performing group consistency analysis based on the vehicle behavior graph to obtain a collaborative offset score, and judging whether an abnormality occurs in a group offset set time window based on the collaborative offset score and the index set of trajectory abnormal points to generate a group acceptable abnormality label set; combining the offset feature, the collaborative offset score, and the group acceptable abnormality label set to determine whether the vehicle behavior is abnormal. 2.The deep learning-based vehicle behavior dynamic detection method of claim 1, wherein, Each site position information in the site sequence represents longitude and latitude, and each site is a site position specified by scheduling; the time window corresponding to the site comprises start and end times. 3.The deep learning-based vehicle behavior dynamic detection method of claim 1, wherein, The rule vector embedding module is an embedding network composed of three linear transformations, each layer of which contains a linear mapping and a hyperbolic tangent activation function, and the output is a rule control vector; The trajectory sequence generation module is composed of two layers of gated recurrent units, each layer of which has a 64-dimensional hidden state, and outputs the spatial coordinates of each trajectory point. 4.The deep learning-based vehicle behavior dynamic detection method of claim 3, wherein, The trajectory generation network further comprises a trajectory rule consistency regularization term; the trajectory rule consistency regularization term is based on the rule control vector, the spatial coordinates, and the time window center moment of the site position specified by scheduling, and generates a time index at which the first point in the generated trajectory closest to the site position specified by scheduling appears; The spatial coordinates of each trajectory point are regularized and constrained based on the trajectory rule consistency regularization term to obtain the necessary trajectory sequence that meets the scheduling rule requirements. 5.The deep learning-based vehicle behavior dynamic detection method of claim 1, wherein, Each trajectory of the actual trajectory of the vehicle is a two-dimensional geographic coordinate of the vehicle at each moment; The average offset represents the average value of the two-dimensional space difference vector between the actual trajectory and the necessary trajectory; the maximum offset represents the maximum value of the two-dimensional space difference vector between the actual trajectory and the necessary trajectory; The site non-coverage loss is calculated based on the spatial distance between the vehicle and the site position specified by scheduling, and is used to represent the risk degree of the vehicle failing to reasonably approach the site position specified by scheduling in the task; the sudden offset density is calculated based on the two-dimensional space difference vector between the future position and the future necessary position, and is used to represent the maximum degree of high offset of the vehicle in a continuous period. 6.The deep learning-based vehicle behavior dynamic detection method of claim 5, wherein, The trajectory anomaly point index set is a set of two-dimensional Euclidean distances of two-dimensional space difference vectors between actual positions and proper positions, which are greater than a system-set offset anomaly judgment threshold. 7.The deep learning-based vehicle behavior dynamic detection method of claim 1, wherein, The edge of the vehicle behavior graph is an edge weight between vehicles, that is, a behavior similarity. Then, the result of the group consistency analysis is generated based on a cosine similarity calculation of feature vectors between vehicles, and is used to measure the direction consistency. The higher the score of the group consistency analysis is, the greater the difference between the target vehicle and other vehicles in behavior offset is, and the more likely the target vehicle is an independent anomaly. 8.The deep learning-based vehicle behavior dynamic detection method of claim 7, wherein, The group-acceptable anomaly label set is: If the target vehicle anomaly time period is in the trajectory anomaly point index set and appears in most reference vehicle anomaly time windows at the same time, the anomaly is marked as group-acceptable, and a group-acceptable anomaly label set is constructed. Otherwise, it is a suspected individual violation. 9.The deep learning-based vehicle behavior dynamic detection method of claim 1, wherein, The abnormality of the vehicle behavior is judged by combining the offset feature, the collaborative offset score, and the group-acceptable anomaly label set, and specifically includes: The abnormality score is calculated by weighted fusion based on the average offset, the maximum offset, the site non-coverage loss, the burst offset density, and the proportion of time in the abnormal time slice that is not exempted by the group. The proportion of time in the abnormal time slice that is not exempted by the group is used to strengthen the punishment of individual anomalies. The final anomaly label is obtained by comparing the abnormality score with a system-set threshold: If the abnormality score is greater than the system-set threshold, the vehicle behavior is determined to be abnormal. Otherwise, the vehicle behavior is determined to be normal.

10. A deep learning based vehicle behavior dynamic detection system, characterized in that, The system includes: A rule condition construction module is configured to obtain a site sequence of a scheduling task, a time window corresponding to a site, and a task number, map the site sequence to a site space projection vector, map the time window corresponding to the site to a site time vector, map the task number to a fixed-length vector through a lookup table, and combine the space projection vector, the site time vector, and the fixed-length vector rule condition representation vector. A proper trajectory generation module is configured to input the rule condition representation vector into a pre-constructed trajectory generation network to generate a proper trajectory sequence that meets the requirements of the scheduling rules. The trajectory generation network includes a rule vector embedding module and a trajectory sequence generation module. An offset feature extraction module is configured to obtain a vehicle actual trajectory, extract offset features between the vehicle actual trajectory and the proper trajectory sequence, and extract a trajectory anomaly point index set. The offset features include an average offset, a maximum offset, a site non-coverage loss, and a burst offset density. A group behavior analysis module is configured to obtain a reference vehicle set of the same line for the current task, combine a target vehicle to construct a vehicle behavior graph, perform group consistency analysis based on the vehicle behavior graph to obtain a collaborative offset score, and judge whether an anomaly occurs in a group offset set time window based on the collaborative offset score and the trajectory anomaly point index set to generate a group-acceptable anomaly label set. An anomaly judgment module is configured to combine the offset feature, the collaborative offset score, and the group-acceptable anomaly label set to judge whether the vehicle behavior is abnormal.

Images