A method for early warning of cheating in grain collection based on space-time trajectory and path time verification

The grain collection anti-fraud early warning method based on spatiotemporal trajectory and path time verification solves the problem that existing technologies cannot automatically identify abnormal vehicle stops and path deviations. It realizes multi-dimensional cross-validation and long-term vehicle behavior analysis, reduces the false alarm rate and improves detection accuracy and anti-fraud capabilities.

CN123069302APending Publication Date: 2026-11-10ZHONGNONG SUNSHINE (JILIN PROVINCE) BIG DATA GROUP CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611586828.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-14
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing grain collection and transportation monitoring technologies cannot automatically identify covert cheating behaviors such as abnormal vehicle stops, route deviations, and abnormal transportation times. They also have a high false alarm rate, lack adaptive training of benchmark models and multi-model fusion mechanisms, and cannot achieve multi-dimensional cross-validation and long-term vehicle behavior analysis.

Method used

The grain harvesting anti-cheating early warning method based on spatiotemporal trajectory and path time verification is implemented through multi-dimensional data collection, benchmark model establishment and adaptive training, real-time multi-dimensional behavior analysis and anomaly judgment, hierarchical early warning and long-term vehicle behavior analysis. It adopts weighted evidence fusion and Bayesian combination strategy to achieve multi-model collaborative decision-making. Combined with a whitelist of legal parking points and dynamic threshold configuration, it reduces the false alarm rate and improves detection accuracy.

Benefits of technology

It achieves multi-dimensional cross-validation, automatically identifies hidden cheating behaviors, reduces false alarm rate, reduces manpower costs, enables real-time proactive early warning and differentiated handling, and improves anti-fraud capabilities and detection accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069302A_ABST
    Figure CN123069302A_ABST
Patent Text Reader

Abstract

A grain collection anti-cheating early warning method based on space-time trajectory and path time verification. It belongs to the technical field of Internet of Things and intelligent monitoring. It solves the technical problem that existing grain collection and transportation monitoring cannot automatically identify hidden cheating behaviors such as abnormal vehicle stay, path deviation, and abnormal transportation time. The method includes: collecting basic space-time data, kinematics data, and transportation task correlation data and returning them to the monitoring center; establishing and training six categories of benchmark sub-models in the monitoring center as a comparison reference for subsequent multi-dimensional behavior analysis; using a weighted evidence fusion and Bayesian combination strategy to realize multi-model collaborative decision-making; the monitoring center receives the multi-dimensional data stream returned by the vehicle, performs multi-dimensional analysis based on the trained benchmark model, generates an abnormal score when any dimension triggers an abnormal condition, generates a warning event after multi-model fusion; according to the final abnormal comprehensive index and the number of triggered dimensions of multi-model fusion, generate warning information of corresponding level.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things and intelligent monitoring technology, specifically relating to a method for early warning of grain harvesting fraud based on spatiotemporal trajectory and path time verification. Background Technology

[0002] In grain purchasing operations, grain transport vehicles deliver grain from farmers' fields to designated warehouses, creating certain blind spots in supervision throughout the process. Due to the lack of effective means of verifying behavior, some drivers use covert methods such as "stopping mid-journey to unload" and "taking detours to sell illegally" to intercept grain, causing economic losses to both enterprises and farmers, and also making it difficult to trace grain quality. Existing location monitoring technologies are mainly divided into the following categories: Simple GPS trajectory playback: Its main function is to record and play back the vehicle's historical driving trajectory. Its limitation is that it can only be viewed after the fact and cannot actively identify suspicious behaviors such as path deviation and abnormal stops. Electronic fence monitoring: Its main function is to determine whether a vehicle enters or leaves a preset area. Its limitation is that it only supports entry and exit triggers and cannot identify hidden behaviors such as detours or unloading during the journey. Manual video monitoring and verification: Its main function is for supervisors to watch the monitoring screen in real time. Its limitations include high labor costs, easy omissions, slow response, and high false alarm rate.

[0003] Therefore, there is an urgent need for a method to prevent cheating in grain collection and transportation that can automatically identify multi-dimensional abnormal behaviors, provide intelligent hierarchical early warnings, and reduce false alarms. Summary of the Invention

[0004] To address the technical problem that existing grain collection and transportation monitoring systems cannot automatically identify covert cheating behaviors such as abnormal vehicle stops, route deviations, and abnormal transportation times, this invention provides a grain collection anti-cheating early warning method based on spatiotemporal trajectory and path time verification.

[0005] The method includes: Multidimensional data acquisition: Collect basic spatiotemporal data, kinematic data, and transportation task-related data and transmit them back to the monitoring center; Benchmark model establishment and adaptive training: Six types of benchmark sub-models are established and trained in the monitoring center as comparison references for subsequent multi-dimensional behavior analysis; weighted evidence fusion and Bayesian combination strategies are adopted to achieve multi-model collaborative decision-making; Real-time multidimensional behavior analysis and anomaly judgment: The monitoring center receives multidimensional data streams transmitted from vehicles in real time, performs multidimensional analysis based on a trained benchmark model, generates an anomaly score when any dimension triggers an anomaly condition, and generates an early warning event after multi-model fusion. Tiered early warning and response: Based on the final comprehensive anomaly index resulting from the fusion of multiple models and the number of triggering dimensions, early warning information of corresponding levels is generated; Long-term grain collection vehicle behavior analysis and trend early warning: Based on real-time detection of single-trip transportation, a long-term behavior profile is established for each grain collection vehicle, and vehicle-level abnormal trends are identified through statistical analysis of historical transportation records.

[0006] Furthermore, the basic spatiotemporal data includes vehicle location coordinates, timestamps, and instantaneous speed; kinematic data includes acceleration, heading angle, and angular velocity; and transportation task-related data includes departure plot number, destination warehouse number, and driver identity information. The above multidimensional data constitute a digital twin data stream of the entire vehicle transportation process, denoted as a single-trip transportation data sequence.

[0007] Furthermore, the six major categories of benchmark sub-models include the electronic fence model for the work area, the standard transportation route model, the theoretical transportation time window model, the legal stop whitelist model, the speed characteristic benchmark model, and the dwelling pattern benchmark model.

[0008] Furthermore, the electronic fence model for the work area is a boundary polygon that delineates the legal grain collection work plots based on a GIS map, consisting of m vertices, which limits the legal starting area. The standard transportation route model is trained using historical trajectory clustering and median trajectory extraction methods. The specific steps are as follows: Historical trajectory preprocessing: Collect a set of historical normal transportation trajectories from the same grain receiving plot to the same target warehouse, and perform coordinate system transformation, trajectory resampling and noise filtering on each trajectory; Trajectory Clustering: The DBSCAN density clustering algorithm is used to cluster historical trajectory sets and identify major transportation routes; Standard path extraction: For each cluster, the median trajectory is extracted as the standard transportation path for that route; for the sampling point position after arc length resampling, the median latitude and longitude of all trajectories in the cluster at that position is taken; the standard path is represented as a broken line sequence, and the path deviation tolerance band width is calculated at the same time.

[0009] Furthermore, the theoretical transportation time window model is obtained by: statistically analyzing the actual transportation time distribution of the same route based on historical normal transportation records, and fitting a probability distribution model to determine a reasonable time window; The legal stop whitelist model is obtained by inputting the location coordinates and effective stopping radius of the locations that are allowed to temporarily stop along the route, forming an expandable whitelist database.

[0010] Furthermore, the specific method for obtaining the velocity characteristic benchmark model is as follows: Road segment division: Divide the standard route into M road segments; Segmented speed statistics: Calculate the historical average speed, standard deviation of speed of the j-th segment, 5th percentile and 95th percentile of speed of the j-th segment from the historical trajectory set; Speed ​​anomaly score: The speed anomaly score of the real-time road segment j is calculated based on the segmented speed statistics. At the same time, the overall speed variance is calculated as an overall speed stability index. The baseline model for the dwell mode is obtained as follows: Dwell Event Extraction: Extract all dwell events from the historical trajectory; Modeling of dwell time distribution: Establish a log-normal distribution model for legal dwell time, and fit a normal distribution after taking the logarithm of historical legal dwell time samples; Modeling of spatial distribution of stopping locations: A Gaussian mixture model is used to model the spatial distribution of legal stopping points.

[0011] Furthermore, the multi-model collaborative decision-making strategy, which combines weighted evidence fusion and Bayesian methods, is specifically implemented as follows: Anomaly score normalization: Normalize the dimensions and range of the anomaly scores output by each sub-model to the [0, 1] interval; Model weight calculation: Assign a dynamic weight to each sub-model, the weights being determined based on the model's detection performance on the historical validation set; Weighted evidence fusion: The weighted summation of all participating sub-models is calculated to obtain the weighted evidence fusion anomaly score for each sub-model; Bayesian combinational decision-making: Based on weighted fusion, a Bayesian combinational strategy is further introduced for probabilistic fusion; Final fusion decision: The final anomaly comprehensive index is obtained by combining the weighted evidence fusion anomaly score and the Bayesian posterior probability.

[0012] Furthermore, real-time multidimensional behavior analysis and anomaly judgment: The monitoring center receives multidimensional data streams transmitted from vehicles in real time, performs multidimensional analysis based on the trained benchmark model, generates an anomaly score when any dimension triggers an anomaly condition, and generates an early warning event after multi-model fusion. Multidimensional analysis includes path deviation analysis, abnormal stay analysis, abnormal transportation time analysis, abnormal speed analysis, abnormal heading analysis, and abnormal spatiotemporal combination analysis. Path deviation analysis is performed by matching and comparing the real-time vehicle trajectory with the standard transportation route on a map. Abnormal stop analysis is performed by identifying the relationship between speed and default speed and determining the spatial distribution of the whitelist and legal docking points after a stop. Anomaly analysis of transportation time is performed by comparing the actual time taken from departure to arrival with the time window. Speed ​​anomaly analysis is performed by comparing real-time speed with the speed benchmark model of the road segment. Anomaly analysis of the heading angle is performed by comparing the real-time heading angle with the tangent direction of the standard path. Spatiotemporal combination anomaly analysis is performed by determining the number of items that simultaneously satisfy the above analysis dimensions.

[0013] Furthermore, the early warning information includes Level 1, Level 2, and Level 3 warning levels; The triggering condition for a Level 1 warning is: the final comprehensive anomaly index ∈ [0.3, 0.5) or an anomaly analyzed in a single dimension; The triggering conditions for a Level II warning are: the final comprehensive anomaly index ∈ [0.5, 0.7) or both dimensions of analysis show anomalies simultaneously; The trigger conditions for a Level 3 warning are: the final comprehensive anomaly index >= 0.7 or all three dimensions of analysis are abnormal.

[0014] Furthermore, the long-term grain collection vehicle behavior analysis and trend early warning are specifically as follows: Vehicle historical behavior profile construction: For each grain collection vehicle, collect all its transportation records over the past N days to construct a vehicle behavior profile vector. Each profile dimension includes: route preference features, time pattern features, stop features, speed features, heading features, and anomaly comprehensive features. Vehicle behavior scoring model: Based on vehicle behavior profiles, calculate the comprehensive vehicle behavior score and classify vehicle behavior score levels based on the score; Seasonal trend analysis: Statistical analysis of the overall fleet anomaly rate and individual vehicle anomaly trends by month or quarter; Long-term data storage and analysis architecture: A hierarchical storage architecture is adopted to store long-term grain receiving vehicle analysis data, including a real-time layer, a near-line layer, an offline layer, and a profiling layer.

[0015] The beneficial effects of the method described in this invention are as follows: Multi-dimensional cross-verification automatically identifies hidden cheating: Breaking through the limitations of traditional GPS monitoring that can only display the trajectory, it uses multi-dimensional data cross-verification, such as location deviation, legality of stops, transportation time window, speed anomalies, and heading angle deviation, to make hidden behaviors such as "stopping midway to unload" and "taking detours to sell privately" nowhere to hide, significantly improving anti-fraud capabilities.

[0016] The whitelist mechanism and multi-model fusion reduce the false alarm rate: By combining a whitelist of legal parking spots, dynamic threshold configuration, and a multi-benchmark model fusion voting mechanism, legal parking behaviors such as normal refueling, traffic congestion, and temporary inspections are filtered out from abnormal judgments. At the same time, cross-validation between models further suppresses the misjudgment of a single model, effectively avoiding false alarms and reducing wrongful accountability for drivers.

[0017] Baseline Model Self-Training and Online Evolution: Based on historical transportation trajectory data, four baseline models—a standard path model, a time window model, a speed feature model, and a dwell pattern model—are automatically trained. Weighted evidence fusion and Bayesian combination strategies are employed to achieve collaborative decision-making among multiple models. The models support online incremental updates, continuously optimizing detection accuracy as business data accumulates.

[0018] Real-time proactive early warning reduces reliance on human resources: The entire process is automatically analyzed and early warnings are pushed out by the method, eliminating the need for supervisors to continuously monitor the screen, greatly saving labor costs, while shortening the response time to anomalies, and upgrading the ability to trace back to the event and intervene in the event.

[0019] Tiered early warning and precise handling: Level 1 to 3 early warnings are implemented according to the severity of the anomaly, and differentiated handling measures are matched (message reminder / SMS warning / remote vehicle locking) to avoid over-handling of minor anomalies and ensure that major cheating behaviors are promptly curbed. Attached Figure Description

[0020] Figure 1 This is a flowchart of the benchmark model training process in an embodiment of the present invention; Figure 2 This is a flowchart of the multi-benchmark model fusion decision-making process in an embodiment of the present invention; Figure 3 This is a flowchart of the multi-dimensional real-time behavior analysis in an embodiment of the present invention; Figure 4 This is a diagram of the long-term vehicle behavior analysis architecture in an embodiment of the present invention; Figure 5 This is a sequence diagram of an anomaly linkage early warning system, representing a typical implementation scenario in this invention. Figure 6 This is a flowchart illustrating the verification process for filtering false alarms of legitimate docking in a typical implementation scenario two of this invention. Figure 7 This is a schematic diagram illustrating the application of the method to a system architecture in an embodiment of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0022] Example 1 Existing technologies for monitoring grain receiving to prevent fraud have the following technical problems: they can only display the trajectory and cannot actively identify suspicious behaviors such as path deviation and abnormal stops; they lack systematic verification of the reasonableness of transportation time and cannot investigate the problem of "hidden overtime detours"; multi-dimensional data are not cross-correlation analyzed, making it difficult to identify complex fraud methods; they lack adaptive training and multi-model fusion mechanisms for benchmark models, and fixed thresholds lead to poor generalization ability; they do not utilize multi-dimensional sensor data such as the speed and heading of transport vehicles, resulting in a single analysis dimension; they lack statistical analysis of the historical behavior of grain receiving vehicles over a long period of time, making it impossible to achieve accurate vehicle-level profiling and anomaly trend prediction; and they rely entirely on manual monitoring and verification, resulting in slow response and a high false alarm rate.

[0023] Therefore, there is an urgent need for a method to prevent cheating in grain collection and transportation that can automatically identify multi-dimensional abnormal behaviors, provide intelligent hierarchical early warnings, and reduce false alarms. This method should also have the capabilities of adaptive training of benchmark models, multi-model fusion decision-making, comprehensive analysis of multi-dimensional sensor data, and long-term vehicle behavior modeling.

[0024] This embodiment aims to address the technical problem in existing grain collection and transportation monitoring that it cannot automatically identify covert cheating behaviors such as abnormal vehicle stops, route deviations, and abnormal transportation times. It provides a grain collection anti-cheating early warning method based on spatiotemporal trajectory and route time verification, specifically addressing the following issues: how to determine in real-time and automatically whether a vehicle deviates from the normal transportation route; how to accurately distinguish between legitimate stops (such as refueling or rest stops) and suspicious stops (covert unloading); how to detect time anomalies such as "detour for private sales" through transportation time window verification; how to reduce false alarm rates by combining multi-dimensional cross-validation and push warnings according to severity levels; how to adaptively train a baseline model based on historical transportation data and achieve multi-model fusion decision-making to improve detection accuracy and generalization ability; how to integrate multi-dimensional sensor data such as vehicle speed and heading angle to construct a more comprehensive anomaly detection system; and how to utilize historical transportation records of long-term grain collection vehicles for statistical analysis to achieve vehicle-level behavioral profiling and anomaly trend early warning.

[0025] Compared with the prior art, the method of this embodiment has the following beneficial effects, as shown in Table 1: Table 1:

[0026] Example 2 This embodiment further defines Embodiment 1, providing a specific technical solution for a grain harvesting anti-cheating early warning method based on spatiotemporal trajectory and path time verification.

[0027] The method includes multi-dimensional data acquisition, benchmark model establishment and adaptive training, real-time multi-dimensional behavior analysis and anomaly detection, tiered early warning and handling, and long-term grain receiving vehicle behavior analysis and trend early warning. Each step is described in detail.

[0028] I. Multidimensional Data Acquisition: GPS / BeiDou positioning terminals and multi-mode sensing devices are installed on grain transport vehicles at fixed time intervals. (Recommended every 5-30 seconds) Collect the following multi-dimensional data in real time and transmit it back to the monitoring center: (1) Basic spatiotemporal data: Vehicle position coordinates (Latitude and longitude, WGS84 coordinate system); Timestamp (UTC time); instantaneous speed (km / h, obtained from GPS Doppler speed measurement); (2) Kinematic data: acceleration (m / s², which can be obtained from velocity difference); Heading angle (Degrees, vehicle direction of travel, 0°~360°, obtained from GPS); angular velocity (degrees per second, reflecting the vehicle's turning rate); (3) Transportation task related data: Starting point number and destination warehouse number; Driver identification information (via IC card); The aforementioned multidimensional data constitutes a digital twin data stream of the entire vehicle transportation process, denoted as a single-trip transportation data sequence: (1) —A multidimensional data sequence of a single transport trip is a collection of data snapshots of the entire transport process; —Vehicle at time The position coordinate vector, ,in Latitude (unit: degrees, WGS84 coordinate system) Longitude (unit: degrees, WGS84 coordinate system); —Vehicle at time The instantaneous velocity (unit: km / h, value range [0, 120]) is obtained by GPS Doppler velocity measurement; —Vehicle at time The acceleration (unit: m / s², range [-5, 5]) is obtained from the velocity difference; —Vehicle at time The heading angle (unit: degrees, range [0°, 360°]), where 0° represents true north, increases clockwise, and is obtained from GPS; —Vehicle at time Angular velocity (unit: degrees / second) reflects the vehicle's turning rate and is calculated from the difference in heading angle between adjacent moments; —Loading and departure time (UTC timestamp), which is the time when the vehicle leaves the grain collection plot; —Arrival time at warehouse (UTC timestamp), which is the time when the vehicle arrives at the entrance of the target warehouse; —Timestamp (UTC time), sampling time index ; —Sampling interval (unit: seconds, recommended value: 5~30 seconds), the time difference between two adjacent data collections.

[0029] II. Baseline Model Establishment and Adaptive Training: The following six categories of benchmark sub-models were established and trained in the monitoring center as comparison references for subsequent multi-dimensional behavior analysis. The benchmark model training flowchart is as follows: Figure 1 As shown, the standard transportation route model, theoretical transportation time window model, speed characteristic benchmark model, and stop pattern benchmark model are the core models adaptively trained based on historical data. The electronic fence model for the work area and the whitelist model for legal stops are auxiliary verification models based on GIS map data and business rule configuration. Each sub-model is trained offline based on historical normal transportation data and supports online incremental updates.

[0030] (1) Electronic fence model of the work area Based on GIS maps, the boundary polygons of legal grain harvesting operation plots are delineated to define the legal starting area. The boundaries of the operation plots are defined by... A polygon composed of vertices is represented as: (2) In the formula: —The electronic fence for the work area is a closed polygonal area composed of ordered vertices, representing the geographical boundary of the legal grain collection operation plot; —Fence polygon number Planar coordinates of each vertex (unit: meters, planar coordinates after Mercator projection). ; —No. The east-west coordinates of each vertex (x-coordinate, unit: meters); —No. The north-south coordinates of each vertex (vertical coordinate, unit: meters); —The total number of vertices of the fence polygon (positive integer, This determines the level of detail in the shape of the fence boundary; Vehicle coordinates The determination of whether something is inside the fence uses the ray casting method, denoted by the indicator function: (3) In the formula: —Indicator function, return value 1 indicates vehicle coordinates. Located in the fence Internal (legal departure area), a return value of 0 indicates Located outside the fence (illegal departure area).

[0031] —The coordinates of the vehicle to be determined. ,in East-west coordinates (unit: meters) North-south coordinates (unit: meters), coordinate system and fence Consistent; Ray—from vehicle coordinates Starting from A horizontal ray extending from the positive axis is used to find the intersection with the boundary of the fence polygon; Crossing count—the total number of intersections between the ray and all edges of the fence polygon; odd numbers indicate... Inside a polygon, an even number indicates outside.

[0032] (2) Standard transportation route model The standard transportation route model is trained using historical trajectory clustering and median trajectory extraction methods. The specific training steps are as follows: 1. Historical trajectory preprocessing: Collect a set of historical normal transportation trajectories from the same grain receiving plot to the same target warehouse. Each trajectory This is a sequence of latitude and longitude coordinates. The following preprocessing is performed on each trajectory: Coordinate system transformation: Convert WGS84 latitude and longitude to plane coordinates (Mercator projection) for easier distance calculation; Trajectory resampling: at equal intervals (Default 50 meters) Resample the arc length of each trajectory to ensure that each trajectory has a uniform number of sampling points; Noise removal: Kalman filtering is used to remove GPS drift noise.

[0033] 2. Trajectory Clustering: The DBSCAN density clustering algorithm was used to cluster historical trajectory sets and identify major transportation routes. The Hausdorff distance metric was used as the core distance metric, with a neighborhood radius... The minimum number of samples (default 100 meters) and the minimum number of samples (MinPts, default 5) are the DBSCAN clustering parameters. (4) In the formula: —Trajectory With trajectory The Hausdorff distance (in meters) between two trajectories measures the maximum mismatch between them; a smaller H value indicates that the two trajectories are more similar. —No. The historical transportation trajectory is a sequence of latitude and longitude coordinates. ,in For the first The first of the trajectories Coordinates of each sampling point; —No. Historical transportation trajectory, defined as the same ; —Trajectory Any sampling point in the sample, ; —Trajectory Any sampling point in the sample, ; -point With point Euclidean distance between them (unit: meters) ; -right Each point in First find Neutral nearest point (Take the inf value), then in all Take the maximum value (take) ),Right now i arrive The distance to Hausdorf; -right Each point in First find Neutral nearest point (Take the inf value), then in all Take the maximum value (take) ),Right now arrive The distance to Hausdorf; —Take the larger of the two directed Hausdorff distances to ensure the symmetry of the distance metric; 3. Standard path extraction For each cluster The median trajectory is extracted as the standard transportation path for this route. For the j-th sampling point after arc length resampling, the median latitude and longitude of all trajectories within the cluster at that location is taken: (5) In the formula: —On the standard path The coordinates of each sampling point It is determined by the median of all trajectories within the cluster; —Sampling point number (positive integer) , This represents the uniform number of sampling points after arc length resampling; — Cluster The Middle The trajectory in the first Latitude values ​​at each sampling point (unit: degrees, WGS84 coordinate system). ; — Cluster The Middle The trajectory in the first Longitude values ​​at each sampling point (unit: degrees, WGS84 coordinate system). ; —The median function takes the median of the set of values ​​within the parentheses, which can resist the influence of outliers; — Cluster The total number of historical trajectories contained therein (positive integer); A standard path is represented as a broken line sequence: (6) In the formula: —Standard transportation routes are... A series of ordered sampling points forms a broken line sequence, which serves as a benchmark for path deviation detection. —On the standard path The coordinates of each sampling point are defined in equation (5). —The total number of sampling points (positive integer) for the standard path, calculated from the arc length resampling interval. The total mileage of the route determines the distance. ; Simultaneously calculate the path deviation tolerance band width, and take the 95th percentile of the maximum vertical distance from each trajectory within the cluster to the standard path: (7) In the formula: —Path deviation tolerance zone width (unit: meters) indicates the allowable range of a vehicle's deviation from the standard path; exceeding this width is considered an abnormal path deviation. —The 95th percentile function takes the 95th percentile value of the set of values ​​in parentheses, which can filter out 5% of extreme outliers. —No. The first of the historical trajectories sampling points to standard path The shortest vertical distance (unit: meters) is calculated using the algorithm for the shortest distance from a point to a polyline segment. —for the first All sampling points of the trajectory Take the maximum value, i.e., the first... The maximum distance by which a trajectory deviates from the standard path; For the maximization operator, It is the operand (argument) of this operator; —Trajectory number (positive integer) Traverse cluster C l All historical trajectories; — Cluster The total number of historical trajectories included; —Standard transport route, defined in equation (6). 3. Theoretical Transportation Time Window Model Based on historical normal transportation records, the actual transportation time distribution of the same route (same grain receiving plot → same warehouse) is statistically analyzed, and a probability distribution model is fitted to determine a reasonable time window.

[0034] (1) Construction of transportation time sample Let the first The actual time taken for a historical transport trip is Collect N valid transport time samples for the same route. .

[0035] (2) Distribution fitting and parameter estimation A normal distribution is used to fit the time-consuming samples. The parameters of the normal distribution are estimated using maximum likelihood estimation. (8) In the formula: —The sample mean of historical transportation time (in minutes) is a location parameter of a normal distribution, reflecting the average transportation time of the route; —Total number of valid historical transport records (positive integer, recommended) ≥ 500), which is the number of samples that took time; —No. Actual time taken for a historical transport trip (unit: minutes). ,in For the first departure time For the first Arrival time of the trip.

[0036] (9) In the formula: —The sample variance of historical transportation time (unit: minutes²) is the scaling parameter of the normal distribution, reflecting the degree of dispersion of transportation time; —Sample standard deviation of historical transportation time (in minutes).

[0037] (3) Time window calculation Based on the fitted distribution, the theoretical transportation time window Determine using the following formula: (10) In the formula: —The lower limit of the theoretical transportation time window (unit: minutes), which is the shortest reasonable time for normal transportation. If it is lower than this value, there may be shortcuts or data anomalies. —The sample mean of historical transportation time (in minutes); —Confidence level coefficient (dimensionless), take The time corresponds to the 95% confidence interval, take When the coefficient of variation corresponds to the 99% confidence interval; At that time, it is recommended to increase the size. To increase tolerance to 2.58; —The sample standard deviation of historical transportation time (in minutes) is obtained by taking the square root of equation (9); (11) In the formula: —The theoretical transportation time window upper limit (unit: minutes), which is the longest reasonable transportation time under normal conditions. Exceeding this value (including the excess percentage) (Afterwards) triggers an abnormal transportation time warning. In addition, based on the total mileage of the standard route and the speed limits of the road sections, the physical upper and lower bounds of the theoretical transportation time can be calculated as an auxiliary verification: (12) In the formula: —The physical shortest transportation time (unit: minutes), which is the theoretical shortest time when traveling at the maximum permissible speed throughout the entire journey, serves as a physical constraint for the lower limit of the statistical window; —The total mileage of the standard route (in kilometers), derived from the standard route in equation (6). The distances between adjacent sampling points are summed to obtain the result. —The maximum permissible average speed (unit: km / h) for the entire route is the weighted average of the speed limits for each section of the road or the legal maximum limit; (13) In the formula: —The physical maximum transport time (in minutes), which is the theoretical maximum time taken under the condition of traveling at the minimum permissible speed throughout the entire journey, serves as a physical constraint for the upper limit of the statistical window; —The minimum permissible average speed for the entire journey (unit: km / h). Speeds below this usually indicate severe congestion or abnormal stops. A speed of 15-20 km / h is recommended.

[0038] The final time window is the intersection of the statistical window and the physical window: (14) In the formula: —Final theoretical transportation time window (unit: minutes), a reasonable transportation time range after considering statistical distribution and physical constraints; —Take the larger of the two values ​​as the final lower limit to ensure that it is not lower than the physical minimum time; —Take the smaller of the two values ​​as the final upper limit to ensure that it does not exceed the physical maximum time.

[0039] 4. Legal docking point whitelist model Enter the location coordinates and effective stopping radius of permitted temporary stops along the route to create an expandable whitelist database: (15) In the formula: —The legal docking point whitelist database is a database of... A collection consisting of records of each stop; —No. A unique number (string) for each valid stop, used for internal system indexing; —No. The coordinates of a legal stop in Latitude (unit: degrees, WGS84 coordinate system) Longitude (unit: degrees, WGS84 coordinate system); —No. The effective parking radius of a legal parking spot (unit: meters, default 200 meters, can be adjusted according to the type of parking spot), the distance of the vehicle's stopping position from the parking spot. The distance does not exceed This is considered to be within the area of ​​that stop; ——No. The type label (enumerated value) of each legal stop point can be set, with values ​​such as "gas station", "checkpoint", "rest area" and "warehouse entrance". Different types can be configured with different allowed stay durations. — Dock number (positive integer) ; —The total number of valid docking points in the whitelist database (positive integer).

[0040] Whitelist matching judgment: Vehicle station coordinates The distance to the nearest stop on the whitelist is ,like If so, it is considered a legal stop.

[0041] (5) Velocity characteristic benchmark model Based on historical normal transportation trajectories, vehicle speed distribution characteristics are statistically analyzed by road segment to establish segmented speed benchmark models.

[0042] 1. Road segment division standard path Divided into Each section of road , No. The arc length interval of each road segment is .

[0043] 2. Segmented speed statistics For the set of historical trajectories falling on the first Speed ​​samples of each road segment Calculate the following statistics: (16) In the formula: —No. The historical average speed (unit: km / h) of each road segment reflects the typical speed level of vehicles driving normally on that road segment; —No. The total number of historical speed samples for each road segment (positive integer), that is, the number of speed sampling points where all historical trajectories fall within that road segment; -right from arrive Summation; —No. The historical trajectory in the first Speed ​​sampling values ​​within each road segment (unit: km / h, value range [0, 120]); —Road segment number (positive integer) , This represents the total number of road segments; (17) In the formula: —No. The standard deviation of speed for each road segment (unit: km / h) reflects the degree of dispersion of vehicle speeds on that road segment; The larger the value, the greater the speed fluctuation.

[0044] (18) In the formula: —No. The 5th percentile (unit: km / h) of the speed of each road segment, that is, 95% of the historical normal driving speed is higher than this value, is used as the lower boundary for judging low speed anomalies. The 5% quantile function takes the 5th percentile value for the velocity sample set within the parentheses.

[0045] (19) In the formula: —No. The 95th percentile (unit: km / h) of the speed of each road segment, that is, 95% of the historical normal driving speed is lower than this value, is used as the upper boundary for judging speeding anomalies. —95th percentile function, taking the 95th percentile value for the velocity sample set within the parentheses; 3. Speed ​​Anomaly Score Real-time speed v(t) on the road segment The velocity anomaly scoring uses a quantile-based scoring function: (20) In the formula: -time The speed anomaly score (dimensionless, range [0, +∞), usually normalized to [0,1]), the larger the value, the more serious the speed deviation from the historical normal range; —Vehicle at time The real-time speed (unit: km / h) is obtained by GPS Doppler speed measurement; -Pick The larger of the values ​​in parentheses and the score is used to ensure that the score is non-negative; when the speed is... This item is within the normal range. ; First item —Low-speed anomaly scoring component, when A positive value indicates that the speed is lower than the historical normal level (such as abnormal stop or slow driving). Second item —Overspeed anomaly score component, when A positive value indicates that the speed is higher than the historical normal level (such as speeding). Simultaneously, the variance of the speed over the entire journey is calculated as an indicator of overall speed stability: (twenty one) in (22), where: —The overall speed variance of a single transport trip (unit: (km / h)²) reflects the stability of the vehicle's speed throughout the entire transport process; The larger the value, the more drastic the speed fluctuation; —The total number of GPS speed sampling points for this transport (positive integer); —For all sampling times from arrive Summation; —Vehicle at time Instantaneous speed (unit: km / h); —The average speed of the entire transport journey (unit: km / h), which is the arithmetic mean of all sampled speeds.

[0046] (6) Baseline model of dwelling pattern Based on historical transportation stop behavior data, a benchmark model of stay patterns is established to distinguish between legal stays and suspicious stays.

[0047] 1. Extraction of dwell events Extract all pause events from the historical trajectory. (23), each dwell event (twenty four).

[0048] In the formula: —The i-th dwell event ( (25), This represents the total number of dwell events extracted from this historical trajectory; each dwell event is described by a 5-tuple of location, start and end times, duration, and type. —A collection consisting of all dwelling events extracted from a single historical trajectory. The total number of events; —No. Coordinates of the location of each dwell event (unit: meters) (26) Plane rectangular coordinates obtained by projection transformation of GPS latitude and longitude; —No. The start time of each dwell event (in seconds, using a Unix timestamp), i.e., the time when the vehicle speed remains below the dwell threshold (e.g., ...). The starting time point; —No. The end time of a dwell event (unit: seconds, Unix timestamp), that is, the point in time when the vehicle speed recovers above the dwell threshold; —No. Duration of each dwell event (in minutes). (27), and ; —No. The event type identifier (enumerated value) for each stop event, such as 1 = legal stop point, 2 = abnormal stop in non-whitelist area, 3 = stop in loading and unloading area, is determined by business rules.

[0049] 2. Modeling the distribution of dwell time Establish a log-normal distribution model for the legal docking duration: (28) In the formula: —Probability density function of legal stop duration (unit: 1 / minute), the larger the output value, the longer the stop duration. It is more common during legal docking; —Duration of stay (in minutes) ), independent variable; —The scaling parameter of the log-normal distribution (dimensionless, >0), reflecting the degree of dispersion of the logarithm of the stay duration; The larger the value, the greater the fluctuation in the duration of stay; —The location parameter (dimensionless) of the log-normal distribution reflects the central tendency of the logarithm of the duration of stay; The median length of stay; —Duration of stay The natural logarithm; —Natural exponential function It is the natural constant (e ≈ 2.71828); —Constant term, π is the value of pi (π ≈ 3.14159); After taking the logarithm of the historical legal parking duration samples, a normal distribution is fitted, and the parameters are estimated as follows: (29) In the formula: —The location parameter (dimensionless) of the log-normal distribution, namely the sample mean of the natural logarithm of the historical legal parking duration; —The scaling parameter (dimensionless) of the log-normal distribution, namely the sample standard deviation of the natural logarithm of the historical legal stop duration; —The total number of samples of historical legal parking events (positive integer); -right from arrive Summation; —No. Duration of stay for each historical legal docking event (in minutes) ); —No. The natural logarithm of the dwell time.

[0050] 3. Spatial distribution modeling of dwelling locations A Gaussian mixture model (GMM) is used to model the spatial distribution of legal docking points: (30) In the formula: —Probability density function of the spatial distribution of legal docking points (unit: 1 / m²), the larger the output value, the better the coordinates. The more likely it is to be a legal parking location; —Planar coordinates of the location to be evaluated (unit: meters, after Mercator projection); —The number of Gaussian components in the GMM, with each component corresponding to a legal docking density region; —No. The mixture weights of Gaussian components (dimensionless) , ), representing the percentage of the k-th docking area among all legal docking areas; —No. The probability density function of a two-dimensional Gaussian distribution, with parameters being the mean vector. Covariance Matrix ; —No. The mean vector of Gaussian components, , indicating the first The center coordinates of a densely packed parking area (unit: meters); —No. The 2×2 covariance matrix (in meters²) of the nth Gaussian component describes the nth Gaussian component. The spatial expansion range and direction of each docking area; --right from arrive Summation is the weighted summation of all Gaussian components.

[0051] 4. Abnormal stay score For real-time dwell events Its abnormality score is: (31) In the formula: —Stay Abnormal Score (dimensionless, range of values) The larger the value, the more abnormal the dwelling behavior; near For legal docking, approach The stay is considered highly suspicious. —Abnormal weighting coefficient for dwell time (dimensionless, default) ), control the proportion of abnormal duration in the overall score; —Weight coefficient for abnormal dwelling location (dimensionless, default) ), control the proportion of location anomalies in the overall score; ) —Duration of stay The probability density value in the distribution of legal parking durations (calculated by equation (28)) indicates that the larger the value, the more common the duration is in legal parking. —Duration of the real-time pause event (in minutes) ); —Location of stay The probability density value in the legal parking space GMM distribution (calculated by equation (30)) indicates that the larger the value, the more likely the location is to be a legal parking area; —The coordinates of the location where the event paused in real time. (Unit: meters) —Duration aberration component, The larger the value, the smaller the component, indicating a more normal duration; —Position anomaly component, The larger the value, the smaller the component, indicating a more normal position.

[0052] 7. Multi-benchmark model fusion The six benchmark sub-models described above each characterize normal transportation behavior from different dimensions. Except for the legal stop whitelist model used for pre-filtering the legality of stopping behavior, the other five models (fence, route, time window, speed, and stop) all participate in anomaly scoring and fusion decision-making. The multi-benchmark model fusion decision-making flowchart is as follows: Figure 2As shown, a single model may suffer from overfitting or insensitivity to specific scenarios. This invention employs a weighted evidence fusion and Bayesian combination strategy to achieve multi-model collaborative decision-making, thereby improving detection accuracy and robustness.

[0053] 1. Abnormal score normalization The anomaly scores output by each sub-model have different dimensions and ranges, and need to be normalized first. The interval. For the first interval. Anomaly scoring of individual sub-models Using Sigmoid mapping: (32) In the formula: —No. The anomaly score after normalization of each sub-model (dimensionless, value range (0, 1)), the larger the value, the higher the degree of anomaly identified by the model; —No. The original anomaly scores for each sub-model (dimensions and ranges vary by model), such as , , , , wait; —No. Sigmoid mapping steepness coefficients of each sub-model (dimensionless) ), controlling the steepness of the normalization curve; The larger the score, the higher the rating. The more drastic the changes in the vicinity; —No. The Sigmoid mapping center of each sub-model (and) (with the same dimensions), that is, the decision boundary point for scoring; the average of historical normal transport scores can be taken, so that the normalized score of normal transport is about 0.5. —Natural exponential function; — Sub-model number (positive integer) , The number of sub-models participating in the fusion; 2. Model weight calculation Each sub-model is assigned a dynamic weight, which is determined based on the model's detection performance (F1-Score) on the historical validation set: (33) (Normalized weights, ) In the formula: —No. The fusion weights of the sub-models (dimensionless) , The higher the weight, the greater the influence of the model in the fusion decision-making process; —No. The F1 score of each sub-model on the historical validation set (dimensionless, ranging from [0, 1]). Where Precision is the precision rate and Recall is the recall rate; -all Sub-model The sum of fractions (dimensionless) is used for normalization; — Sub-model number (positive integer) ; —The total number of sub-models participating in the fusion (positive integer).

[0054] 3. Weighted Evidence Fusion The fusion anomaly score is calculated as a weighted average of the anomaly scores of each sub-model: (34) In the formula: —Weighted evidence fusion anomaly score (dimensionless, value range [0, 1]), the larger the value, the more abnormal the transportation behavior; = 0 indicates that all models are judged as completely normal. This indicates that all models are judged as completely abnormal.

[0055] 4. Bayesian Combinatorial Decision Making Building upon weighted fusion, a Bayesian combination strategy is further introduced for probabilistic fusion. Based on Bayes' theorem and the conditional independence assumption (Naive Bayes), the anomalous posterior probability after fusion is: (35) In the formula: —Normalized scores for all known submodels arrive Under the given conditions, the posterior probability (dimensionless, range [0, 1]) of the transportation behavior being abnormal is given; the larger the value, the higher the probability of abnormality. —Abnormal events in transportation behavior; —Normal events in transportation activities, i.e. The supplementary event; —The prior probability of the abnormal event (dimensionless, range [0, 1]), which can be taken as the abnormality rate of historical transportation (e.g., 5%), representing the basic probability of transportation abnormality in the absence of any evidence; —Prior probability of a normal event (dimensionless). ; —Under the condition that the transportation behavior is indeed abnormal, the first Each sub-model outputs a score The conditional probability (obtained through statistical analysis of historical labeled data); —Under the condition that the transportation activity is indeed normal, the first Each sub-model outputs a score The conditional probability (obtained through statistical analysis of historical labeled data); —for all from arrive The product of the conditional probabilities of each sub-model (based on the conditional independence assumption). —The total number of sub-models participating in the fusion (positive integer).

[0056] 5. Final Integration Decision By combining the weighted evidence fusion anomaly score and the Bayesian posterior probability, the final anomaly comprehensive index is obtained: (36) In the formula: —The final anomaly composite index (dimensionless, value range [0, 1]) is the comprehensive anomaly judgment value after integrating information from all sub-models; ≥ A fusion warning is triggered when the value is 0.6 (default). The specific warning format is shown in Table 2. — Fusion adjustment coefficient (dimensionless, value range [0, 1], default) ), controlling the relative weight of the weighted fusion score and the Bayesian posterior probability; It relies entirely on weighted fusion. It relies entirely on Bayesian decision-making; —Weighted evidence fusion anomaly score (dimensionless), calculated by equation (34) —The Bayesian posterior probability (dimensionless) is calculated by equation (35); —Fusion warning trigger threshold (dimensionless, default 0.6), when The system will trigger an alert if this value is exceeded.

[0057] Table 2:

[0058] 6. Online incremental update of the model Once a new transportation task is completed and manually confirmed as normal or abnormal, the data is added to the training set, triggering an incremental model update. The update strategy uses a sliding window mechanism: retaining the most recent W trips (default 2000 trips) of transportation data, and gradually discarding older data outside the window. The incremental update frequency is configurable (daily or every 100 trips of new data is recommended). After the update, the model weights are recalculated. and conditional probability .

[0059] III. Real-time multidimensional behavior analysis and anomaly detection: Multi-dimensional real-time behavior analysis flowchart as follows Figure 3 As shown, the monitoring center receives multi-dimensional data streams from vehicles in real time. Based on a trained baseline model, it performs multi-dimensional behavioral analysis from the following six dimensions in Table 3. An anomaly score is generated when any dimension triggers an abnormal condition. After multi-model fusion, a warning event is generated. Table 3:

[0060] The newly added formula for calculating heading deviation is: (37) In the formula: -time The heading deviation (unit: degrees, range [0°, 180°]) represents the angle between the vehicle's actual driving direction and the tangent direction of the standard path; the larger the Δθ, the more serious the deviation of the vehicle from the standard driving direction. —Vehicle at time The real-time heading angle (unit: degrees, range [0°, 360°]) is obtained from GPS, where 0° represents true north; —Standard path at current arc length position The tangent direction angle at the point (unit: degrees, range [0°, 360°]), determined by the standard path. The coordinate difference between adjacent sampling points is calculated. —Vehicle at time The arc length position along the standard path (unit: meters), that is, the cumulative mileage corresponding to the projection point of the vehicle's real-time position on the standard path; — Angle normalization function, which folds the angle difference to Interval; specifically implemented as: if the difference Subtract If the difference Then add (This step does not take the absolute value; the absolute value is determined by the outer layer of equation (37)) (Operator complete) —Absolute value operator; for The absolute value of the output is taken, and the heading deviation is ultimately mapped to... The range, i.e., in equation (37) .

[0061] IV. Tiered Early Warning and Response: Based on the multi-model fusion anomaly comprehensive index S_final and the number of triggering dimensions, the system automatically generates warning information of corresponding levels and triggers different levels of response measures, as shown in Table 4: Table 4:

[0062] All warning events are recorded in the system log, including complete trajectory snapshots, timestamp sequences, velocity curves, heading deviation sequences, multi-model fusion decision logs (including scores, weights, and fusion scores of each sub-model), and the reasons for triggering the warnings, for post-event accountability and evidence collection.

[0063] V. Long-term Grain Collection Vehicle Behavior Analysis and Trend Early Warning Module: Based on real-time monitoring of single-trip transport, a long-term behavioral profile is established for each grain collection vehicle. By statistical analysis of historical transport records, vehicle-level abnormal trends are identified, achieving a leap from single-anomaly detection to long-term behavioral early warning capabilities.

[0064] 1. Vehicle historical behavior profile construction For each grain collection vehicle Collect its past All transportation records within the default 90 days are used to construct a vehicle behavior profile vector: (38) In the formula: -vehicle The behavioral profile vector, consisting of 6 feature sub-vectors, is a comprehensive digital description of the vehicle's long-term transportation behavior. —No. A unique identifier (string) for each grain collection vehicle; —Path preference feature sub-vector, defined in equation (39); —Time regularity feature sub-vector, defined as shown in equation (40); —Stay feature vector, defined in equation (41); —The velocity feature vector is defined in equation (42); —The heading feature vector is defined in equation (44); —The anomaly comprehensive feature sub-vector is defined in equation (45); —Statistical time window for long-term analysis (unit: days, default 90 days); The dimensions of each portrait are defined as follows: (1) Path preference characteristics : Vehicle statistics Frequency distribution of transport on each standard route: (39) In the formula: —A set of route preference features, recording the frequency distribution of vehicle transportation on each standard route; —Standard route number (string), which identifies a standard transportation route from the grain collection site to the warehouse; -vehicle In the statistics window The set of all standard routes actually traveled within the area; -vehicle On the route Historical transport counts (positive integers); —The proportion of total transportation trips along this route (dimensionless, range of values) ), , To count the total number of transports within the window; The route is considered a low-frequency route, and a vehicle suddenly switching to a low-frequency route is marked as an abnormal path preference. —Statistics Window Inner vehicle Total number of transports (positive integer); (2) Characteristics of time patterns : Historical distribution and trends of vehicle transportation time: (40) In the formula: —Time-based feature subvectors, describing the historical statistical characteristics and trends of vehicle transportation time; -vehicle The historical average transportation time (unit: minutes) is the arithmetic mean of all historical transportation times for this vehicle. -vehicle The historical standard deviation of transportation time (in minutes) reflects the degree of fluctuation in the transportation time of the vehicle. —The slope of the transportation time trend (unit: minutes / trip) is obtained by linear regression of the historical time series of the vehicle; This indicates that the time taken increases with each trip (there may be increasingly serious detours). This indicates that the time taken decreases with each trip; —Historical timeout rate (dimensionless, range [0, 1]). = Number of timeouts / Total number of shipments; Timeout is defined as actual time taken exceeding [a certain limit]. ;like and Then it is marked as an anomaly in the time trend. (3) Dwelling characteristics : Statistical analysis of vehicle historical dwelling behavior characteristics: (41) In the formula: —Stay feature subvector, describing the statistical characteristics of a vehicle's historical stay behavior; —The average number of stops per trip by a vehicle (dimensionless, value range ≥ 0), which is the total number of all stop events within the statistical window divided by the total number of trips; —The average duration of each vehicle stop (in minutes), which is the total duration of all stop events within the statistics window divided by the total number of stop events; —The spatial distribution characteristics (coordinate set) of the historical stopping locations of vehicles reflect the preferred stopping areas of vehicles; —Historical illegal stay rate (dimensionless, range [0, 1]). = Number of illegal stays / Total number of stays; if A sustained increase indicates a deterioration in dwell behavior; (4) Velocity characteristics : Statistical analysis of vehicle speed behavior characteristics: (42) In the formula: —A speed feature subvector that describes the statistical characteristics of a vehicle's historical speed behavior; —The historical average speed of the vehicle over the entire journey (unit: km / h), which is the arithmetic mean of all historical transport speeds; —Vehicle historical speed variance (unit: (km / h)²), reflecting the degree of fluctuation in vehicle speed; —Historical velocity anomaly rate (dimensionless, range [0, 1]). = Number of speed anomalies / Total number of transport trips; —The vehicle's highest instantaneous speed in history (unit: km / h) is used to assess whether there is serious speeding behavior; —The total number of speeding events within the statistics window (positive integer), that is, the number of times the speed exceeded the speed limit of the road segment; (5) Heading characteristics : Statistical analysis of vehicle heading deviation behavior: (43) In the formula: —Heading feature subvector, describing the statistical characteristics of the vehicle's historical heading deviation behavior; —Vehicle historical average heading deviation (unit: degrees, range) ), which is the arithmetic mean of the heading deviations during all transportation processes; —Historical heading anomaly rate (dimensionless, range of values) ), = Number of course deviations / Total number of transports; —Vehicle's historical maximum heading deviation (unit: degrees, range) This reflects the most severe degree of directional deviation of the vehicle; (6) Abnormal comprehensive characteristics : Statistics on historical vehicle warning triggering: (44) In the formula: —Anomaly comprehensive feature sub-vectors describe the statistical characteristics and trends of vehicle historical warning triggers; —Total number of warnings triggered by vehicles within the statistics window (positive integer); —Early warning level distribution (three-dimensional vector), recording the number of times each level 1, level 2, and level 3 early warnings occurred. For example, (5, 3, 1) represents 5 times for level 1, 3 times for level 2, and 1 time for level 3. —Early warning rate (dimensionless, range [0, 1]). , Total number of transports; The higher the value, the more abnormal the overall vehicle behavior; —Recent early warning trend (dimensionless), based on recent... Linear regression was performed on the sequence of warning frequencies for each transport trip; positive values ​​indicate an increase in warning frequency, and negative values ​​indicate a decrease. —Statistics Window Inner vehicle Total number of transports (positive integer).

[0065] 2. Vehicle Behavior Scoring Model Based on vehicle behavior profiles, calculate a comprehensive vehicle behavior score (0-100 points, 100 points being completely normal): (45) In the formula: -vehicle Comprehensive behavioral score (dimensionless, range of values) ), The score indicates that everything is perfectly normal. The score indicates an extreme anomaly; —No. The unique identifier for each grain collection vehicle; —The maximum score represents completely normal behavior; -right from arrive Summation, that is, summation of... Weighted summation of each portrait dimension; —No. The weights of each portrait dimension (dimensionless) , The higher the weight, the greater the impact of that dimension on the overall score; suggested values ​​are shown in Table 5 below. —No. Normalized outliers (dimensionless, range of values) of each portrait dimension feature vector ), This indicates that the dimension is completely normal. This indicates that the dimension is extremely abnormal; it is obtained by normalizing each sub-indicator of the dimension after comparing it with the historical normal range. —No. Each portrait dimension feature sub-vector , respectively corresponding , , , , , .

[0066] Table 5:

[0067] The vehicle behavior rating levels are shown in Table 6: Table 6:

[0068] 3. Seasonal trend analysis Grain transportation exhibits a clear seasonal pattern. The system provides monthly / quarterly statistics on the overall fleet anomaly rate and individual vehicle anomaly trends. (1) Monthly abnormality rate statistics of the fleet: (46) In the formula: —The fleet anomaly rate for a given month (dimensionless, percentage, range) This reflects the overall frequency of abnormal events in the fleet during that month; —The number of transport trips in which the fleet triggered the warning in that month (positive integer), that is, the number of transport trips in which at least one warning was triggered among all vehicles in that month; —The total number of transport trips by the fleet in that month (positive integer), which is the total number of transport trips completed by all vehicles in that month; —Statistical months (January to December).

[0069] (2) Monthly abnormal trends of individual vehicles: Calculate the monthly anomaly rate for each vehicle. Time series analysis is used to predict the anomaly rate for the following month. If the predicted value exceeds the historical average... This triggers a trend warning.

[0070] (3) The team-level statistical indicators are shown in Table 7: Table 7:

[0071] 4. Long-term data storage and analysis architecture Long-term grain collection vehicle analysis data adopts a hierarchical storage architecture: Real-time layer: Raw trajectory and sensor data for the past 7 days, stored in an in-memory database (such as Redis), supporting real-time querying; Nearline layer: Transportation records and early warning logs for the most recent 90 days, stored in a time-series database (such as InfluxDB), supporting rapid statistical analysis; Offline layer: Historical data of more than 90 days, stored in a columnar database, supporting large-scale offline analysis and model training; Profile layer: Vehicle behavior profiles and scores are stored in a relational database and updated daily / weekly / monthly.

[0072] Long-term data analysis provides a data source for incremental training of the benchmark model, forming a closed-loop evolutionary system of "data collection → model training → real-time detection → long-term analysis → model update".

[0073] Example 2 This embodiment is a further explanation of Embodiment 1. For example... Figure 7 As shown, the system architecture of this invention consists of the following three layers: Terminal layer: GPS / BeiDou positioning terminals (including SIM cards) and multi-mode sensing devices (IMU, OBD interface modules) installed on grain receiving vehicles are responsible for periodically reporting raw data such as position, speed, acceleration, and heading angle; Platform Layer: The monitoring center server, deployed in the cloud or privately, is responsible for data reception, baseline model management and training, multi-model fusion decision-making, real-time multi-dimensional trajectory analysis, early warning generation and push notifications, and long-term vehicle behavior analysis; the long-term vehicle behavior analysis architecture diagram is shown below. Figure 4 As shown; Application layer: The administrator's web-based dashboard and the driver's mobile app are used for viewing early warnings / handling / long-term analysis reports and uploading / confirming on-site photos, respectively.

[0074] Implementation Scenario 1: Autumn Harvest Grain Transportation (Multi-Dimensional Anomaly Linkage Early Warning), the anomaly linkage early warning sequence diagram for Implementation Scenario 1 is as follows: Figure 5 As shown.

[0075] Application Background: During the autumn harvest, a grain purchasing company's grain collection vehicle A departs from farmer's plot B, travels along County Road S101, and is transported to the target warehouse C. The standard transportation distance is approximately 28 kilometers, and the theoretical transportation time window is 40-60 minutes (based on a time window model trained using historical data from 200 trips). , , , ).

[0076] Baseline model training process: Before the autumn harvest began, the system used over 500 normal transport records accumulated from the previous quarter to complete the following model training: Standard route model: DBSCAN clustering of historical trajectories for a 28-kilometer route ( , One main route was identified, and the median trajectory was extracted as the standard path. Deviation from tolerance band width rice; Time window model: Historical duration follows a normal distribution ,Pick have to , ; Speed ​​characteristic model: Divide the road into segments at 50-meter intervals and statistically analyze the mean and quantile of the speed in each segment; Dwelling pattern model: Log-normal distribution parameters of legal stop duration , The spatial distribution of GMM contains 5 components; Model fusion: The F1 scores of each model on the validation set are as follows: Normalized weights are .

[0077] Implementation process: (1) After vehicle A departs, the system begins to record the multidimensional data stream and starts timing the time window; (2) At kilometer 18, the vehicle deviated from the S101 main road and entered a country road, with a continuous deviation of 350 meters (more than 100 meters). (meters) and continues for 8 minutes (exceeding the 5-minute threshold), path deviation score. The system triggers a path deviation anomaly (Level 1 warning) and pushes an APP message to the administrator; (3) The vehicle then stayed at the deviation point for 12 minutes. The distance from the stopping coordinates to the nearest whitelist point was 850 meters (>200 meters). The spatial distribution probability of GMM is... (Below the 0.1 threshold), abnormal stay score The abnormal stop was triggered (Level 1 warning overlay), upgraded to Level 2 warning, and the driver was required to upload photos of the scene. (4) The driver did not respond to the photo upload request, and the final actual total transportation time reached 95 minutes, exceeding the theoretical upper limit of 60 minutes by 30% (i.e., the 78-minute upper limit), resulting in an abnormal transportation time score. ; (5) All three dimensions are abnormal, and the combined abnormality index is calculated. The system automatically escalates to a Level 2 warning, continuously pushes alerts, and notifies the administrator to pay attention. (6) Afterwards, the transport data was marked as abnormal and added to the training set, triggering an incremental update of the model. Vehicle A's behavior score dropped from 82 to 58, and the level was downgraded to a warning. The system suggested that the driver be interviewed and that training be strengthened.

[0078] Implementation Scenario 2: Validation of False Alarm Filtering for Legitimate Docking. The flowchart for Implementation Scenario 2, validation of false alarm filtering for legitimate docking, is as follows: Figure 6 As shown.

[0079] Application background: A grain collection vehicle D departs from a farmer's plot of land and heads to a warehouse. Due to insufficient fuel, it stops at a pre-designated gas station for refueling, with a stop time of approximately 15 minutes.

[0080] Implementation process: (1) When vehicle D arrives at the gas station, its speed decreases to 0 and it stops. The system detects the stopped state. (2) The system automatically searches the whitelist of legal stopping points and calculates the distance between the stopping coordinates and the gas station locations in the whitelist as 45 meters. The meter threshold is used to confirm a legal docking; simultaneously, the GMM spatial distribution probability is used. (Above the 0.1 threshold), double verification passed; (3) Even if the stay lasts for more than 10 minutes, the system will still mark the stay as a legal behavior and give an abnormal stay score. This avoids triggering abnormal stop warnings and prevents false alarms. (4) After refueling, vehicle D continued driving and eventually arrived at the warehouse normally. The transportation time was within the theoretical window, and the entire process was integrated with the comprehensive abnormal index. (Far below the 0.3 threshold), no warnings were generated throughout the process, verifying the effectiveness of the whitelist filtering and multi-model fusion mechanism.

[0081] Key Algorithm Description (1) Path deviation judgment: The shortest distance algorithm from point to polyline segment is adopted to determine the real-time coordinates of the vehicle. Compared with standard path polyline Perform map matching and calculate the minimum offset distance. If continuous All sampling points satisfy and duration (Default 5 minutes) If the path deviates from the target range, it is considered an abnormal path deviation. Path deviation anomaly score: (47) In the formula: —Path deviation anomaly score (dimensionless, value range [0, 1]), the larger the value, the more serious the path deviation; Indicates no deviation. This indicates a serious deviation; — An upper bound truncation function restricts the value to a certain range. Within the specified range, prevent score overflow; —The maximum vertical distance (in meters) from the vehicle to the standard path during the period of continuous deviation is obtained by taking the maximum value among all deviation sampling points; —Path deviation tolerance band width (unit: meters), obtained by training with equation (7), represents the allowable path deviation range; —The deviation range percentage (dimensionless) indicates the relative degree to which the tolerance zone is exceeded; —Duration of consecutive deviations (in minutes), from the first one exceeding The time span from the sampling point to the last off-sampling point; —Path deviation duration threshold (unit: minutes, default 5 minutes). Deviations exceeding this value are considered valid anomalies. (2) Judgment of abnormal stop: monitoring vehicle speed (Default 0.5km / h) and duration At a certain time, the location is marked as stationary. Then, a double validity check is performed: ① The list of valid docking points is searched; if the distance from the stationary coordinates to the nearest whitelisted point is... If the parking is illegal, it is considered an illegal stop; ② Calculate the probability of the stopping location in the GMM spatial distribution. ,like If so, it is further confirmed as an abnormal location. If the duration of stay is... (Default 10 minutes) triggers an abnormal stay warning. The abnormal stay score is shown in equation (31).

[0082] (3) Verification of transportation time window: based on the timestamp of the vehicle leaving the grain receiving plot after loading. Timestamp of arrival at the target warehouse entrance fence Calculate the actual time consumed .like ,in If the over-limit percentage is configurable (default 30%), a transit time anomaly warning will be triggered. Transit time anomaly score: (48) In the formula: —Transportation time anomaly score (dimensionless, value range [0, 1]), the larger the value, the more abnormal the transportation time; This indicates the time taken within the window. Indicates a serious timeout —The upper limit truncation function restricts the score to Within the range; —Lower bound cutoff function to ensure non-negative scores; when The time score is 0; —Actual transportation time (unit: minutes) ,in For the departure time, The arrival time; —The upper limit of the theoretical transportation time window (unit: minutes) is determined by equation (11) or equation (14); —Excess ratio coefficient (dimensionless, default 0.3, i.e., 30%), indicating the allowable excess. The proportion; actual timeout threshold = ; —Timeout percentage (dimensionless), representing the proportion of actual timeout to the allowable timeout range; (4) Speed ​​anomaly judgment: real-time speed and the road section speed benchmark The speed anomaly score is shown in equation (20). Meanwhile, if the overall speed variance... Exceeding the historical normal range This also triggers speed anomalies.

[0083] (5) Heading anomaly judgment: Calculate the deviation between the real-time heading angle and the standard path tangent direction. ,like Continue to exceed This triggers a heading anomaly. Heading anomaly score: (49) In the formula: —Heading anomaly score (dimensionless, range of values) The larger the value, the more severe the course deviation; Indicates no deviation; — An upper bound truncation function restricts the value to a certain range. Within the range; —Maximum deviation angle during the duration of heading deviation (unit: degrees, range) ), calculated by equation (37) and taking the maximum value during the continuous anomaly period; —The maximum possible deviation in heading (in degrees), i.e., traveling in the completely opposite direction; — Deviation amplitude ratio (dimensionless), representing the proportion of the deviation to the maximum deviation; —The heading deviation continues to exceed The time (in minutes) is counted from the moment the deviation first exceeds the threshold. —Tail deviation duration threshold (unit: minutes, default 3 minutes). Deviations exceeding this value are considered valid anomalies. —Duration ratio (dimensionless), representing the proportion of the deviation duration to the threshold.

[0084] (4) Multi-model fusion judgment: The system maintains the abnormal state vector of each transportation task. Each dimension is marked with a flag position of 1 when it is abnormal. After the abnormal scores of each dimension are normalized, they are weighted and fused according to formula (34), and the final abnormal comprehensive index is calculated according to formula (36). .when Exceeding the threshold or number of abnormal dimensions When the warning level is raised, it will be automatically upgraded.

Claims

1. A method for early warning against cheating in grain harvesting based on spatiotemporal trajectory and path time verification, characterized in that, The method includes: Multidimensional data acquisition: Collect basic spatiotemporal data, kinematic data, and transportation task-related data and transmit them back to the monitoring center; Benchmark model establishment and adaptive training: Six types of benchmark sub-models are established and trained in the monitoring center as comparison references for subsequent multi-dimensional behavior analysis; weighted evidence fusion and Bayesian combination strategies are adopted to achieve multi-model collaborative decision-making; Real-time multidimensional behavior analysis and anomaly judgment: The monitoring center receives multidimensional data streams transmitted from vehicles in real time, performs multidimensional analysis based on a trained benchmark model, generates an anomaly score when any dimension triggers an anomaly condition, and generates an early warning event after multi-model fusion. Tiered early warning and response: Based on the final comprehensive anomaly index resulting from the fusion of multiple models and the number of triggering dimensions, early warning information of corresponding levels is generated; Long-term grain collection vehicle behavior analysis and trend early warning: Based on real-time detection of single-trip transportation, a long-term behavior profile is established for each grain collection vehicle, and vehicle-level abnormal trends are identified through statistical analysis of historical transportation records.

2. The method for early warning against cheating in grain harvesting based on spatiotemporal trajectory and path time verification according to claim 1, characterized in that, The basic spatiotemporal data includes vehicle location coordinates, timestamps, and instantaneous speed; kinematic data includes acceleration, heading angle, and angular velocity; transportation task-related data includes departure plot number and destination warehouse number, and driver identity information; the above multidimensional data constitute a digital twin data stream of the entire vehicle transportation process, denoted as a single-trip transportation data sequence.

3. The method for early warning of grain receiving fraud based on spatiotemporal trajectory and path time verification according to claim 2, characterized in that, The six major categories of benchmark sub-models include the electronic fence model for the work area, the standard transportation route model, the theoretical transportation time window model, the legal stop whitelist model, the speed characteristic benchmark model, and the dwelling pattern benchmark model.

4. The method for early warning against cheating in grain harvesting based on spatiotemporal trajectory and path time verification according to claim 3, characterized in that, The electronic fence model for the work area is a boundary polygon that delineates the legal grain collection work plots based on a GIS map. It consists of m vertices and defines the legal starting area. The standard transportation route model is trained using historical trajectory clustering and median trajectory extraction methods. The specific steps are as follows: Historical trajectory preprocessing: Collect a set of historical normal transportation trajectories from the same grain receiving plot to the same target warehouse, and perform coordinate system transformation, trajectory resampling and noise filtering on each trajectory; Trajectory Clustering: The DBSCAN density clustering algorithm is used to cluster historical trajectory sets and identify major transportation routes; Standard path extraction: For each cluster, the median trajectory is extracted as the standard transportation path for that route; for the sampling point position after arc length resampling, the median latitude and longitude of all trajectories in the cluster at that position is taken; the standard path is represented as a broken line sequence, and the path deviation tolerance band width is calculated at the same time.

5. The method for early warning of grain harvesting fraud based on spatiotemporal trajectory and path time verification according to claim 4, characterized in that, The theoretical transportation time window model is obtained by: statistically analyzing the actual transportation time distribution of the same route based on historical normal transportation records, and fitting a probability distribution model to determine a reasonable time window; The legal stop whitelist model is obtained by inputting the location coordinates and effective stopping radius of the locations that are allowed to temporarily stop along the route, forming an expandable whitelist database.

6. The method for early warning of grain harvesting fraud based on spatiotemporal trajectory and path time verification according to claim 5, characterized in that, The specific method for obtaining the velocity characteristic benchmark model is as follows: Road segment division: Divide the standard route into M road segments; Segmented speed statistics: Calculate the historical average speed, standard deviation of speed of the j-th segment, 5th percentile and 95th percentile of speed of the j-th segment from the historical trajectory set; Speed ​​anomaly score: The speed anomaly score of the real-time road segment j is calculated based on the segmented speed statistics. At the same time, the overall speed variance is calculated as an overall speed stability index. The baseline model for the dwell mode is obtained as follows: Dwell Event Extraction: Extract all dwell events from the historical trajectory; Modeling of dwell time distribution: Establish a log-normal distribution model for legal dwell time, and fit a normal distribution after taking the logarithm of historical legal dwell time samples; Modeling of spatial distribution of stopping locations: A Gaussian mixture model is used to model the spatial distribution of legal stopping points.

7. The method for early warning of grain harvesting fraud based on spatiotemporal trajectory and path time verification according to claim 6, characterized in that, The specific steps for achieving multi-model collaborative decision-making using a weighted evidence fusion and Bayesian combination strategy are as follows: Anomaly score normalization: Normalize the dimensions and range of the anomaly scores output by each sub-model to the [0, 1] interval; Model weight calculation: Assign a dynamic weight to each sub-model, the weights being determined based on the model's detection performance on the historical validation set; Weighted evidence fusion: The weighted summation of all participating sub-models is calculated to obtain the weighted evidence fusion anomaly score for each sub-model; Bayesian combinational decision-making: Based on weighted fusion, a Bayesian combinational strategy is further introduced for probabilistic fusion; Final fusion decision: The final anomaly comprehensive index is obtained by combining the weighted evidence fusion anomaly score and the Bayesian posterior probability.

8. The method for early warning against cheating in grain harvesting based on spatiotemporal trajectory and path time verification according to claim 7, characterized in that, Real-time multidimensional behavior analysis and anomaly judgment: The monitoring center receives multidimensional data streams transmitted from vehicles in real time, performs multidimensional analysis based on a trained benchmark model, generates an anomaly score when any dimension triggers an anomaly condition, and generates an early warning event after multi-model fusion. Multidimensional analysis includes path deviation analysis, abnormal stay analysis, abnormal transportation time analysis, abnormal speed analysis, abnormal heading analysis, and abnormal spatiotemporal combination analysis. Path deviation analysis is performed by matching and comparing the real-time vehicle trajectory with the standard transportation route on a map. Abnormal stop analysis is performed by identifying the relationship between speed and default speed and determining the spatial distribution of the whitelist and legal docking points after a stop. Anomaly analysis of transportation time is performed by comparing the actual time taken from departure to arrival with the time window. Speed ​​anomaly analysis is performed by comparing real-time speed with the speed benchmark model of the road segment. Anomaly analysis of the heading angle is performed by comparing the real-time heading angle with the tangent direction of the standard path. Spatiotemporal combination anomaly analysis is performed by determining the number of items that simultaneously satisfy the above analysis dimensions.

9. A method for early warning against cheating in grain harvesting based on spatiotemporal trajectory and path time verification according to claim 8, characterized in that, The early warning information includes Level 1, Level 2, and Level 3 warning levels; The trigger condition for a Level 1 warning is: the final comprehensive anomaly index. Or analyze anomalies using a single dimension; The trigger condition for a Level II warning is: the final comprehensive anomaly index. Or both dimensions of analysis show anomalies; The trigger condition for a Level III warning is: the final comprehensive anomaly index. Or all three dimensions of analysis are abnormal.

10. A method for early warning against cheating in grain harvesting based on spatiotemporal trajectory and path time verification according to claim 9, characterized in that, The long-term grain collection vehicle behavior analysis and trend early warning are as follows: Vehicle historical behavior profile construction: For each grain collection vehicle, collect all its transportation records over the past N days to construct a vehicle behavior profile vector. Each profile dimension includes: route preference features, time pattern features, stop features, speed features, heading features, and anomaly comprehensive features. Vehicle behavior scoring model: Based on vehicle behavior profiles, calculate the comprehensive vehicle behavior score and classify vehicle behavior score levels based on the score; Seasonal trend analysis: Statistical analysis of the overall fleet anomaly rate and individual vehicle anomaly trends by month or quarter; Long-term data storage and analysis architecture: A hierarchical storage architecture is adopted to store long-term grain receiving vehicle analysis data, including a real-time layer, a near-line layer, an offline layer, and a profiling layer.