Vehicle positioning and intelligent risk control analysis method and readable storage medium

Abstract

The application discloses a kind of based on vehicle positioning and intelligent risk control analysis method and readable storage medium, the method takes vehicle terminal multi-source original data generation standardization high-precision driving track sequence;Through multidimensional feature point extraction and self-adaptive compression, remove redundant data, combined with space gridding and time slicing constructs composite space-time index, generates lightweight searchable storage level track set;Track set is input into pre-training neural network, and the feature is extracted by combining mechanical behavior parameter model, and the risk mode is identified by the dual engine of rule hard constraint and model soft matching, and the structured risk control evaluation report is generated by multi-factor scoring model;Linkage GIS completes data visualization, and output interactive vehicle asset monitoring risk early warning view.The application realizes the efficient management and control of track data whole process, significantly improves the identification performance of abnormal driving behavior, achieves the scientization, quantification and visualization of risk assessment, and provides an integrated risk control scheme for vehicle operation management.

Description

Technical Field

[0001] This invention relates to the field of vehicle positioning technology, and in particular to a vehicle positioning and intelligent risk control analysis method and a readable storage medium. Background Technology

[0002] With the rapid development of vehicle-to-everything (V2X) technology, the dimensions of vehicle positioning and status data that can be collected by in-vehicle terminals are constantly enriching. Vehicle positioning and driving behavior risk control have become core requirements for transportation and vehicle operation management. However, in existing technologies, the multi-source raw data collected by in-vehicle terminals suffers from problems such as high noise, positioning drift, and inconsistent sampling frequencies, resulting in low accuracy of trajectory data. At the same time, the storage cost of massive trajectory data is high and the retrieval efficiency is low, making it difficult to support efficient risk control analysis. In addition, traditional driving behavior risk control often relies on single rule judgments or simple model recognition, which is insufficient in recognizing complex and abnormal driving behaviors, and the risk control results lack intuitive visualization, failing to meet the actual needs of enterprises for vehicle asset monitoring and risk warning. Summary of the Invention

[0003] To address the aforementioned issues, this invention proposes a vehicle positioning and intelligent risk control analysis method and system based on a multi-level architecture. By standardizing the processing of multi-source raw data, lightweight storage of trajectory data, intelligent risk control analysis that integrates physical rules and deep learning, and visual interaction of risk control results, this invention achieves high-precision vehicle positioning and efficient intelligent risk control, thereby improving the level of intelligence in vehicle operation management.

[0004] Specifically, this invention discloses a method for vehicle positioning and intelligent risk control analysis, characterized by the following steps: Collect multi-source raw positioning data and vehicle status data of the vehicle, process the multi-source raw positioning data and vehicle status data to obtain vehicle driving trajectory and status data sequence; Based on the vehicle driving trajectory and state data sequence, redundant data is removed using a feature point extraction algorithm and an adaptive compression mechanism, and a multi-dimensional spatiotemporal index is constructed to generate trajectory set data. The trajectory set data is input into a pre-trained neural network risk control model, and combined with a formulaic behavior parameter model, abnormal driving behavior features are extracted and risk patterns are matched, and a structured risk control assessment report containing risk level and abnormality type is output. In response to external risk control query commands, the system uses a distributed retrieval engine to retrieve and visualize related data based on the structured risk control assessment report. Specifically, the step of inputting the trajectory set data into a pre-trained neural network risk control model and combining it with a formulaic behavioral parameter model to extract abnormal driving behavior features and match risk patterns includes: The neural network risk control model is used to extract deep nonlinear spatiotemporal features from the trajectory set data. At the same time, a formulaic behavior parameter model is used to calculate driving behavior indicators based on physical rules, and the final hybrid feature representation is output. The final hybrid feature representation is compared with a pre-set abnormal driving behavior database in multiple dimensions, and a classifier or similarity algorithm is used to identify potential risk patterns such as sudden acceleration and deceleration and route deviation.

[0005] In a second aspect, the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions for causing a computer to execute the above-described method for generating a dynamic electronic fence for automobiles.

[0006] This invention employs a full-link data governance technique that combines multi-source data standardization processing with lightweight storage and efficient retrieval. It achieves high-precision trajectory generation through Kalman filtering and HMM map matching, and relies on multi-dimensional feature extraction, adaptive compression mechanisms, and composite spatiotemporal index construction to completely solve the industry pain points of low accuracy of raw positioning data, high cost of massive trajectory storage, and low retrieval efficiency. Under the premise of ensuring the integrity of driving semantics, it realizes the benefits of efficient management and control of the entire process of vehicle trajectory data from collection, cleaning, storage to retrieval.

[0007] This invention employs a dual-channel intelligent risk control and visualization interaction technology that integrates physical rules and deep learning. By extracting physical features and deep semantic features in parallel, it constructs a dual risk identification engine with hard rule constraints and soft model matching. Combined with a confidence-calibrated quantitative scoring model and GIS-linked rendering, it significantly improves the accuracy and generalization ability of abnormal driving behavior identification, realizing the scientific, quantitative, and visual nature of risk assessment. This provides vehicle operation management with an integrated risk control solution that combines interpretability and intelligent decision-making capabilities.

[0008] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1A flowchart of the vehicle positioning and intelligent risk control analysis method provided in the embodiments of this disclosure.

[0011] Figure 2 A flowchart illustrating a method for generating a lightweight and easily searchable storage-level trajectory set, as provided in embodiments of this disclosure.

[0012] Figure 3 This is a flowchart illustrating a method for extracting abnormal driving behavior features and matching risk patterns, provided in an embodiment of this disclosure.

[0013] Figure 4 This is a flowchart illustrating a method for extracting abnormal driving behavior features and matching risk patterns, provided in an embodiment of this disclosure.

[0014] Figure 5 This is a flowchart illustrating a method for achieving effective alignment of multi-source heterogeneous features and outputting the final hybrid feature representation, as provided in an embodiment of this disclosure.

[0015] Figure 6 A flowchart illustrating a risk warning method for monitoring vehicle assets provided in this embodiment.

[0016] Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present disclosure. Detailed Implementation

[0017] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0018] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0019] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0020] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The drawings only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0021] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0022] See Figure 1 As shown in the figure, the present disclosure of an embodiment of a vehicle positioning and intelligent risk control analysis method specifically includes the following steps: Step S1: Obtain multi-source raw positioning data and sensor status data collected by the vehicle terminal, perform noise removal and cleaning, drift correction and spatiotemporal alignment processing to obtain a standardized high-precision vehicle driving trajectory sequence. Figure 2 This is a flowchart illustrating the method for obtaining a standardized high-precision vehicle trajectory sequence from an in-vehicle terminal, as disclosed in this embodiment. (Specifically, it is combined with...) Figure 2 It can be seen that this step specifically includes: Step S101: Establish a connection with the vehicle terminal (such as T-BOX or OBD device) via a wireless communication link, and receive and parse the data packets uploaded by the terminal at a preset frequency in real time. In a specific embodiment, the collected content includes reading positioning packets (including statements such as $GPGGA and $GPRMC) that conform to the NMEA-0183 standard protocol from the high-precision GNSS positioning module through a serial communication interface (such as UART or SPI), parsing and extracting latitude and longitude coordinates, altitude, and UTC timestamp; collecting the vehicle's triaxial acceleration and angular velocity raw data through the microelectromechanical system (MEMS) inertial measurement unit (IMU) integrated inside the terminal, and calculating the vehicle's instantaneous velocity vector and heading angle using a Kalman filter algorithm; and connecting to the vehicle controller through the OBD-II physical interface or by directly connecting the vehicle's CAN-High and CAN-Low signal lines. The local area network (CANBus) uses a CAN transceiver to listen for and capture broadcast data frames that conform to the SAE J1939 or ISO15765 communication protocol. It matches and decodes the message IDs according to the preset DBC (DatabaseCAN) parsing rule file to obtain the real-time vehicle sensor status such as ACC ignition status, engine speed, throttle opening, and diagnostic fault codes (DTC). For example, the system is configured with a sampling timer to synchronously trigger the reading and encapsulation of the above multi-source data at a frequency of 1Hz, generating a structured real-time data stream containing "longitude 114.05, latitude 22.54, speed 65km / h, heading 270°, ACCON" and their corresponding timestamps.

[0023] Step S102: Use filtering algorithms to remove outliers, noise, and duplicate / invalid data from the data. In a specific embodiment, Kalman filtering or bidirectional filtering algorithms are applied to smooth the original GNSS coordinate sequence. Physical constraint thresholds are set (e.g., judging an instantaneous speed exceeding 180 km / h or a single-second displacement exceeding 50 meters as an abnormal jump) to filter outliers caused by signal interference. At the same time, the uniqueness of the uploaded data is verified based on the timestamp and message ID to remove duplicate / invalid records caused by network retransmission.

[0024] The Kalman filter algorithm is applied to smooth the original GNSS coordinate sequence, and a state equation is constructed. With observation equation ;in, express The vehicle state vector at any given time (including longitude, latitude, and corresponding velocity components). Here is the state transition matrix. for The observation vector at that moment (i.e., the original GNSS coordinates). The observation matrix maps the state vector to the observation space. and These represent process noise and observation noise, respectively. The Kalman gain is calculated... Using the update formula The trajectory points are optimized and corrected, where For prior estimates, based on The predicted value of the optimal estimate at time; for Posterior state estimation at time step; Kalman gain , The covariance matrix is ​​used for prior estimation, representing the uncertainty of the prior estimate; To observe the noise covariance matrix and characterize GNSS positioning accuracy, , This is the square of the standard deviation of GNSS longitude / latitude. Covariance update. ,in It is an identity matrix.

[0025] Simultaneously, physical constraint thresholds are set to filter out outliers caused by signal interference. In a specific embodiment, the Euclidean distance between the current positioning point and the previous valid point is calculated. and time difference If the calculated instantaneous velocity Exceeding 180km / h, or Displacement at 1s If the distance exceeds 50 meters, the point is considered an abnormal jump and is removed. In addition, the uniqueness of the uploaded data is verified based on the timestamp and message ID. If the unique identifier ID of the current message is the same as the data ID received in the previous frame and the timestamp is consistent, it is determined to be a duplicate invalid record caused by network retransmission and is discarded.

[0026] Step S103: Combining road network topology information, a map matching algorithm is used to perform trajectory correction processing on the cleaned positioning data to correct positioning drift caused by signal obstruction.

[0027] Here, a path matching mechanism based on a Hidden Markov Model (HMM) can be introduced to address the location point drift phenomenon that occurs when vehicles pass under overpasses, in tunnels, or in densely populated areas with tall buildings (such as the location point falling into non-road areas). By calculating the geometric distance and driving direction matching degree between the location point and surrounding road segments, the drifting point is forcibly mapped and absorbed to the road centerline with the highest probability. A probabilistic graphical model containing observed states and hidden states is constructed; in a specific embodiment, the GPS positioning trajectory point sequence is defined as the observation sequence. ,in

[0028] Let t be the latitude and longitude coordinates; the candidate road segment sequence in the road network is defined as the hidden state sequence.

[0029] ,in Let t be the actual road segment where the vehicle is located at time t. Indicates The set of neighboring road segments centered at a radius r = 50m.

[0030] First, calculate the emission probability. Used to characterize positioning points Located on the road section The geometric likelihood is calculated using the following formula: ,in Indicates the location point To candidate road sections Vertical projection distance of the centerline The standard deviation of the positioning error is set (e.g., 10 meters). Next, the transition probability is calculated. Used to characterize the vehicle's movement from the previous time segment of the road Move to the current road segment The topological rationality is calculated using the following formula: ,in The great circle distance between two consecutive positioning points. For road segments in the road network topology arrive The shortest path length, These are the road network constraint parameters. Finally, the Viterbi algorithm is used for dynamic programming to solve for the global posterior probability. The largest optimal state sequence The recursive formula is: Observation points at each time point Vertical projection to the corresponding optimal road segment On the center line, the corrected high-precision trajectory points are obtained, thereby correcting the positioning deviation caused by the multipath effect.

[0031] Step S104: Synchronize the positioning data with the sensor data in time and match them spatially to generate a standardized high-precision vehicle trajectory sequence. In a specific embodiment, to address the inconsistent sampling frequencies of multi-source data, the GNSS module uses... The frequency outputs positioning data. The IMU sensor outputs inertial data at a frequency greater than or equal to 100Hz, and the CAN bus outputs vehicle status data at a frequency greater than or equal to 10Hz. Using GNSS satellite timing signals as a unified clock source, timestamps are applied to the data packets from each sensor, and a linear interpolation algorithm is used to resample the IMU and CAN data. Frequency is used to achieve time alignment of multi-source data rather than synchronous triggering. Simultaneously, vehicle operating attributes acquired by sensors (such as engine speed and instantaneous fuel consumption) are precisely mapped spatially to geographic coordinates, ultimately outputting a standard trajectory point sequence containing a unified UTC timestamp, corrected latitude and longitude, and full state information. For example, on a city expressway, a vehicle equipped with a smart in-vehicle terminal begins driving at 8:30 AM on January 15, 2024. The vehicle carries three types of data sources: a GNSS positioning module outputs latitude and longitude coordinates and speed information once per second; an IMU inertial sensor records vehicle acceleration 100 times per second; and a CAN bus collects operating data such as engine speed and fuel consumption every 100 milliseconds. These three devices use different time bases: GNSS uses satellite-synchronized UTC time, while the IMU and CAN bus rely on local crystal oscillators, resulting in varying degrees of clock deviation. After starting, the vehicle maintains a constant speed of 65 km / h. In the first two seconds, the engine remained stable at 2500 rpm, with fuel consumption of approximately 8.5 liters per hour and longitudinal acceleration close to zero, indicating smooth road conditions. In the third second, the driver began braking, an action keenly captured by the IMU sensor, recording a maximum deceleration of 3.2 m / s². The CAN bus also showed the engine speed plummeting to 1800 rpm, fuel consumption significantly reduced, and GNSS speed decreasing from 65 km / h to 45 km / h. In the fourth second, deceleration continued to 20 km / h, and in the fifth second, the system came to a complete stop, with the engine idling at 750 rpm, and fuel consumption dropping to 0.5 liters per hour. Due to significant differences in the sampling frequencies of the three data types, the system required time synchronization. First, using GNSS UTC time as an absolute reference, a clock deviation of 200 milliseconds ahead for the CAN bus and 150 milliseconds ahead for the IMU was calibrated. Then, linear interpolation was used to downsample the 10 Hz CAN data to 1 Hz, and the average and peak acceleration within a one-second window were calculated for the 100 Hz IMU data. After spatiotemporal alignment, six standard trajectory points were generated. Each point contains a unified UTC timestamp, corrected road centerline coordinates, and fused full-state information including speed, engine speed, fuel consumption, and acceleration. Particularly noteworthy is the trajectory point at the third second, which simultaneously recorded a speed of 45 km / h, an engine speed of 1800 rpm, fuel consumption of 6.2 liters per hour, and an average deceleration of -2.85 meters per second squared. This multi-source data corroborated each other, accurately reconstructing the critical driving event of braking initiation, providing comprehensive data support for subsequent risk identification.

[0032] Step S2: Based on the standardized high-precision vehicle driving trajectory sequence, redundant data is removed using feature point extraction algorithms and adaptive compression mechanisms, and a multi-dimensional spatiotemporal index is constructed to generate a lightweight and easily searchable storage-level trajectory set. Figure 3 This is a schematic flowchart of the method for generating a lightweight and easily searchable storage-level trajectory set disclosed in this embodiment, specifically combined with... Figure 3 It can be seen that this step specifically includes: Step S201: Utilize a feature point extraction algorithm to traverse the standardized driving trajectory sequence, identifying and retaining key feature points such as steering, speed changes, and parking, ensuring that the key semantics of the trajectory are not lost. In a specific embodiment, a multi-dimensional feature point adaptive extraction algorithm that integrates heading angle change rate detection, speed gradient monitoring, and spatiotemporal dual threshold judgment is used to scan and analyze the driving trajectory sequence point by point. This algorithm first initializes the feature point set and includes the starting and ending points of the trajectory sequence into the set, and then traverses and logically judges the intermediate sequence points one by one: for steering features, calculate the current trajectory point. Compared with the previous preserved feature point The absolute value of the difference in heading angle between the points is used to determine if the difference exceeds a preset steering threshold (e.g., set to 15°), or if the local radius of curvature formed by three consecutive trajectory points is less than the set value. In this case, the vehicle is determined to have undergone a significant change of direction, and the current point is adjusted accordingly. Key steering points are marked and retained; for gear shifting characteristics, the velocity gradient within adjacent time windows is monitored in real time, and instantaneous acceleration is calculated using differential calculations. ,like

[0033] If the acceleration exceeds a preset threshold, the point is captured to record the rapid acceleration / deceleration behavior. For parking characteristics, the vehicle's parking state is determined by combining spatial displacement radius and time span thresholds. If the vehicle is detected parking within a continuous time period... If the displacement fluctuation range within the area is always within the geofence with the starting point as the center and a preset radius R, it is determined to be a stationary event. Only the center point or the beginning and end points of this time period are extracted as representative nodes. In this way, while filtering out non-critical data generated by straight-line uniform speed driving, the core nodes reflecting the vehicle's driving intention are accurately retained, ensuring that the key semantics of the trajectory are not lost.

[0034] For example, in a specific embodiment, an improved strategy based on a sliding window or the Douglas-Peucker algorithm is used to perform point-by-point scanning and feature discrimination on the trajectory sequence. A vehicle equipped with an intelligent in-vehicle terminal departs from point A and travels to point B, with the entire journey lasting 1000 seconds. The in-vehicle system records 1000 original trajectory points at a frequency of once per second. The system first enters a highway section. In the first 600 seconds, the vehicle maintains a stable speed of approximately 110 kilometers per hour, with the heading angle fluctuation consistently less than 3 degrees and the acceleration not exceeding 1 meter per second squared. Since this section of the journey has neither significant steering nor obvious speed changes, nor any stops, it is determined that this section lacks key driving semantics. Only the start point, end point, and sampling anchor points set every 60 seconds are retained, compressing the 600 original points into 12 feature points. After exiting the highway and entering the ramp turning area, the vehicle's heading angle gradually changed from 0 degrees to 90 degrees within 30 seconds from 600 to 630 seconds, and its speed decreased from 110 km / h to 40 km / h. The system detected a peak heading change rate of 20 degrees per second, exceeding the preset threshold of 15 degrees. Although the peak deceleration of 2.5 m / s² did not reach the 3 m / s² threshold, it was close to the critical value. Considering the changes in motion state, three key nodes were retained: the straight-line driving point before the turn at 600 seconds, the turning apex at 615 seconds with a heading of 45 degrees, and the trajectory point that stabilized again after the turn at 630 seconds. Subsequently, the vehicle entered a city intersection and encountered a red light and stopped within 70 seconds from 630 to 700 seconds. From 635 to 665 seconds, the vehicle remained stationary within a 10-meter range. Although the duration did not reach the 3-minute dwell time threshold, the system, combined with the ACC ignition status, determined it to be a valid dwell event. After the green light turned on, the vehicle started moving, and the system detected an acceleration of 4 meters per second squared, exceeding the speed change threshold. This section ultimately retained four feature points: the starting point of the stop at 635 seconds, the center of the dwell time at 650 seconds, the point of rapid acceleration at 665 seconds, and the trajectory point after the speed stabilized again at 700 seconds. In the last 300 seconds, the vehicle entered the parking lot, hovering at low speed searching for a parking space, and finally stopped at approximately 850 seconds, maintaining an idle speed for 20 minutes. Although this journey included several minor turns, the system determined it as a significant dwelling feature based on a 30-meter spatial range and a 20-minute dwell time, and also selected three key turning points at 720 seconds, 780 seconds, and 820 seconds to record the vehicle's path. This section was compressed from 300 original points to 5 feature points, including the entry point, the center of the dwell time, and the three key turning points. The original 1,000 trajectory points were reduced to 24 feature points, but the key driving behavior semantics were completely preserved: the continuity of high-speed cruising, the geometry of ramp turns, the dynamic changes of starting and stopping at intersections, and the spatiotemporal characteristics of parking lot stays were all accurately captured, providing sufficient data support for subsequent risk identification and trajectory reconstruction.

[0035] Step S202: Based on the adaptive compression mechanism, redundancy removal is performed on intermediate points in smooth road sections or stationary states, significantly reducing the amount of data while ensuring trajectory fitting accuracy. In a specific embodiment, the vertical distance limit method or the simultaneous Euclidean distance (SED) algorithm is applied to perform secondary filtering on the non-feature point sequence. For smooth road sections where vehicles travel in a straight line, a recursive segmentation strategy is adopted. First, the first and last points of the trajectory segment are connected to form a reference straight line. The vertical Euclidean distance from each intermediate sampling point to this straight line is calculated. If the calculated maximum vertical distance is less than a preset accuracy threshold (e.g., 5 meters), the intermediate point group is determined to be linear redundant data and is removed as a whole, retaining only the first and last key points. Conversely, if the maximum distance exceeds the threshold, the extreme point is retained as a new segmentation node, and the above process is repeated for the generated sub-segments to ensure fitting accuracy. When a vehicle is in an ACCOFF or stationary state for an extended period of time, the system monitors the sensor status and speed values ​​in real time. When it detects that the vehicle has been stationary within the same geographical coordinate range (e.g., a radius of less than 10 meters) for more than a set threshold (e.g., 180 seconds), it triggers a timing merging logic. Only snapshot data of the start and end times of the stationary period are recorded, and redundant duplicate positioning messages in the middle are discarded. This effectively reduces data storage costs while ensuring the spatiotemporal continuity of trajectory playback.

[0036] Step S203: Establish a multi-dimensional spatiotemporal index for the compressed trajectory data, and optimize data retrieval efficiency through spatial gridding and time sharding techniques. In a specific embodiment, a spatial filling curve (such as GeoHash or Google S2 algorithm) is used to map two-dimensional latitude and longitude coordinates into one-dimensional grid codes. The continuous geographic space is divided into hierarchical discrete grids through a recursive binary search method. The latitude and longitude values ​​of trajectory points are converted into binary strings and then alternately merged or mapped by Hilbert curves to achieve dimensionality reduction storage of high-dimensional geographic data. On this basis, a hierarchical index structure is designed in conjunction with the time dimension to construct a composite index key that includes entity identifiers, time windows, and spatial codes. For example, a composite primary key (RowKey) in the form of "vehicle ID reverse_time period_GeoHash code" is designed. The unique vehicle ID is reversed to achieve hashing, preventing database write hotspots (RegionHotspot) caused by consecutive serial numbers. The time period is set to hourly or daily granularity to guide the physical sharding of data storage. When retrieving the trajectory of a specific vehicle in a certain time period, the system uses the database's RowKey prefix matching mechanism (PrefixFilter) or sets the StartRow and StopRow scanning range to quickly filter irrelevant data blocks and directly locate the target storage unit, avoiding the performance loss caused by full table scans. Thus, the single query response time is controlled to the millisecond level when dealing with hundreds of millions of data points.

[0037] A certain online car-hailing platform has 100,000 operating vehicles, generating hundreds of millions of trajectory data every day. It is necessary to store them efficiently and query the historical trajectories of any vehicle quickly. Spatial grid processing is carried out. The vehicle passed through the Science and Technology Park area of Nanshan District, Shenzhen on a certain day, and the latitude and longitude coordinates of a certain trajectory point are 113.9436 degrees east longitude and 22.5489 degrees north latitude. The system uses the GeoHash algorithm to map the two-dimensional coordinates into a one-dimensional grid code: first, encode the latitude of 22.5489 degrees into the binary string "1010110001", and encode the longitude of 113.9436 degrees into "1101010100", then merge the alternating bits to get "1110010011110000", and finally convert it into the Base32 string "wsqq7". This code represents a geographical grid of approximately 500 meters × 500 meters, and all the trajectory points of the vehicle within this area share the previous prefix code, achieving the dimensionality reduction storage of high-dimensional geographical data. Build a composite index key. For the trajectory data of the vehicle from 8:00 to 9:00 on January 15, 2024, the system generates the RowKey as "54321B Yue_2024011508_wsqq7". Among them, "54321B Yue" is the string reversal of the original license plate "Yue B12345", which scatters the consecutive license plate numbers into different hash intervals to avoid hot spots formed in the same database partition when a large number of vehicles are written simultaneously; "2024011508" represents the time period accurate to the hour level, guiding the data to be physically sharded and stored by hour; "wsqq7" is the aforementioned spatial grid code. This three-segment structure of "vehicle ID reversal_time period_GeoHash code" makes the relevant data dispersed and aggregated in distributed storage, taking into account both writing balance and query efficiency. When the operation management personnel need to retrieve the driving trajectory of the vehicle from 8:30 to 8:35 on that day, the system performs an efficient search: first, convert the query time range into the RowKey prefix "54321B Yue_2024011508", and use the PrefixFilter mechanism of HBase to quickly filter out the data blocks of other hours and other vehicles; at the same time, calculate the StartRow as "54321B Yue_2024011508_00000" and the StopRow as "54321B Yue_2024011508_zzzzz" according to the time range to further limit the scanning range. Since the data has been sharded by hour, the system only needs to scan a small number of data blocks within a single Region, without touching the massive historical records of the other 90,000 vehicles. The single query response time is controlled within 50 milliseconds, which is two orders of magnitude faster than the several-second delay of a full-table scan. The finally returned trajectory data is organized according to the GeoHash spatial proximity, which is convenient for directly rendering on the map in the order of the driving path without additional sorting.

[0038] Step S204: Logically encapsulate and physically serialize the compressed trajectory key point data and the constructed spatiotemporal index information to generate a compact, storage-level trajectory dataset with fast retrieval capabilities. In a specific embodiment, efficient serialization protocols such as Protocol Buffers or Avro are used. First, a data structure schema is defined, encapsulating the vehicle's unique identifier ID, the spatiotemporal index key (RowKey), and the differentially encoded list of trajectory points (specifically including the time offset relative to the reference point, latitude and longitude coordinate increments, and compression status bitmask) into a compact binary data block. Then, this binary data block is written to the corresponding cell in a distributed columnar database (such as HBase or Cassandra), where the RowKey uses the format "vehicle ID reversed_time window". The design avoids writing hotspots; simultaneously, it generates header information containing metadata descriptions (such as data version number, compression algorithm identifier, and integrity check code) at the beginning of the data block, and finally outputs a standardized storage-level trajectory set object; in a specific embodiment, this object is instantiated as a serialized entity containing "Header" and "Payload" fields, where the "Header" field stores the version identifier and check code required for parsing, and the "Payload" field encapsulates the compressed trajectory point binary data stream. This object can be directly read and parsed by subsequent risk control model modules through the API interface without repeated preprocessing.

[0039] Step S3: Input the lightweight, storage-level trajectory set into the pre-trained neural network risk control model, combine it with the formulaic behavior parameter model to extract abnormal driving behavior features and match risk patterns, and output a structured risk control assessment report containing risk level and abnormality type. Figure 4 This is a schematic diagram of the method for extracting abnormal driving behavior features and matching risk patterns disclosed in this embodiment, specifically combined with... Figure 4 It can be seen that this step specifically includes: Step S301: Load the lightweight storage-level trajectory set into the analysis and computing engine, and perform data format conversion to adapt to the requirements of the model input layer. In a specific embodiment, a parallel data processing pipeline is built using a distributed computing framework (such as Spark or Flink). The storage-level trajectory set is read in batches from the columnar database according to the RowKey range by configuring the TableInputFormat interface; the deserialization interface (such as ProtobufParser) is called to decode the binary data block to restore the original trajectory point sequence containing longitude, latitude, instantaneous velocity, heading angle, timestamp, and ACC status bit; then feature vectorization and normalization processing are performed, selecting velocity, acceleration, and heading angle change rate as key feature dimensions, using the Z-Score normalization algorithm to make the continuous numerical features dimensionless, and performing One-H normalization on discrete features such as ACC status. Ot encoding is used to map discrete trajectory points into a multi-dimensional feature matrix. Based on this, according to the input layer specifications of the pre-trained neural network model, a sliding window mechanism is used to slice long trajectory sequences. The time step is set to 128 and the stride is set to 64 to enhance sample coverage. For time slices with a length of less than 128, a zero-padding strategy is used to pad the end of the sequence with zeros to unify the length. Finally, a standard input data stream conforming to the model tensor format (i.e., a three-dimensional array with the shape of BatchSize×128×FeatureDimension) is constructed and efficiently delivered to the inference engine through shared memory or TFRecord format.

[0040] Step S302: Extract deep nonlinear spatiotemporal features of the trajectory using a pre-trained neural network model in parallel, and calculate driving behavior indicators based on physical rules by combining a formulaic behavior parameter model.

[0041] In a specific embodiment, on the one hand, a deep learning feature extraction channel is constructed, and the adapted trajectory tensor is input into a pre-trained bidirectional long short-term memory network (Bi-LSTM). A deep neural network architecture containing an input layer, two stacked Bi-LSTM layers, and a fully connected output layer is built based on a deep learning framework (such as TensorFlow or PyTorch) to extract high-dimensional features from time-series trajectory data. The deep neural network architecture is obtained through the following steps: (1) Collect no less than 500,000 real driving trajectories as a training set, and divide them into training / validation / test sets in a ratio of 8:1:1; (2) Hire more than 3 professional driving safety officers to label the trajectories for risks, including rapid acceleration, rapid deceleration, sharp turns, fatigued driving, and serpentine driving, and use majority voting to determine the final labels; (3) Construct a network architecture containing two layers of Bi-LSTM (128 hidden units) + a fully connected layer (128 neurons), and train it using the Adam optimizer (initial learning rate 0.001). Training is stopped when the validation set loss does not decrease for 5 consecutive epochs; (4) Store the trained model weight file in the vehicle terminal ROM or cloud server, and load and call it through the API interface. First, configure the input layer to receive shapes as The standardized trajectory tensor is defined, where Batch_Size is the batch size (e.g., set to 64) and 128 is the time step. The number of hidden units in both the first and second Bi-LSTM layers is set to 128, using orthogonal initialization. The strategy initializes the cyclic weight matrix to maintain gradient norm stability, and initializes the bias term to 0. During computation, for each time step... The LSTM unit performs gating operations internally: calculating the forget gate. This is used to decide whether to discard historical information. (Calculate the input gate.) With candidate state To update cell state (* indicates element-wise multiplication), and passed through the output gate. Get the current hidden state .

[0042] The model uses forward LSTM units to process the trajectory sequence in order from time step t=1 to 128, capturing the cumulative influence of historical states on the current state; it uses backward LSTM units to process the sequence in reverse order from t=128 to 1, capturing the feedback constraints of future states; then it processes the forward hidden states at the same time step (denoted as...). ) and backward hidden state (denoted as The features are concatenated along the channel dimension to form a 256-dimensional feature vector containing bidirectional contextual information. To prevent overfitting, a dropout layer with a dropout rate of 0.5 is introduced between the two Bi-LSTM layers and before the fully connected layer, and an L2 regularization term (with a regularization coefficient set to...) is added to the loss function. ).

[0043] Finally, the concatenated hidden state vector of the second-layer Bi-LSTM at the last time step is extracted and input into a fully connected layer containing 128 neurons. A ReLU activation function is used for nonlinear mapping and dimensionality compression, outputting a deep semantic feature vector with a dimension of 128. This accurately captures the long-distance temporal dependencies and nonlinear spatiotemporal correlations implicit in the trajectory sequence (such as the latent vector representation of complex patterns like frequent lane changes and serpentine driving). During the model pre-training stage, the dataset is divided into training, validation, and test sets in an 8:1:1 ratio. The Adam optimizer (initial learning rate set to 0.001, first-moment estimation exponential decay rate beta1=0.9, second-moment estimation exponential decay rate beta2=0.999) is used with a learning rate decay strategy (e.g., decaying by 0.1 every 10 epochs). Backpropagation iteration is performed using binary cross-entropy as the loss function. An early stopping mechanism is also introduced: training is automatically terminated and the optimal weight parameters are saved when the validation set loss value does not decrease significantly within 5 consecutive epochs, thus obtaining a feature extractor with strong generalization capabilities.

[0044] On the other hand, an explicit feature extraction channel based on physical rules is constructed, and a parametric driving behavior model is built based on classical Newtonian mechanics and kinematic principles. Multi-dimensional numerical calculations and statistical analyses are performed on the instantaneous velocity, acceleration, and rate of change of heading angle of the trajectory points. In a specific embodiment, the input trajectory point sequence is first preprocessed, converting the velocity units to meters per second (m / s), and a five-point cubic smoothing algorithm is applied to denoise the velocity curves to reduce acceleration calculation jitter caused by GPS positioning errors. Next, the longitudinal acceleration between adjacent trajectory points is calculated. Set a two-way threshold range (e.g., set the rapid acceleration threshold to 3 m / s² and the rapid deceleration threshold to -4 m / s²), and when it is detected... When the value exceeds the safe range, the system not only accumulates the frequency of rapid acceleration / deceleration within that time window, but also simultaneously calculates the mean and variance of acceleration under this abnormal state as longitudinal risk characteristic components; at the same time, it calculates the lateral acceleration. ,in To account for the change in heading angle, a periodic angle correction logic is introduced during the calculation process. The angle difference crossing the 0° / 360° threshold is standardized (e.g., the change from 359° to 1° is corrected to +2° instead of -358°). When the absolute value exceeds a set threshold (e.g., 0.5g), it is marked as a sharp turning behavior indicator, and the sideslip risk coefficient is calculated by weighting it with the current vehicle speed. In addition, statistics such as the percentage of vehicle speeding time, the frequency of long-term idling, and the duration of continuous driving are also collected.

[0045] Finally, in order to eliminate the differences in dimensions and distribution between physical features and depth features, and to achieve effective alignment of multi-source heterogeneous features, the explicit physical statistical feature vectors calculated above are first processed. (Assuming the dimension is) For example, 16 statistical indicators (including the mean longitudinal acceleration and the frequency of sharp lateral turns) are subjected to Min-Max normalization, and the calculation formula is as follows: All physical feature values ​​are mapped to the [0,1] interval to avoid large numerical features dominating gradient updates; subsequently, the normalized physical feature vectors are... Deep semantic feature vectors extracted by neural networks (dimension is) For example, a 128-dimensional vector is concatenated along the channel dimension to generate a vector with dimension 1. joint eigenvectors Next, a multilayer perceptron (MLP) is constructed as the feature fusion module. This module contains at least two fully connected layers (DenseLayer). The first fully connected layer has 128 neurons, uses a He normal distribution to initialize the weight matrix, and introduces the ReLU activation function to increase non-linear expressive power. The dropout rate is set to 0.3 to prevent overfitting. The second fully connected layer further maps the feature dimensions to the target space (e.g., 64 dimensions), outputting the final hybrid feature representation. Through the above process, the model can adaptively learn the complementary relationship between physical rules and deep semantics, and construct a high-dimensional hybrid feature set that takes into account both physical interpretability and data-driven generalization ability, providing robust input support for subsequent risk pattern classification.

[0046] Step S303: Characterize the extracted final mixed features. A multi-dimensional comparison is performed with a pre-set abnormal driving behavior database to identify potential risk patterns such as sudden acceleration / deceleration and route deviation using classifiers or similarity algorithms. In a specific embodiment, a dual matching engine based on rule-based judgment and vector similarity calculation is constructed. This engine adopts a hybrid decision mechanism that combines hard rule constraints with soft model matching. First, for the physical statistical features output by the formulaic model (such as the number of sudden brakings, speeding duration, and frequency of long-term idling), parameterized rule verification logic is designed, and a pre-set business rule set is loaded (e.g., threshold constraints defined in JSON format: {"indicator":"longitudinal acceleration","operator":"<","threshold":-4,"unit":"m / s²"}). The real-time calculated physical feature values ​​are logically compared item by item with the standard thresholds in the behavior database. Once the triggering condition is met, the explicit violation behavior is immediately locked and a basic risk label is generated. Second, for the deep semantic feature vectors extracted by the neural network, similarity retrieval and probability inference based on vector space are performed. The system pre-builds an abnormal driving behavior feature library, maps historically confirmed typical risk samples (such as fatigue driving trajectory clusters and malicious detour patterns) into standard feature vectors through an encoder, and stores them in a vector database (such as Faiss); during the inference phase, it calculates the cosine similarity between the semantic feature vector of the current trajectory and the sample in the feature library ( Alternatively, the data can be input into a fully connected layer and a Softmax classifier to output the posterior probability distribution of each risk category. Finally, a multi-source evidence fusion strategy is introduced to obtain a list of risk events and parameters (abnormal behavior type - trigger frequency - confidence level - spatiotemporal / feature parameters). For example, when the similarity between the trajectory feature vector of a certain time period and the "snake-like driving" sample library exceeds 0.85, and the classifier predicts "aggressive lane change" with a confidence level greater than 0.7, while the corresponding lateral acceleration statistics show high-frequency oscillations although not exceeding the limit, the system comprehensively determines that the segment matches the "dangerous lane change" risk pattern and associates the corresponding risk label with the confidence score.

[0047] Step S304: Quantitatively score the risk based on the comprehensive risk event list and parameters, determine the risk level, and label the specific anomaly type, ultimately generating a structured risk control assessment report. In a specific embodiment, a multi-factor risk scoring model based on linear weighted superposition and confidence level calibration is constructed, and the comprehensive risk scoring formula is defined as follows: Where N is the total number of detected abnormal behavior types, For the first Preset hazard weighting coefficients for abnormal behaviors (e.g., setting weights for rapidly accelerating behaviors). The weight of fatigue driving The weight of sharp turns ), The frequency of this type of behavior within the evaluation time window. The corresponding confidence probability value output by the model (range 0 to 1); the system calculates the cumulative risk score S in real time and limits its saturation to ensure that the score is within the range of [0,100]. The calculated score is judged according to the preset grade threshold matrix. A structured intelligent risk control assessment report conforming to the preset data structure (with JSON as the core serialization format) is obtained. At the same time, the report can realize message push / database persistence for downstream system calls. The complete content of the report includes two core modules: (1) metadata module: vehicle unique identifier (VIN), assessment time window (Start / EndTimestamp), comprehensive risk score S (0~100 score after linear weighting and saturation limit), risk level (safe / low risk / medium risk / high risk determined according to the score). (2) Payload business data module: Encapsulates a detailed list of abnormal behavior events identified by S303. Each abnormal behavior element in the list includes: trigger timestamp, latitude and longitude of the occurrence location, abnormal type label, feature snapshot data (such as maximum longitudinal acceleration: 4.5m / s²), and is associated with the rating-related parameters (frequency, confidence level) of the behavior.

[0048] For example, if a threshold vector T = [30, 60, 85] is set, if S < 30, it is considered "safe", if 30 ≤ S < 60, it is considered "low risk", if 60 ≤ S < 85, it is considered "medium risk", and if S ≥ 85, it is considered "high risk". Finally, a structured risk control report is generated using serialization technology. In a specific embodiment, a JSON data contract structure is defined, with the root node containing two sub-objects: metadata and payload. The metadata field stores the vehicle unique identifier (VIN), the assessment time window (Start / EndTimestamp), the comprehensive risk score (Score), and the assessment level (Level). The payload field encapsulates a detailed list of abnormal behavior events. Each element in the list includes a trigger timestamp, the latitude and longitude coordinates of the location, an anomaly type label, and corresponding feature snapshot data (such as "maximum longitudinal acceleration: 4.5m / s²"). The system publishes this JSON document to a specified topic through a message middleware (such as Kafka or RabbitMQ) or persists it to a relational table in the business database via a JDBC interface for downstream risk control dashboards or real-time alarm systems to subscribe to and access.

[0049] Step S4: Respond to external risk control query commands, and use a distributed retrieval engine to retrieve and visualize related data based on the structured risk control assessment report, outputting a real-time risk warning view of vehicle asset monitoring status. Figure 6 This is a schematic diagram of the risk warning method for vehicle asset monitoring status disclosed in this embodiment, specifically combined with... Figure 6 It can be known that this step specifically includes: Step S401: Configure the service interface to receive the risk control query request sent by the external business system in real time, parse, authenticate, and extract parameters from the request message to clarify the query target and scope. In a specific embodiment, a Web service listening module is built on the server side based on the HTTP / HTTPS application layer protocol, and a standardized RESTful API interface is exposed externally (such as defining the interface path as ` / api / risk / monitor` and using the POST request method) to receive the query instructions initiated by the risk control management platform or the third-party supervision system in real time; the system kernel calls the JSON parsing engine to deserialize the received HTTP message payload (Payload) to parse out the query parameters in the form of key-value pairs, specifically extracting the unique identifier of the vehicle to be queried (the field name is `vehicleId`, and the format is a 17-bit VIN code or license plate number) and the monitoring time window (the field names are `startTime` and `endTime`, using the Unix timestamp format). For example, when the background receives a JSON data packet with the content `{"vehicleId":"粤B12345","startTime":1698300000,"endTime":1698386400}`, the system immediately triggers the parameter verification logic chain: First, perform format and logic verification, use regular expressions to match the VIN code rules, and calculate the time difference $\Delta t = endTime - startTime$. If $\Delta t$ is less than 0 or exceeds the maximum query span allowed by the system (such as 7 days), a parameter error response is returned; Second, perform security authentication, extract the digital signature (Signature) and timestamp (Timestamp) from the HTTP request header, calculate the local hash value using the HMAC-SHA256 algorithm in combination with the pre-shared key and compare it to prevent the request from being tampered with or replayed; Only when all verifications pass, the system instantiates the above parameters into an internal retrieval context object (SearchContext), which encapsulates the filtering conditions and pagination cursor and serves as the input parameter for the subsequent retrieval engine.

[0050] Step S402: Retrieve the corresponding structured risk control assessment report from the database using a distributed search engine, and retrieve the relevant storage-level trajectory set. In a specific embodiment, an inverted index is built using a distributed search engine such as Elasticsearch, and an index mapping is configured containing "vehicle_id" (keyword type), "timestamp" (date type), and "risk_level" (integer type). Using the parsed vehicle unique identifier (VIN code) and time window as joint query conditions, a Boolean query is constructed. A TermQuery is used to precisely match the vehicle ID, and a RangeQuery is used to filter the time range. The structured risk control assessment report (including comprehensive score and abnormal event list) for that time period is quickly retrieved and extracted from the risk control result database. Simultaneously, the time periods of abnormal behavior recorded in the report are analyzed. Based on the RowKey generation rule defined in step S203 above (i.e., "reverse VIN_time period"), the corresponding HBase scan range is automatically calculated. StartRow is set to "reverse VIN_start time" and StopRow is set to "reverse VIN_end time". A range query (RangeScan) is initiated to the underlying columnar database (such as HBase) to retrieve the storage-level trajectory set data blocks corresponding to the spatiotemporal time in batches. Finally, at the service layer, a hash mapping (HashMap) or multi-threaded parallel merging strategy is used, with UTC timestamps as keys, to perform ID alignment and logical aggregation of risk labels in the risk control assessment report and high-precision trajectory data points, constructing a complete data package object containing a two-way association between "risk conclusion - trajectory facts", providing full data support for subsequent visualization and reproduction.

[0051] Step S403: Deeply integrate the aggregated risk control assessment data with the Geographic Information System (GIS), and use the front-end rendering engine to map the abstract trajectory coordinate sequence and risk indicators onto the electronic map layer to achieve intuitive reproduction of the entire vehicle driving process and highlight display of abnormal behavior. In a specific embodiment, the front-end visualization component builds a single-page application based on the Vue.js or React framework. During the DOM mounting phase, it instantiates a WebGIS map object (e.g., via the `newAMap.Map` or `L.map` interface), configures the tile layer source (TileLayer) to load the vector road network base map, and sets the initial center coordinates and scaling level. It receives the trajectory dataset and risk control report JSON object returned by the backend, iterates through the trajectory point sequence, and uses a coordinate transformation algorithm (e.g., by importing the `gcoord` library) to batch convert the original WGS84 coordinates to a coordinate system adapted to the base map (e.g., GCJ02), constructing a standardized coordinate point array. Subsequently, it parses the abnormal event set in the risk control report, uses a segmented rendering strategy to draw a trajectory polyline, and cuts the trajectory sequence into several sub-segments based on the timestamp interval. For time periods marked as "normal," the line color is configured to green (e.g., `#28C76F`) and the line width is 4 pixels. For time periods marked as "speeding" or "high risk," the line color is configured to red (e.g., `#EA5455`), the line width is thickened to 6 pixels, and a dashed line style is applied. The `map.add(overlays)` method is called to render all line segment objects to the vector layer. Simultaneously, for discrete risk events such as "sudden braking," "sharp turning," or "abnormal parking," the latitude and longitude of their occurrence locations are extracted, and Marker objects are created. The corresponding SVG icon resources are dynamically loaded according to the risk type, and a `click` or `mouseover` event listener is bound to each Marker object. In the callback function, an HTML template string containing the trigger time, instantaneous value (e.g., "deceleration: -0.8g"), and risk description is constructed. An InfoWindow object is instantiated, and the `open` method is called to pop up a details pop-up window at the corresponding coordinates on the map. This helps managers quickly locate and intuitively perceive the operational risks of vehicles in specific times and spaces.

[0052] Step S404: The rendered visualization layer and associated risk control data are encapsulated into an interactive monitoring interface, outputting a comprehensive view including real-time vehicle status, historical trajectory playback, and risk level prompts for managers to conduct in-depth analysis. In a specific embodiment, a web front-end monitoring dashboard based on HTML5 and CSS3 Flex layout is constructed, and a component-based development model is adopted to divide the interface logic into a map display main area, a vehicle attribute sidebar, and a risk event statistics bottom bar. In the main map display area, a Canvas-based timeline playback controller is integrated. It calculates time offsets by listening to mouse drag events and uses the `requestAnimationFrame` animation loop mechanism to implement trajectory playback. The system performs linear interpolation calculations on the trajectory point sequence based on the current playback timestamp, dynamically calling the map API's `setPosition` and `setAngle` methods to update the coordinates and heading of vehicle markers. This supports drag-and-drop positioning, multi-speed playback, and frame-by-frame viewing. In the attribute sidebar, a WebSocket long connection or polling mechanism is established to subscribe in real-time to the vehicle status data stream pushed from the backend. Utilizing the two-way data binding feature of the MVVM framework, DOM nodes are updated in real-time to display vehicles. The system displays the VIN code, dashboard speed value, ACC ignition status icon, and risk score value. It also includes built-in front-end threshold monitoring logic. When the risk score exceeds a set threshold (e.g., 85 points), it dynamically switches CSS class names to activate keyframe animations, triggering a flashing red warning on the interface. A modal component is instantiated to pop up a risk details card at the top level of the view, using list rendering commands to iterate and display the specific abnormal behavior sequence (e.g., "10:05 rapid acceleration," "10:08 yaw"). Furthermore, the card integrates "remote command issuance" and "report export" interactive buttons. By binding click events, it triggers asynchronous Axios requests, sending control commands containing the vehicle ID and operation code to the backend interface. This assists administrators in quickly performing remote vehicle control or evidence solidification operations for risk control reports.

[0053] This invention employs Kalman filtering for noise reduction combined with map matching techniques based on Hidden Markov Models (HMM) and Viterbi algorithms. This effectively corrects positioning drift in areas with signal obstruction, such as elevated roads and tunnels, and accurately snaps the positioning point to the road centerline, significantly improving the accuracy of vehicle trajectory data and geographical matching.

[0054] This invention employs an adaptive extraction algorithm that integrates heading angle and velocity features, combined with lightweight techniques such as recursive segmentation using vertical distance limit method and time-series merging of static states. This achieves the benefits of maximizing the elimination of linear and static redundant data while preserving key driving semantics, thereby significantly reducing the storage cost and transmission overhead of massive trajectory data.

[0055] This invention employs a composite index key design technique of "vehicle ID reversal_time period_spatial encoding", which effectively avoids the write hotspot problem of distributed databases, enables millisecond-level spatiotemporal range retrieval of hundreds of millions of trajectory data, and significantly improves query efficiency in big data scenarios.

[0056] This invention employs a technique of parallel extraction and fusion of Bi-LSTM deep features and Newtonian kinematic physical features, achieving both generalization recognition capability that takes into account complex behaviors and interpretability of physical rules, thus solving the technical pain points of single-model "black box" decision-making and poor generalization of single rules.

[0057] This invention employs a dual matching engine technique that combines hard rule constraints with soft matching based on vector similarity / classifiers. This achieves the precise identification of explicit violations and the intelligent inference of implicit and complex risks. The fusion of multi-source evidence effectively improves the accuracy and recall rate of abnormal driving behavior identification.

[0058] This invention employs a linear weighted multi-factor risk scoring model with confidence level calibration, which organically combines the objective severity and frequency of risk with the confidence level of the model's judgment, outputting continuous and quantifiable risk scores. This makes the risk level classification more scientific and more in line with actual business risk control needs.

[0059] This invention employs a multi-level architecture technique that decouples the distributed computing framework from the message middleware, achieving the benefits of supporting high-concurrency data processing for massive numbers of vehicles, with each module deployed independently and working collaboratively, possessing strong system scalability and fault tolerance, and adapting to different scale operation scenarios from ride-hailing to logistics fleets.

[0060] This invention employs GIS visualization rendering technology that links risk control reports with original trajectory data, achieving a precise mapping between risk conclusions and spatiotemporal locations. Through segmented rendering and anomaly marking, it intuitively reproduces the risk process, providing operational managers with the benefits of an efficient and intuitive risk control assessment and decision-making tool.

[0061] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

[0062] A computer device according to embodiments of the present disclosure includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0063] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the computer device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory, causing the computer device to perform all or part of the steps of the learning outcome prediction method based on learning behavior data mining of the foregoing embodiments of this disclosure.

[0064] like Figure 7 This is a schematic diagram of a computer device provided for an embodiment of the present disclosure. It illustrates a structural schematic diagram suitable for implementing the computer device in the embodiments of the present disclosure. Figure 7 The computer device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0065] like Figure 7 As shown, a computer device may include a processor (such as a central processing unit, graphics processing unit, etc.), which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) or programs loaded from storage devices into random access memory (RAM). The RAM also stores various programs and data required for the operation of the computer device. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0066] Typically, the following devices can be connected to the I / O interface: input devices, such as sensors or visual information acquisition devices; output devices, such as displays; storage devices, such as magnetic tapes or hard drives; and communication devices. Communication devices allow the computer device to communicate wirelessly or wiredly with other devices (such as edge computing devices) to exchange data. Although a computer device with various devices is illustrated, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0067] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processor, all or part of the steps of the learning outcome prediction method based on learning behavior data mining according to embodiments of this disclosure are performed.

[0068] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

Claims

1. A method for vehicle positioning and intelligent risk control analysis, characterized in that, Includes the following steps: Collect multi-source raw positioning data and vehicle status data of the vehicle, process the multi-source raw positioning data and vehicle status data to obtain vehicle driving trajectory and status data sequence; Based on the vehicle driving trajectory and state data sequence, redundant data is removed using a feature point extraction algorithm and an adaptive compression mechanism, and a multi-dimensional spatiotemporal index is constructed to generate trajectory set data. The trajectory set data is loaded into the analysis and computing engine, and the data format is converted to adapt to the input layer requirements of the neural network risk control model. The trajectory tensor, converted from data format, is input into the neural network risk control model to obtain deep nonlinear spatiotemporal features; wherein the neural network risk control model is a deep neural network architecture containing an input layer, two stacked bidirectional long short-term memory network layers, and a fully connected output layer; An explicit feature extraction channel based on physical rules is constructed, and a formulaic behavior parameter model is built based on classical Newtonian mechanics and kinematics principles. Multi-dimensional numerical calculations and statistical analyses are performed on the instantaneous velocity, acceleration, and heading angle change rate of trajectory points to obtain driving behavior indicators. The driving behavior index is subjected to Min-Max normalization to obtain the normalized physical feature vector; The normalized physical feature vector is concatenated with the deep nonlinear spatiotemporal feature to generate the final hybrid feature representation. Statistical analysis is performed on the normalized physical feature vector in the final hybrid feature representation. Parameterized rule verification logic is designed, a pre-set business rule set is loaded, and the instantaneous velocity, acceleration, and heading angle change rate of the trajectory points calculated in real time are logically compared with the standard thresholds in the behavior library. If the triggering condition is met, explicit violations are immediately locked and basic risk labels are generated. For the deep nonlinear spatiotemporal features of the neural network risk control model, similarity retrieval and probability inference based on vector space are performed; A multi-source evidence fusion strategy is introduced to obtain a list of risk events and parameters based on similarity retrieval and probability inference of basic risk labels and vector space; The risk is quantitatively scored based on a comprehensive list of risk events and parameters, the risk level is determined and the specific anomaly type is marked, and finally a structured risk control assessment report is generated. In response to external risk control query commands, the system uses a distributed retrieval engine to retrieve and visualize related data based on the structured risk control assessment report.

2. The method according to claim 1, characterized in that, The processing of the multi-source raw positioning data and vehicle status data specifically includes: A filtering algorithm is used to remove outliers, noise, and duplicate invalid data from the multi-source original positioning data to obtain cleaned positioning data. By combining road network topology information, a map matching algorithm is used to perform trajectory correction processing on the cleaned positioning data to obtain corrected positioning data. The corrected positioning data is synchronized in time and matched spatially with the vehicle status data to generate the vehicle's driving trajectory and status data sequence.

3. The method according to claim 2, characterized in that, The processing of the multi-source raw positioning data and vehicle status data further includes: The Kalman filter algorithm is applied to smooth the multi-source raw positioning data, and physical constraint thresholds are set to filter out outliers caused by signal interference. The uploaded multi-source raw positioning data and vehicle status data are uniquely verified based on timestamps and message IDs to eliminate duplicate and invalid records caused by network retransmission. To address the issue of location point drift, a path matching mechanism based on a hidden Markov model is used. By calculating the geometric distance and driving direction matching degree between the location point and surrounding road segments, the drifting point is forcibly mapped and absorbed to the road centerline with the highest probability. A unified clock source is set up, and a linear interpolation algorithm is used to resample the multi-source data to the frequency of the unified clock source to achieve time alignment. At the same time, the vehicle operating condition attributes are spatially matched with the geographical coordinates to obtain the vehicle driving trajectory and status data sequence.

4. The method according to claim 1, characterized in that, The process of generating trajectory set data based on the vehicle's driving trajectory and state data sequence, using feature point extraction algorithms and adaptive compression mechanisms to remove redundant data and construct a multi-dimensional spatiotemporal index, specifically includes: The feature point extraction algorithm is used to traverse the vehicle driving trajectory sequence, identify and retain key feature points, and ensure that the key semantics of the vehicle driving trajectory sequence are not lost. The intermediate points of the vehicle trajectory sequence in a smooth road section or a stationary state are deredundant based on an adaptive compression mechanism. A multidimensional spatiotemporal index is established for the vehicle trajectory data after redundancy removal, and data retrieval efficiency is optimized through spatial gridding and temporal slicing techniques. The vehicle trajectory data after redundancy removal and the constructed spatiotemporal index information are logically encapsulated and physically serialized to generate a compact trajectory set data with fast retrieval capabilities.

5. The method according to claim 4, characterized in that, The step of using a feature point extraction algorithm to traverse the vehicle trajectory sequence, identify and retain key feature points, and ensure that the key semantics of the vehicle trajectory sequence are not lost specifically includes: A multi-dimensional feature point adaptive extraction algorithm that integrates heading angle change rate detection and velocity gradient monitoring is used to scan and analyze the vehicle's driving trajectory sequence point by point to obtain key feature points for steering, gear shifting, and parking. The vehicle's parking status is determined by combining the spatial displacement radius and time span thresholds. If it is determined to be a stationary parking event, only the center point or the first and last points are extracted as representative nodes.

6. The method according to claim 4, characterized in that, The redundancy removal process for the intermediate points of the vehicle trajectory sequence on smooth road sections or in stationary states based on the adaptive compression mechanism specifically includes: For the smooth road segment where the vehicle travels in a straight line, the vertical Euclidean distance from the midpoint to the reference line of the first and last points is calculated by a recursive segmentation strategy. Whether the intermediate redundant points are removed is determined based on whether the maximum vertical Euclidean distance exceeds a preset threshold. If the threshold is exceeded, the extreme point is retained and the operation is repeated for the generated sub-segment. For vehicles that are stationary for extended periods, the system monitors the state data sequence and vehicle speed values, retaining only the start and end snapshot data of the stationary period and discarding redundant intermediate positioning messages.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the vehicle positioning and intelligent risk control analysis method as described in any one of claims 1-6.

Images