Commercial vehicle route risk assessment method based on spatio-temporal coupling of multi-source sensing data

Abstract

The application provides a commercial vehicle route risk assessment method based on multi-source sensing data space-time coupling, comprising: cutting a transport route into continuous road sections and establishing a route archive; collecting static attribute data of each road section and dynamic sensing data of vehicle operation within a statistical period; performing cleaning and differential time window deduplication on alarm events; associating alarm events to specific road sections through a map matching algorithm and calculating per-100-kilometer normalized index values, while aligning time progress of alarm events within a trip with road section expected driving time windows to generate fatigue accumulation stage markers; calculating a route risk coefficient by performing multi-dimensional weighted summation after scoring each sub-index using a linear decay scoring model; performing independent safety verification on road sections to correct the risk coefficient; determining a route risk level according to the corrected risk coefficient and generating a safety management strategy. The application realizes space-time fusion of vehicle-mounted dynamic sensing data and road static attributes.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and logistics safety technology, and in particular to a method for assessing the risk of commercial vehicle routes based on the spatiotemporal coupling of multi-source sensor data. Background Technology

[0002] Risk assessment of commercial vehicle transportation routes is a core aspect of logistics fleet safety management. Currently, Driver Monitoring Systems (DMS) and Advanced Driver Assistance Systems (ADAS) are widely deployed in commercial vehicles, capable of real-time monitoring of driver fatigue, distraction, and events such as sudden deceleration and lane departure. Simultaneously, road infrastructure databases based on Geographic Information Systems (GIS) provide static road information such as road grade, curve curvature, and longitudinal slope. However, these two types of data are processed through different data chains in existing technologies: DMS and ADAS alarm data are stored at the vehicle or driver level, uploaded to the fleet management platform via onboard terminals, and are typically used only for single-vehicle, single-driver driving behavior scoring or real-time alerts, without being associated with road geographic location information; GIS road data, on the other hand, is stored independently in the map service provider's database and is used only for route planning and static road condition queries.

[0003] Because the two types of data differ fundamentally in storage dimensions, coordinate reference systems, and data sampling frequencies—DMS / ADAS alarm data is indexed by the vehicle and recorded in a vehicle motion coordinate system with a sampling frequency at the millisecond level; GIS road data is indexed by geographic latitude and longitude and recorded in a geodetic coordinate system with an update frequency at the grade level—current technologies lack the means to unify these two heterogeneous data into a single spatiotemporal reference system and perform fusion processing. Specifically, a transportation route traverses multiple administrative divisions and road sections of different grades. The road conditions encountered by the same vehicle at different stages of its journey vary significantly, and scattered individual vehicle alarm records cannot establish spatial attribution to specific road sections on the route. Furthermore, static GIS road data cannot reflect the dynamic risk signals generated by vehicles during actual driving. Therefore, current technologies cannot aggregate massive amounts of DMS / ADAS vehicle-mounted sensor alarm data by route segment dimension, nor can they integrate dynamic driving behavior data with static road characteristics into the same evaluation framework for fusion analysis.

[0004] This technological deficiency leads to the following dilemmas for fleet managers in actual operation: on the one hand, a large amount of collected vehicle sensor alarm data can only be used for the assessment of individual drivers and cannot be reused for risk assessment at the route level; on the other hand, the decision to select transportation routes still mainly relies on the manager's experience-based judgment on static information such as road grade and mileage, and lacks quantitative assessment methods based on actual operating data. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for assessing the risk of commercial vehicle routes based on the spatiotemporal coupling of multi-source sensor data.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for assessing the risk of commercial vehicle routes based on the spatiotemporal coupling of multi-source sensor data includes the following specific steps: S1. Definition of Evaluation Unit: S11. Route planning: The system obtains the origin and destination coordinates of the transportation task input by the user, calls the map navigation service interface, plans and outputs a recommended driving route from the origin to the destination, generates a complete geometric trajectory of the route consisting of an ordered sequence of latitude and longitude coordinates, and records the total mileage of the route. If the map service returns multiple candidate routes, the user confirms the target route for this evaluation through the interactive interface, and the system uses the confirmed route as the processing object for all subsequent steps.

[0007] S12, Road segmentation: The ordered coordinate sequence along the route is used to divide the entire route into multiple continuous segments based on road infrastructure characteristics and assessment accuracy constraints. The segmentation locations include, but are not limited to: (1) Road grade change points: The connection points between different road grades such as expressways and national highways, national highways and provincial highways, etc., need to be independently assessed because there are essential differences in the design standards, traffic capacity and risk characteristics of different road grades; (2) Administrative division boundaries: The boundaries of district and county-level administrative regions are divided according to the administrative regions, since meteorological and environmental data and historical accident data are based on administrative regions as the basic statistical units, so it is convenient to match subsequent data. (3) Start and end points of large structures: The entrance and exit locations of tunnel groups and large bridges are significantly different from ordinary road sections due to their internal environment (lighting, signal coverage, and refuge space); (4) Location of service areas and rest areas: As the interruption point of driver fatigue accumulation, they have a key impact on the assessment of operational load.

[0008] Between the aforementioned feature points, the length of a single road segment shall not exceed 20km. If the distance between adjacent dividing points exceeds this threshold, supplementary dividing points shall be inserted with a distance not exceeding 20km. If a road segment is less than 3km in length and has the same road grade as its adjacent road segment, it shall be merged into the adjacent road segment to avoid insufficient mileage within the statistical period and loss of statistical significance of normalized indicators due to excessively short road segments.

[0009] S13. Road segment attribute recording and route file creation: Generate a unique and permanent road segment number for each segmented road segment (coding rule: route number + segmentation sequence number), and record relevant static attributes, including road segment ID, starting point latitude and longitude, ending point latitude and longitude, road segment mileage, road grade, design speed, and number of lanes; establish a route file, recording the route number, route name, starting point name, ending point name, total mileage, list of administrative regions passed through (district / county level), and the sequence of road segment IDs contained in the route (sorted by driving direction).

[0010] S2, Multi-source data acquisition: S21. Static infrastructure data collection: Using the road segment list and route file output by S1 as indexes, static attribute data for each road segment is collected from the following data sources: (1) Road geometric feature data: By calling map service interfaces (such as the Gaode Maps Road Geometry Query API) and the road survey database, the following attributes of each road segment are obtained: road grade, number of lanes, design speed, number of curves and minimum turning radius, maximum longitudinal slope, number and total length of tunnels, number and total length of bridges, number of interchanges, number of intersections, and number and location of speed limit change points. The above data is stored by road segment dimension and corresponds one-to-one with the road segment IDs output by S1. (2) Traffic flow characteristic data: Obtain historical traffic flow data for each road segment within the statistical period from the traffic management department's data interface or navigation service platform (such as Gaode Traffic Platform), including: average daily traffic volume (vehicles / day), average daily number of large vehicles (vehicles / day), average weekly number of congestion events, and the location and number of accident black spots; (3) Meteorological and environmental data: Meteorological statistics for at least three consecutive calendar years were obtained from the meteorological service data interface for each district and county along the route, including: average annual number of rainy days (daily precipitation ≥10mm), average annual number of foggy days (visibility <200m), average annual number of freezing days (daily minimum temperature <0℃ and accompanied by precipitation), average annual number of high-temperature days (daily maximum temperature ≥35℃), and monthly meteorological index fluctuation values ​​(used to calculate the seasonal risk fluctuation index). (4) Historical accident data: Obtain historical traffic accident records for the past 12 months for each section along the route from the public security traffic management department or insurance claims platform, including: number of accidents, accident type (minor / general / major / extremely serious), severity of injuries or fatalities, and accident cause classification. If a section has no accident records during the statistical period, this field should be set to 0 and must not be left blank. (5) Communication signal coverage data: Obtain the mobile communication signal coverage status of each section along the line from the network coverage data interface of the telecommunications operator, and mark the start and end positions and continuous length of the signal coverage blind spots.

[0011] S22. Dynamic vehicle-mounted sensor data acquisition: Using the route files output by S1 as the search scope, extract the original sensor records of all vehicles that have run on the route within the statistical period from the fleet operation platform data warehouse. (1) DMS alarm data: Retrieve raw records of the following alarm events from the Driver Monitoring System (DMS) backend database: fatigued driving, closed eyes, yawning, excessive fatigue, inattention, frequent head-down movement, mobile phone use, camera obstruction, and camera rotation. Each record must include: vehicle ID, driver ID, alarm type, trigger timestamp, GPS coordinates at the time of triggering, and vehicle speed at the time of triggering.

[0012] (2) ADAS alarm data: Retrieve raw records of the following types of alarm events from the Advanced Driver Assistance System (ADAS) backend database: sudden deceleration, left lane departure, right lane departure, forward collision warning, and close proximity warning. Each record must include: vehicle ID, driver ID, alarm type, trigger timestamp, GPS coordinates at the time of triggering, and vehicle speed at the time of triggering.

[0013] (3) Vehicle alarm data: Extract the original records of overspeed alarm events from the vehicle CAN bus data recording system or the vehicle terminal (T-Box). Each record must include: vehicle ID, driver ID, overspeed start time stamp, overspeed end time stamp, average GPS coordinate trajectory during the overspeed period, and maximum instantaneous vehicle speed.

[0014] (4) GPS trajectory data: Extract the complete trajectory point sequence of each vehicle for each run within the statistical period from the vehicle-mounted GPS terminal or T-Box. Each trajectory point must include: timestamp, latitude and longitude coordinates, instantaneous vehicle speed, and cumulative mileage.

[0015] S3. Data Cleaning and Deduplication: S31. Data quality cleaning: The following cleaning rules are applied to the static attribute dataset and dynamic sensing data stream output by S2 to remove unqualified data records and ensure the data quality for subsequent evaluation processes. S311, Static attribute data integrity verification: Each static attribute of each road segment output by S2 is checked for null values, and missing values ​​are handled according to different cases. Among them, road grade, road segment mileage, and design speed are mandatory fields. If they are missing, the road segment will be marked as "invalid data" and will not participate in subsequent scoring calculations. If geometric feature fields such as curve density, longitudinal slope, tunnel-bridge ratio, interchange density, and speed limit change frequency are missing, an attempt will be made to supplement them through the map API. If the supplementation fails, the sub-indicator will be marked as "insufficient data" for the road segment and will not participate in the subsequent scoring of the corresponding sub-indicator. For meteorological data items (such as the average number of severe weather days per year), if data for a certain district or county is missing, the average value of the same indicator in adjacent districts and counties will be used as a substitute, and "interpolation substitution" will be marked in the evaluation report.

[0016] S312. Verification of the legality of dynamic sensor data records: For each alarm record and GPS track point output by S2, the following checks are performed. If any check fails, the entire record is discarded and will not be considered a valid sample for subsequent calculations: The completeness of required fields is verified. The judgment rule is that the vehicle ID, alarm type, trigger timestamp, GPS coordinates, and vehicle speed must all be non-empty. If any of these fields are empty, the record is removed.

[0017] The validity of GPS coordinates is verified by determining whether the GPS coordinates are within the range of latitude [15°N, 55°N] and longitude [72°E, 136°E]. If the coordinates are not within this range, the record is removed.

[0018] The logic of the timestamp is verified. The judgment rule is that the trigger timestamp is within the start and end dates of the statistical period. If it is not within the range, the record is removed.

[0019] The vehicle speed is checked for logical consistency, and the rule is that the vehicle speed must be within the maximum speed limit for commercial vehicles; if it is not within this range, the record is discarded. S313, Vehicle mileage reliability verification: For each vehicle included in the statistics, calculate the total mileage of its single trip and compare it with the total mileage of the route recorded in S1. If the deviation between the single trip mileage and the total route mileage exceeds ±20%, all alarm data and GPS trajectory data for that trip will be marked as "abnormal" and will not be included in subsequent aggregation calculations.

[0020] S314, Section mileage determination: For each road segment, the sum of the effective mileage of all valid vehicles within the statistical period is calculated. If this value is less than 10km, the road segment is marked as "insufficient data", all its sub-indicators are removed in subsequent scoring calculations, and the system also marks the road segment as "insufficient sample size, low confidence level" in the evaluation report.

[0021] S32. Deduplication via Differentiated Time Window: The deduplication rule is as follows: for the same vehicle and the same alarm type, if the interval between the trigger timestamp of the later alarm and the timestamp of the previous valid alarm is less than or equal to the corresponding time window value, then the later alarm is merged into the previous one and counted as the same valid event; if the interval is greater than the time window value, then the later alarm is counted as a new independent valid event.

[0022] Specifically, the time window settings and technical basis for each alarm type are as follows: If the alarm type is a DMS driving status type, such as fatigued driving, closing eyes, yawning, excessive fatigue, distraction, frequent head-down, or playing with a mobile phone, the time window is set to 5 seconds. The technical basis for setting the window is that the human fatigue state has physiological continuity. The duration of a single fatigue sign (such as closing eyes) is usually 1 to 3 seconds. When the time interval between consecutive alarms is less than 5 seconds, it is considered to be a repeated trigger of the same physiological event. If the alarm type is DMS device status, such as camera obstruction or camera twisting, the time window is set to 10 seconds. The technical basis for setting the window is that abnormal device status is usually caused by a physical operation or fault, and continuous alarms within 10 seconds are repeated reports of the same event. If the alarm type is ADAS emergency deceleration, the time window is set to 10 seconds. The technical basis for setting the window is that the mechanical completion cycle of a single braking behavior (from the driver pressing the brake pedal to the vehicle fully responding and returning to stability) is about 6 to 10 seconds. Continuous alarms within this window are multiple triggers of the same braking event. If the alarm type is ADAS lane departure, the time window is set to 5 seconds. The technical basis for setting the window is that the single correction cycle from lateral deviation of the vehicle to returning to the normal trajectory is about 3 to 5 seconds. Continuous alarms for deviation in the same direction within 5 seconds belong to the same deviation event. If the alarm type is ADAS forward collision warning or too close following distance, the time window is set to 10 seconds. The technical basis for setting the window is that the continuous monitoring cycle of following distance is related to the driver's reaction time. Continuous warnings within 10 seconds are repeated triggers of the same dangerous following scenario. If the alarm type is speeding, no time window is set. The time from the start of speeding to the return of speed to normal is counted as 1 time. If speeding is a continuous violation, the complete time period during which the speed continuously exceeds the speed limit is counted as 1 time. Instead of merging by time window, the alarms are merged by continuous status segments.

[0023] S4. Data Spatiotemporal Fusion and Normalization: S41. Spatial attribution determination based on HMM map matching: In scenarios such as urban elevated roads running parallel to ground-level roads, tunnel entrances and exits, and mountain curves, GPS positioning for ordinary freight vehicles suffers from a random drift error of 10-30 meters. Relying solely on shortest Euclidean distance matching is prone to mismatches (e.g., misclassifying a vehicle on an elevated road as a ground-level road). This invention employs a Hidden Markov Model (HMM) for map matching, transforming road segment matching into a sequence decoding problem with topological constraints, as detailed below: S411. Candidate road segment selection (observation probability calculation): GPS track point sequences sorted by time for the same trip For each trajectory point Using the latitude and longitude of this point as the center and a search radius of 50 meters (this radius corresponds to the upper limit of the horizontal positioning accuracy range of commercial GPS in open environments), all candidate road segments falling within this radius are retrieved from the S1 road segment list to form a candidate road segment set. ( (The number of road segments within that radius). For each candidate road segment Calculate the observation probability Its meaning is "if the vehicle is actually located on the road section". GPS point observed The probability of "is" is calculated using the following formula: ; In the formula, For trajectory points To candidate road sections Vertical projection distance, This represents the standard deviation of GPS observation noise (default value is 15 meters, corresponding to the horizontal positioning accuracy of a typical commercial GPS receiver). The physical meaning of this formula is: the closer a trajectory point is to a candidate road segment, the higher the probability that a vehicle is located on that road segment, following a normal distribution.

[0024] S412. Calculation of inter-segment transition probability (topological constraints): Two adjacent GPS track points and The transition between candidate road segments is constrained by the road network topology. If a vehicle transitions from a candidate road segment... Transfer to candidate road section This requires that there be a connecting path between the two road segments in the actual road network (either directly connected or connected via an intermediate road segment). Transition probability The calculation formula is: ; In the formula, Let be the shortest path length between the two candidate road segments in the road network. GPS track points and The spherical distance between them This is the scaling parameter (default value 50 meters). If two candidate road segments are inaccessible in the road network, the transition probability is set to 0, thus excluding the jump path.

[0025] S413, Viterbi algorithm for decoding optimal road segment sequences: Based on the observation probability and transition probability, the Viterbi Algorithm is used to solve for the hidden state sequence with the maximum probability, that is, to find a set of road segment sequences that maximizes the following expression. : ; By using dynamic programming recursive calculation to avoid the exponential complexity of exhaustive search, the optimal matching road segment ID for each GPS trajectory point is output. The road segment ID matched by the GPS trajectory points corresponding to the same alarm event is assigned as the "belonging road segment ID" for that alarm event.

[0026] S414, Matching confidence markers: For each matching result, calculate the confidence level: if the ratio of the highest observed probability to the second highest observed probability of a trajectory point is greater than 1.5, mark it as a "high confidence match"; otherwise, mark it as a "low confidence match". Calculate the percentage of "low confidence matches" among all matching results for a road segment. If it exceeds 20%, mark it in the scoring report for that road segment for user review.

[0027] S42. Driving timeline feature alignment (spatiotemporal coupling): For alarm events that have been matched to road segments, perform the following time alignment: S421. Calculation of relative driving time within the trip: For this trip, the effective GPS track point sequence sorted by time is defined with the departure time (the moment when the vehicle departs from the starting parking lot and its speed first continuously exceeds 5 km / h) as the time origin. Calculate each alarm event relative to relative time (Unit: hours), and normalize the total travel time of this trip to 100%, then the time progress position of the alarm event in the trip is determined. .

[0028] S422, estimated arrival driving time window for the section: For each road segment defined in S1, the estimated time for the vehicle to reach that segment is calculated based on the proportion of the segment's starting point within the total route mileage. This timeframe is jointly estimated using the total route mileage, the proportion of each road segment's mileage, and the historical average vehicle speed (from GPS statistics collected in S2), thus obtaining the "estimated driving time window" for that road segment during the journey. (unit:%).

[0029] S423, alarm-road segment-driving time triplet generation: The relative time progress of alarm events The estimated driving time window for the road segment where the alarm is located. The comparison is performed, and a classification label named "Fatigue Accumulation Stage Marker" is generated for this alarm event, with a value belonging to one of the following three categories: Early days: At this point, the driver is at the beginning of the journey and has plenty of energy; Mid-term: At this point, the driver is in a continuous driving phase, and fatigue gradually accumulates; Final stage: At this point, the driver is nearing the end of the journey and is at his most fatigued.

[0030] S43. Multi-dimensional aggregated statistics by line, section, and alarm type: After spatial matching by S41, all valid alarm events have obtained their "Road Segment ID". Using the three fields of Route ID, Road Segment ID, and Alarm Type as grouping keys, the cleaned and deduplicated valid alarm events are grouped and aggregated. The total number of valid alarm events within each group (i.e., the deduplicated count) is calculated, and the sum of the effective mileage of all vehicles on that road segment within the statistical period is also summarized (road segment statistical mileage). The data table output by the aggregation operation includes Route ID, Road Segment ID, Alarm Type, Total Number of Valid Alarms, Road Segment Statistical Mileage, Fatigue Accumulation Stage Marker, and Data Validity Label. Among these, Route ID, Road Segment ID, and Alarm Type are the grouping keys; the total number of valid alarms is the aggregated statistical value; the road segment statistical mileage is the aggregated summary value; and the remaining fields are additional attributes.

[0031] S44. Normalized calculation of sub-indicators per 100 kilometers: For each record in the S43 aggregate table, calculate the normalized index value per 100 kilometers for that alarm type. To eliminate the impact of differences in statistical mileage across different road sections on the evaluation: ; In the formula, This represents the total number of valid alarms of a certain type on a certain road section (derived from the count after deduplication in S3). This is the statistical mileage (unit: km) of the road segment, which is the sum of the cumulative mileage traveled by all vehicles on the road segment within the statistical period. The unit is "times per 100 kilometers".

[0032] For static infrastructure sub-indicators, there is no need to normalize by mileage. The raw values ​​collected by S2 (such as the maximum longitudinal slope of 4.2%, in %) or inherent density values ​​(such as the curve density of 3.5 curves / 100 km, whose denominator is the road segment mileage, which is itself in density form) can be directly used as input to the S5 linear decay scoring model. value.

[0033] If a road segment is marked as "insufficient data" (statistical mileage < 10km), then all dynamic alarm normalized values ​​for that road segment will be set to NULL and will not participate in the S5 scoring.

[0034] S5. Risk Quantification Calculation: S51, Sub-index Linear Decay Score: For each valid sub-indicator output by S4, a linear decay scoring model is used to convert its normalized value into a score within the range of 0 to 100. A higher score indicates a lower risk for that sub-indicator (i.e., a safer indicator dimension). The scoring formula is: ; In the formula: This refers to the normalized or original value of the sub-indicator. For dynamic alarm sub-indicators, Normalized value per 100 kilometers calculated for S44 (Unit: times / 100 km); For static attribute sub-indicators, The raw or density values ​​collected for S2; This is the sensitivity coefficient of the sub-index, dimensionless, and its physical meaning is: per unit The deduction amount resulting from the increase. For example... This means that for every additional alarm frequency per 100 kilometers, 25 points will be deducted from the score; Ensure the score is non-negative, and truncate the lower limit to 0 points.

[0035] Furthermore, The calibration method for the values ​​is based on historical accident causation data and uses the Pearson correlation coefficient. Specifically, this involves collecting historical data from a large number of road sections and calculating the correlation coefficient between the normalized values ​​of each sub-indicator and the historical accident rate of that road section. The stronger the correlation ( The closer the value is to 1, the higher the sensitivity of the indicator to security risks, and the greater the value assigned to it. This value gives it a higher marginal impact in the scoring. For example, the correlation coefficient between the number of historical accidents and the accident rate of a road segment. It has the highest sensitivity coefficient among all sub-indicators, therefore it is assigned the highest sensitivity coefficient. Correlation coefficient between the number of risk interventions and the accident rate The correlation is weak, therefore a low sensitivity coefficient is assigned. .default The values ​​are calibrated and fixed in the system based on publicly available datasets from the entire industry. Users can then adjust them within a range of ±50% through the configuration interface to adapt to different fleet management characteristics.

[0036] The preferred five dimensions and 22 sub-indicators and their default values. The value configuration is as follows: R1 section characteristic risks (w1 weight 15%) Curve density (number per 100 km) 8 S2 Static Road Geometry Maximum longitudinal slope (%) 12 S2 Static Road Geometry Percentage of tunnels and bridges (%) 1.0 S2 Static Road Geometry Interchange density (number per 100 km) 6 S2 Static Road Geometry Frequency of speed limit changes (times per 100 km / h) 10 S2 Static Road Geometry R2 traffic flow risk (w2 weight 25%) Daily traffic volume deviation rate (%) 0.8 S2 traffic flow data Large vehicle mix rate (%) 1.2 S2 traffic flow data Traffic congestion frequency (times / week) 15 S2 traffic flow data Accident hotspot density (number per 100 km) 20 S2 Historical Accident Data R3 Environmental and Climate Risks (W3 weighting 15%) Average number of severe weather days per year (days) 5 S2 meteorological and environmental data Visibility failure rate (%) 1.5 S2 meteorological and environmental data Seasonal risk volatility index 10 S2 meteorological and environmental data R4 operational load risk (w4 weight 15%) Total route length (100 kilometers) 3 S1 Line Archives Estimated driving time (hours) 8 S1 Line Archives Percentage of nighttime driving (%) 1.5 S2 historical operation data Rest stop coverage rate (inverse) 2.0 S2 Static POI Data Fatigue Cumulative Index 18 S2 historical operation data R5 historical operational risks (w5 weight 30%) DMS alarm frequency (times / 100 km) 25 S4 Normalized Dynamic Alarm ADAS alarm frequency (times / 100 km) 22 S4 Normalized Dynamic Alarm Historical accident count (times per 100 kilometers) 30 S2 Historical Accident Data Number of risk interventions (per 100 km) 7 S4 Normalized Dynamic Alarm Number of high-risk events (per 100 km) 9 S4 Normalized Dynamic Alarm ; If a sub-indicator is marked as "insufficient data" or "invalid data" in S4, then that sub-indicator will not participate in the scoring and will be removed during dimension aggregation.

[0037] S52, Dimensional Score Aggregation and Handling Insufficient Data: Scores of all valid sub-indicators within the same dimension Take the arithmetic mean to obtain the dimensional score for that dimension. : ; Constraint on the number of valid sub-indicators within a dimension: If the number of valid sub-indicators within a dimension is less than 2, the dimension is marked as "insufficient data" and will not participate in the S53 LRC weighted calculation. In this case, the original weights of the dimension are proportionally redistributed to other valid dimensions, specifically as follows: Let the original weights be... ( (Total 100%), the effective dimension set is (i.e., dimensions not marked as "insufficient data"), then the dimension Correction weights for: ; S53. Weighted Calculation of Line Risk Coefficient: The original value of the line risk coefficient is calculated using a weighted summation method. (Raw Line RiskCoefficient) ; In the formula to Scoring is given for each of the five dimensions. to The weights for the corresponding dimensions (default weights or weights modified by S52).

[0038] The value ranges from 0 to 100, with higher values ​​indicating a safer route (lower risk). If all five dimensions are marked as "insufficient data," the system cannot calculate. The system will output "Insufficient data, unable to evaluate" and terminate the process.

[0039] S6. Veto Security Verification: S61. Definition of rejection trigger condition: The system performs the following four checks on each segment of the route: (1) Fatal accident rejection: Check whether there are any fatal traffic accidents (i.e., traffic accidents that cause death) on this road section in the past 12 months as determined by the public security traffic management department. This condition is based on historical facts, and once triggered, it indicates that there is an indisputable risk to life safety on this road section. No additional threshold is set.

[0040] (2) No signal coverage is a rejection criterion. Check whether there is a continuous area with no mobile communication signal coverage exceeding 20km in length on the road section. The 20km threshold is set based on the following: According to the technical requirements for the timeliness of emergency alarm information upload in the transportation industry standard JT / T 794-2019 "Communication Protocol and Data Format of Satellite Positioning System Terminal for Road Transport Vehicles", vehicles must complete the alarm signal upload within 30 seconds after an emergency occurs; combined with the average travel speed of commercial vehicles in this type of road section (mountainous or remote areas) of 60km / h, the travel distance within 30 seconds is approximately 0.5km; considering the typical distance for base station signal recovery after an interruption and the triggering cycle of the system reconnection mechanism, the threshold of 20km is taken as the minimum length for judging "long-term, large-area signal blind spots" based on comprehensive engineering experience. No signal areas exceeding this length will prevent vehicles from sending alarm signals to the fleet monitoring platform or rescue agency through the mobile communication network in emergency situations such as accidents, malfunctions, or sudden health problems of the driver, constituting an unacceptable safety risk.

[0041] (3) High-risk events are rejected. The historical frequency of high-risk events for the road segment is calculated as follows: Number of high-risk events (times / 100 km) = Total number of high-risk events within the statistical period of the road segment ÷ Statistical mileage of the road segment × 100. If this value exceeds 5 times / 100 km, a rejection is triggered. Among them, "high-risk event" is defined as an event in S5 where the corresponding sub-index score is 0, that is, the original value of the sub-index has exceeded the effective scoring range of the linear decay scoring model ( × A score of ≥100 indicates that the risk level of the road segment on this indicator has become so severe that it has been truncated by the scoring model.

[0042] (4) Sharp bends and long slopes are rejected. The road section must be checked to see if there are sections with a continuous downhill length exceeding 10km and an average longitudinal slope exceeding 4%. The combined setting of the 10km and 4% thresholds is based on the physical mechanism of thermal fade in commercial vehicle braking systems: During continuous downhill driving, the brakes convert the vehicle's potential energy into heat energy through friction. The temperature of the brake drum or brake disc increases continuously with the downhill mileage, the coefficient of friction decreases accordingly, and the braking torque decreases. According to vehicle braking safety engineering experimental data, under continuous downhill conditions with an average longitudinal slope of 4%, using the conventional operation method of combining engine braking and service braking, the vehicle brake drum temperature approaches the failure critical temperature (approximately 400℃) after about 8-10km of continuous downhill driving. Afterward, the braking torque decreases significantly, the braking distance increases dramatically, posing a serious safety hazard. Therefore, setting the combined condition of a 10km continuous downhill slope and a 4% average longitudinal slope as the rejection threshold has an objective basis from a braking safety engineering perspective.

[0043] S62. Definition of rejection trigger condition: For all road segments defined in S1, sequentially perform the four checks in S61, maintaining a rejection trigger list that records all triggered rejection types, their corresponding road segment IDs, and specific check values. After all road segments have been checked, execute the following correction logic: ; This invention will A score below 45 is defined as an extremely high-risk route. A veto will be forcibly triggered. The score is set at 44 points, one point lower than the extremely high-risk threshold, ensuring that any line triggering a veto will inevitably fall into the "extremely high-risk" category. Simultaneously, one point of physical semantic space is reserved on the scoring scale to avoid overestimating the risk. The problem of complete information loss due to truncating directly to 0 points (44 points still retains some quantitative difference information). If the score is already below 44 (i.e., extremely low risk score), then maintain... The assessment results will remain unchanged, and no upward adjustment will be implemented, so as not to weaken the already extremely low assessment results.

[0044] S7, Level Determination and Strategy Output: S71. Line Risk Level Determination: According to the final The risk level of the line is determined based on the following threshold values, each with a clear physical and managerial meaning: Low risk LRC≥85 green The risks across all aspects of the route are at a low level, and standard safety management procedures can be followed. Medium risk 60≤LRC<85 yellow The route has localized or individual risk factors, requiring appropriate attention in scheduling and monitoring. High risk 45≤LRC<60 orange color The line presents significant risks across multiple dimensions, necessitating enhanced management and the allocation of additional security resources. Extremely high risk LRC<45 red The line has serious safety hazards and requires the activation of the highest level of safety control measures. ; The LRC score has an upper limit of 100 points and a lower limit of 0 points; there are no legal values ​​outside this range. 85 points serves as the dividing line between low and medium risk, corresponding to the ideal safety state where "all five dimensions score no less than 85 points"; 60 points corresponds to the state where "the average score of the five dimensions is just above the passing mark," and values ​​below this indicate significant deficiencies in most dimensions; 45 points is the dividing line between high and extremely high risk, corresponding to the mandatory correction target value (44 points) for S6 veto, meaning that lines triggering veto are forced to fall into the extremely high risk range.

[0045] S72. Generation of risk profiles for each road segment: For each road segment defined in S1, output the following information: (1) Basic information of the road segment: road segment ID, start and end locations, mileage, road grade; (2) Rejection flag: Whether the road segment has triggered any of the rejection conditions in S6. If it has, list the specific rejection type and the judgment basis value. (3) Scores in each dimension: This road segment in to The detailed scores of sub-indicators under each dimension (taking the score values ​​of each sub-indicator to which the road segment belongs) make it easier to identify the main risk dimensions of the road segment; (4) Top 3 High-Frequency Alarms: The three alarm types that occurred most frequently during the statistical period of this road section and their normalized values ​​per 100 kilometers, which are derived from the S4 aggregation results; (5) Static risk feature labels: Automatically labeled based on the static attributes collected by S2, including but not limited to “accident black spots”, “sharp bends”, “long downhills”, “tunnel groups”, “no signal coverage”, “high traffic flow”, etc. (6) Distribution of fatigue accumulation stage: The proportion of alarm events in the three stages of “initial / middle / final” of this road section is derived from the fatigue accumulation stage markers attached to each alarm event by S42.

[0046] The aforementioned profile data is output in a structured format (JSON / table) for front-end visualization, providing managers with a detailed risk view at the road segment level.

[0047] S73. Generation of safety management strategies and issuance of equipment control commands: Based on the route risk level and scores across various dimensions, differentiated fleet safety management strategies are generated. These differentiated fleet safety management strategies include driver qualification and scheduling management, monitoring frequency control instructions, and rest stop layout rules.

[0048] (1) Driver qualification and scheduling management: Driver eligibility criteria and scheduling rules are set according to risk levels, as follows: Low risk ≥1 year No requirements No requirement Medium risk ≥2 years Safety score ≥ 60 points in the past 6 months No requirement High risk ≥3 years Safety score ≥ 75 points in the past 6 months Road sections that trigger a veto require two drivers Extremely high risk ≥5 years Safety score ≥ 85 points in the past 6 months Mandatory dual-driver rotation throughout the entire process ; The above rules are stored in the system as parameterized conditional expressions. The system automatically matches the corresponding conditions according to the risk level of the route and filters the list of drivers who meet the conditions in the fleet scheduling module for the dispatcher to confirm.

[0049] (2) Monitoring frequency control command: Real-time monitoring strategies are set at different levels based on risk levels, and parameters are sent to the monitoring and scheduling platform in the form of configuration commands: Preferably, low-risk areas are inspected every 4 hours; medium-risk areas are inspected every 2 hours; high-risk areas are inspected every 1 hour, with random video spot checks enabled; and extremely high-risk areas are monitored in real time throughout the process.

[0050] (3) Monitoring frequency control command: Based on operational load risk dimension score The mandatory spacing of rest stops is dynamically adjusted. A lower score indicates a higher operational load risk (reflected by five sub-indicators: total mileage, estimated driving time, percentage of nighttime driving, rest stop coverage, and fatigue accumulation index), requiring more frequent rest stop arrangements. Preferred, In this case, a mandatory rest stop will be set every 2 hours or every 200 kilometers; Rest stops will be set up every 3 hours or every 300 kilometers (based on who arrives first). Follow the minimum standards stipulated by law (take a 20-minute break after 4 hours of continuous driving).

[0051] All the above strategies are output in a structured data format, including the following fields: Line ID, Risk Level, Strategy Type, Strategy Parameters, Target, and Execution Status.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) A multi-level data processing link from raw sensor data to structured road segment risk indicators was constructed, solving the technical problem that vehicle sensor data cannot be directly used at the route level. In the prior art, DMS / ADAS alarm data is stored at the vehicle or driver level, which is isolated from road infrastructure data and cannot be directly used for route-level risk assessment. This invention discretizes continuous geographic space into the smallest computable assessment unit by segmenting road segments, establishes a spatial index relationship between vehicle sensor data and road segments through multi-source data acquisition, eliminates the interference of high-frequency noise from sensors on statistical accuracy by deduplication through differentiated time windows, anchors discrete alarm events to specific road segments through GPS spatial matching and time axis alignment, and finally eliminates the evaluation bias caused by the difference in statistical mileage of different road segments through normalization processing per 100 kilometers. The above data processing link gradually transforms unstructured sensor time series data into structured road segment risk indicators, enabling the raw sensor data to be effectively utilized at the route level.

[0053] (2) The HMM map matching algorithm solves the problem of road segment ambiguity caused by GPS positioning drift, and improves the accuracy of spatial fusion of dynamic alarm data and static road data. In scenarios such as urban elevated and ground parallel road sections, tunnel entrances and exits, and mountain curves, commercial GPS has a random positioning error of 10 to 30 meters. The traditional shortest Euclidean distance matching method is prone to misidentifying vehicles on elevated roads as ground roads. This invention models road segment matching as a sequence decoding problem with road network topology constraints, and uses the spatial connectivity constraints between adjacent GPS trajectory points to correct the observation probability, so that the matching results of isolated single points with positioning errors conform to the globally optimal path of the entire trajectory, effectively eliminating mismatches caused by GPS positioning drift.

[0054] (3) A differentiated time window deduplication mechanism was established to solve the problem of risk statistics distortion caused by repeated alarms triggered by the same dangerous state in a short period of time. A driver's fatigue and closing of eyes once may trigger multiple alarms from the DMS within a few seconds, and a sudden braking may trigger multiple sudden deceleration alarms from the ADAS. If the original alarms are counted directly, the same physical event will be repeatedly counted in the statistics, resulting in a systematically high risk value for the road segment. This invention sets differentiated deduplication windows for different sensor types, such as the physiological persistence of fatigue signs in the DMS and the mechanical response cycle of ADAS braking, to merge repeated triggers of the same physical event into a single valid event, so that the normalized alarm frequency accurately reflects the true risk density, rather than being a byproduct of the sensor sampling frequency.

[0055] (4) A driving timeline feature alignment mechanism was established, coupling alarm events with the driving fatigue accumulation process in the time dimension, solving the technical defect in the existing technology that only spatially superimposes "dynamic alarms" and "static road segments" without temporal correlation. When assigning DMS alarms to road segments, the existing scheme only uses GPS coordinates for spatial matching, ignoring the fact that the alarm risk of the same road segment is closely related to its time progress in the journey—curves with the same radius of curvature pose completely different composite risks to drivers when they appear at the beginning and end of the journey. This invention calculates the relative time position of alarm events within the journey, mapping the mileage progress of each road segment in the total journey to the expected driving time window, realizing the dual correlation between the road segment to which the alarm event belongs and its time position in the driving fatigue accumulation process, providing a data foundation for subsequent analysis of the composite risk of "fatigue accumulation + specific road segment".

[0056] (5) A veto mechanism was established as a security verification layer for the weighted scoring model, solving the technical problem that multi-dimensional weighted scoring might dilute the key physical risks of a single road segment. When existing technologies use the analytic hierarchy process or multi-objective optimization for path selection, the scores of each dimension are weighted and summed to obtain a comprehensive score. If the overall historical operating data of a certain route is good, but it contains a sharp bend, long slope, or communication blind spot, the physical risk of that road segment may be diluted by the excellent scores of other road segments in the weighted average, resulting in the comprehensive score of the entire route failing to reflect the safety hazards of key road segments. This invention sets up a veto verification layer independently outside the weighted scoring model. When any road segment has hard constraint risks at the physical level such as sharp bends, long slopes, or lack of signal coverage, the risk coefficient of the route is forcibly corrected to the extremely high risk range, so that the mathematical calculation results of the weighted scoring model give way to the objective judgment of physical hard constraints.

[0057] (6) By normalizing the data per 100 kilometers, the technical problem of incomparable risk indicators between road segments of different lengths is solved. The risk density of the same number of alarms is drastically different on short and long road segments. This invention uses "per 100 kilometers" as a unified unit for normalization, making the alarm frequencies between different road segments horizontally comparable, and at the same time providing input parameters with consistent units of measurement for the weighted summation of subsequent route risk coefficients. Attached Figure Description

[0058] Figure 1 This is a flowchart of the commercial vehicle route risk assessment method based on spatiotemporal coupling of multi-source sensor data according to the present invention. Detailed Implementation

[0059] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0060] This embodiment uses a transportation route from Nanjing to Hangzhou (hereinafter referred to as the "Nanjing-Hangzhou Line") operated by a logistics company as the evaluation object to describe in detail the overall operation process of the technical solution of this invention. The total mileage of this route is approximately 280km, mainly passing through the G25 Changshen Expressway (Nanjing-Hangzhou section). The fleet consists of 42 commercial vehicles operating on this route in the past 30 days, with a total mileage of approximately 11,760km. The system's default statistical period is the 30 calendar days prior to the evaluation date.

[0061] S1. Definition of Evaluation Unit: S11. Route planning: The system administrator input the starting point "a logistics park in Jiangning District, Nanjing City, Jiangsu Province" (longitude 118.78°E, latitude 32.06°N) and the destination "a distribution center in Yuhang District, Hangzhou City, Zhejiang Province" (longitude 120.12°E, latitude 30.27°N) through the interactive interface. The system called the Gaode Map navigation API to plan a recommended driving route, returning a total route distance of 278.6km. The route trajectory is south along the G25 Changshen Expressway, passing through Nanjing—Lishui—Yixing—Huzhou—Hangzhou. The administrator confirmed that this route is the target route for this evaluation.

[0062] S12, Road segmentation: The system segments the road along the above path according to the following rules: (1) Road grade change point: The junction of G25 Expressway and urban expressway in Nanjing section (near the starting point) is used as the dividing point; (2) Administrative division boundary: The boundary between Jiangsu and Zhejiang provinces (the border between Yixing and Huzhou, at a distance of about 145km) is used as the dividing point; (3) Start and end points of large structures: The entrance (at about 120km) and the exit (at about 128km) of a tunnel group (5 consecutive tunnels with a total length of about 8km) in Yixing are each used as the dividing point; (4) Service area locations: Lishui Service Area (approximately 55km away) and Taihu Service Area (approximately 180km away) along the route serve as dividing points.

[0063] Based on the above-mentioned dividing points, the system divides the route into several continuous segments, which are 6 segments in this embodiment. Each segment is no longer than 20km, and the segments are connected end to end without overlapping.

[0064] NJ-HZ-01 Nanjing Logistics Park Lishui Service Area 55.2km high speed NJ-HZ-02 Lishui Service Area Yixing Tunnel Entrance 64.8km high speed NJ-HZ-03 Yixing Tunnel Entrance Yixing Tunnel Exit 8.0km Highway (tunnel group) NJ-HZ-04 Yixing Tunnel Exit Jiangsu-Zhejiang provincial border 17.0km high speed NJ-HZ-05 Jiangsu-Zhejiang provincial border Taihu Service Area 35.0km high speed NJ-HZ-06 Taihu Service Area Hangzhou Distribution Center 98.6km Expressway / Urban Expressway ; S13. Road segment attribute recording and route file creation: The system generates a unique ID for each road segment (format: "Route Abbreviation-Serial Number"), records the starting and ending latitude and longitude coordinates, mileage, road grade, and design speed of each road segment, and establishes a "Nanjing-Hangzhou Line" route file.

[0065] Output the above road segment list and route file to S2.

[0066] S2, Multi-source data acquisition: S21. Static infrastructure data collection: The system uses the road segment ID as an index to call the following data sources in parallel: (1) Road geometry features (Gaode Road Geometry Query API): Get the number of curves, maximum longitudinal slope, total length of tunnels and bridges, number of interchanges, and number of speed limit change points for each road segment; (2) Traffic flow characteristics (Gaode Traffic Platform): Obtain the average daily traffic volume of each road segment; (3) Meteorological environment (Meteorological Service API): Obtain meteorological data for the past 3 years for each district and county along the route; (4) Historical accident data (traffic management data interface): Obtain accident records for each road section over the past 12 months; (5) Communication signal coverage data (operator network coverage interface): obtain the signal coverage status along the line.

[0067] S22. Dynamic vehicle-mounted sensor data acquisition: The system uses route files as the search scope and extracts raw sensor data from the fleet operation platform's data warehouse for 42 vehicles that operated on that route over the past 30 days, including: (1) DMS alarm data (DMS system backend): Extract alarm records such as fatigue driving, closing eyes, yawning, and distraction; (2) ADAS alarm data (ADAS system backend): Extract alarm records such as sudden deceleration, lane departure, and forward collision warning; (3) Overspeed alarm data (T-Box): Extract overspeed event records; (4) GPS trajectory data (T-Box): Extract the complete trajectory point sequence of each vehicle for each run, with a timestamp interval of 5 seconds. Each trajectory point includes latitude and longitude, instantaneous vehicle speed, and cumulative mileage.

[0068] S23. Statistical Period and Data Validation: The system confirmed that within a 30-day statistical period, the number of vehicles operating on the route was 42 (≥3 vehicles), and the total number of trips was approximately 168 (≥5 trips), meeting the minimum sample size requirement. The statistical mileage for each route was greater than 10km (for route NJ-HZ-03, due to its shorter mileage, the statistical mileage is approximately 42 vehicles × 8km / trip × approximately 6 trips ≈ 2016km, which is greater than 10km), and there was no "insufficient data" marker.

[0069] The aforementioned static attribute dataset and dynamic sensing data stream are output to S3.

[0070] S3. Data Cleaning and Deduplication: S31. Data quality cleaning: The system performs the following checks in sequence: (1) Dynamic record legality verification: Illegal records collected within 30 days will be removed, such as records with GPS coordinates outside the reasonable range, records with timestamps outside the statistical period, and records with vehicle speeds >150km / h. (2) Verification of the reliability of single trip mileage: Among the 42 vehicles, if the deviation between the single trip mileage and the total route mileage is greater than 20%, all alarm data and GPS trajectory data of that trip will be marked as "abnormal" and removed, and will not be included in subsequent aggregation. (3) Road segment mileage verification: The mileage of each road segment is ≥10km and there is no "insufficient data" mark. S32. Deduplication via Differentiated Time Window: The system performs deduplication on the cleaned alarm records according to alarm type. Taking the DMS fatigue driving alarm for road segment NJ-HZ-03 (tunnel section, 8km) as an example: A total of 320 original DMS fatigue driving alarm records were collected for this road section during the statistical period. Deduplication was performed using a 5-second time window. For the same vehicle (taking vehicle V-018 as an example), 18 fatigue driving alarms were triggered consecutively within 2 minutes of traversing the tunnel group section (9:32:05 to 9:34:00). The first record at 9:32:05 was marked as a valid event; 9:32:08, 3 seconds (≤5 seconds) from the first record, was merged; 9:32:14, 9 seconds (>5 seconds) from the first record, was marked as a new valid event; and so on.

[0071] After deduplication of all 320 records for road segment NJ-HZ-03, 186 valid events were output, with a merging rate of 41.9%.

[0072] For example, ADAS emergency deceleration alarms are processed in 10-second windows: if the same vehicle triggers an emergency deceleration alarm consecutively with a 6-second interval, it is counted as one alarm; if the interval is 15 seconds, it is counted as two independent valid events.

[0073] After deduplication, each valid event is output with a "Deduplicated" label and the original merge count.

[0074] The output data from S3 is passed to S4.

[0075] S4. Data Spatiotemporal Fusion and Normalization: S41. Spatial attribution determination based on HMM map matching: Taking vehicle V-018's journey from 9:30 AM to 10:30 AM on June 15, 2026 as an example, the GPS trajectory point sequence for this journey contains approximately 720 points. The system performs HMM matching on each trajectory point: (1) Candidate road segment selection: For the initial trajectory point (119.58°E, 31.22°N), the candidate road segments within a 50-meter search radius are NJ-HZ-03 and NJ-HZ-02 (this point is located about 30 meters before the entrance of Yixing Tunnel, and there is GPS drift). Observation probability: 12m vertical distance to NJ-HZ-03 (high probability), 38m vertical distance to NJ-HZ-02 (low probability).

[0076] (2) Calculation of transition probability: The subsequent trajectory point sequence continues to move south. According to the Viterbi algorithm, although the first point is closer to NJ-HZ-02, the road network topology constraint of the subsequent trajectory points (vehicles must pass through the NJ-HZ-03 tunnel to reach the south side from the north) makes the algorithm re-determine all the positioning points of this trip as the NJ-HZ-03 road segment.

[0077] (3) Matching results: Among all the alarm events of vehicle V-018 on this trip, 15 DMS fatigue alarms and 4 ADAS lane departure alarms were matched to NJ-HZ-03, and the matching confidence was "high confidence" (the ratio of the highest probability to the second highest probability is 2.1>1.5).

[0078] For elevated and ground-level road sections (such as the overlapping area of ​​urban expressways and ground-level roads near Nanjing South Railway Station), the HMM algorithm successfully avoids mismatching vehicle trajectories on elevated roads to ground-level road sections through topological constraints.

[0079] S42, Driving Timeline Feature Alignment: Taking vehicle V-018 as an example: departure time 9:00:00 (speed first > 5km / h), arrival at the tunnel entrance 9:32:05 (approximately 32 minutes travel time), total journey time approximately 4.5 hours (9:00 to 13:30). The relative time progress at the tunnel entrance is 32 minutes ÷ 270 minutes = 11.9%, marked as the "initial" stage. Most DMS fatigue alarms triggered by the vehicle within the tunnel group occur in the "initial" stage of the journey, indicating that the alarms are less correlated with fatigue accumulation (the driver's journey has just begun) and more correlated with physiologically low arousal caused by the monotonous visual environment within the tunnels.

[0080] S43. Multi-dimensional aggregated statistics by line, section, and alarm type: The system groups alarms by line ID, segment ID, and alarm type, and summarizes the total number of valid events for each alarm type in each segment: L-2026-07 NJ-HZ-01 DMS fatigue driving 42 2318km Initial 15% / Mid-term 55% / Final 30% L-2026-07 NJ-HZ-01 ADAS rapid deceleration 18 2318km Initial 10% / Mid-term 50% / Final 40% L-2026-07 NJ-HZ-02 DMS fatigue driving 98 2722km 8% in the early stage / 52% in the mid-term / 40% in the final stage L-2026-07 NJ-HZ-03 DMS fatigue driving 186 2016km Initial 85% / Mid-term 10% / Final 5% L-2026-07 NJ-HZ-03 ADAS rapid deceleration 52 2016km Initial 78% / Mid-term 15% / Final 7% L-2026-07 NJ-HZ-04 DMS fatigue driving 152 2380km Initial 5% / Mid-term 35% / Final 60% ... ... ... ... ... ... ; S44, Normalized calculation per 100 kilometers: For each record in the above aggregated table, calculate the normalized value per 100 kilometers. Taking the DMS fatigue driving in the NJ-HZ-03 segment as an example, the normalized value per 100 kilometers for DMS fatigue driving in the NJ-HZ-03 segment is... The calculation is as follows:

[0081] The normalized values ​​for each road segment are summarized below. This example only lists the DMS fatigue driving alarm; other alarm types are similar. NJ-HZ-01 42 2318km 1.81 NJ-HZ-02 98 2722km 3.60 NJ-HZ-03 186 2016km 9.23 NJ-HZ-04 152 2380km 6.39 NJ-HZ-05 75 2450km 3.06 NJ-HZ-06 44 2890km 1.52 ; Although the total number of valid alarms for road segment NJ-HZ-03 (186) is lower than that for road segment NJ-HZ-04 (152), the normalized alarm rate of 9.23 alarms / 100km for NJ-HZ-03 is significantly higher than that of NJ-HZ-04 (6.39 alarms / 100km). This is because the NJ-HZ-03 road segment is shorter (8km), resulting in a smaller cumulative mileage for all vehicles (2016km) and an extremely high alarm density (alarm frequency per unit mileage). Without normalization per 100km, although the absolute number of alarms for the short road segment may seem low, the risk density per unit mileage could be severely underestimated. Without deduplication, the normalized value for road segment NJ-HZ-03 would reach 15.87 alarms / 100km (320÷2016×100). Deduplication reduces this value by approximately 42%, preventing an overestimation of the risk for this road segment.

[0082] The normalized index value is passed to S5.

[0083] S5. Risk Quantification Calculation: S51, Sub-index Linear Decay Score: Taking road segment NJ-HZ-03 as an example, some key sub-indicators are scored, and the calculation formula is as follows: ; Curve density 2.8 units / 100 kilometers 8 78 Maximum longitudinal slope 3.2% 12 62 tunnels and bridges as a percentage 100% 1.0 0 DMS alarm frequency 9.23 times / 100 kilometers 25 0 ADAS Lane Departure 2.58 times / 100 kilometers 22 43 ; Among the scores of the above sub-indicators, the tunnel-bridge ratio score is 0 (because NJ-HZ-03 is a tunnel group section, the tunnel-bridge ratio is 100%), and the DMS alarm frequency score is 0 (9.23×25=230.75>100, exceeding the linear attenuation range), indicating that this road section is at the highest risk level in these two indicators.

[0084] S52, Dimensional Score Aggregation: Calculate the dimension score for each of the five dimensions (taking the average score of all valid sub-indicators within that dimension): R1 section characteristics avg (curve density 78, maximum longitudinal slope 62, tunnel and bridge ratio 0, interchange density 85, speed limit change frequency 88) 62.6 R2 traffic flow avg (average daily traffic volume deviation 75, large vehicle mixing rate 62, congestion frequency 78, accident black spot density 95) 77.5 R3 Environment and Climate avg (average number of severe weather days per year: 52; visibility failure rate: 88%; seasonal risk fluctuation: 70%) 70.0 R4 Operating Load avg (Total route mileage 91, estimated driving time 64, night driving percentage 73, rest stop coverage 45, fatigue accumulation index 58) 66.2 R5 historical operation avg(DMS alarm frequency 0, ADAS lane departure 43, historical accident count 82, risk intervention count 88, high-risk event count 73) 57.2 ; S53. Calculation of Line Risk Coefficient: ; Set RawLRC=66 Output to S6.

[0085] S6. Veto Security Verification: The system performs four rejection checks on each of the six road segments defined in S1: NJ-HZ-01 none none no no no NJ-HZ-02 none none no no no NJ-HZ-03 none none no Yes (11.2km of continuous downhill slope, with an average longitudinal slope of 4.1%) yes NJ-HZ-04 none none no no no NJ-HZ-05 none Yes (no signal for 22km) no no yes NJ-HZ-06 none none no no no ; Details of the veto decision: According to the geographical feature data collected by S2, the NJ-HZ-03 section (Yixing Tunnel Group Section) is not only a continuous tunnel group (8km), but also has a continuous downhill section of about 11.2km before entering the tunnel group (from the end of NJ-HZ-02 section to NJ-HZ-03 section), with an average longitudinal slope of 4.1%. This exceeds the threshold for rejection of sharp bends and long slopes with "continuous downhill > 10km and average longitudinal slope > 4%", thus triggering the rejection of sharp bends and long slopes.

[0086] According to the communication signal coverage data collected by S2, there is a continuous area of ​​about 22km without mobile communication signal coverage in the section of NJ-HZ-05 from about 185km to 207km, which exceeds the 20km threshold and triggers the rejection due to no signal coverage.

[0087] Execution of the veto modification logic: The RawLRC value of 66 > 45 was corrected to a final LRC of 44; the system outputs a final LRC of 44. If relying solely on the S5 weighted scoring model, the route's RawLRC score of 66 would classify it as "medium risk." However, the sharp bends and long slopes of section NJ-HZ-03 and the lack of signal coverage on section NJ-HZ-05 both constitute unacceptable physical risks. A veto was applied, forcibly revising the route from "medium risk" to "extremely high risk" (LRC=44). This more accurately reflects the serious safety hazards present in key sections of the route and prevents the critical risks of a single section from being diluted by the excellent performance of other sections.

[0088] S7, Level Determination and Strategy Output: S71. Line Risk Level Determination: Based on the final LRC=44, the route risk level is determined to be an extremely high-risk route according to the threshold.

[0089] S72, Risk Profile for Each Road Section: Taking the NJ-HZ-03 road segment, which triggered a veto, as an example, the system outputs a profile of this segment. 93.6% of the DMS fatigue driving alarms for segment NJ-HZ-03 are concentrated in the "initial" stage of the journey (less than 30 minutes after the driver has started driving), contrary to the usual pattern of fatigue accumulation. Combined with the static characteristic of this segment being a group of tunnels, this suggests that the fundamental reason for the high alarm frequency is the monotonous visual environment inside the tunnels leading to physiologically low driver alertness, rather than fatigue accumulation from prolonged driving. This finding provides a basis for the precise formulation of subsequent safety management strategies.

[0090] S73. Generation of safety management strategies and issuance of equipment control commands: The system generates and distributes fleet safety management strategies based on the extremely high risk level and scores across various dimensions.

[0091] Specifically, for high-risk routes, drivers must have more than 5 years of driving experience and a safety score of more than 85 points in the past 6 months, with mandatory dual-driver rotation throughout the entire route; the entire route will be monitored in real time (with continuous video signal transmission); and rest stops will be set up every 3 hours or every 300 kilometers. Since the total route length is 278.6km < 300km, only one mandatory rest stop is required at the Taihu Service Area, with a rest duration of no less than 20 minutes.

[0092] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. A method for assessing the risk of commercial vehicle routes based on the spatiotemporal coupling of multi-source sensor data, characterized in that: Includes the following steps: S1. Definition of the evaluation unit; The transportation route to be evaluated is divided into multiple continuous segments according to the characteristics of road infrastructure. A unique identifier is established for each segment and the static attributes of the segment are recorded. A route file containing the segment identifier sequence is established. S2, Multi-source data acquisition; Using the road segment identifier as an index, static attribute data of each road segment is collected; at the same time, dynamic sensing data of all vehicles running on the route within the statistical period is collected, and the dynamic sensing data includes at least alarm events and corresponding GPS trajectory point sequences. S3, Data Cleaning and Deduplication; Data cleaning is performed on the alarm events to remove unqualified records; after cleaning, the alarm events are deduplicated and merged according to the alarm type by setting a differentiated time window, and repeated triggers of the same physical state are merged into a single valid event; S4. Data spatiotemporal fusion and normalization; The deduplicated alarm events are associated with specific road segments using a map matching algorithm, and the normalized index value per 100 kilometers for each alarm type on each road segment is calculated. At the same time, the time progress position of the alarm event within its respective trip is aligned with the expected driving time window of the corresponding road segment to generate a stage marker for the alarm event in the process of driving fatigue accumulation. S5. Risk Quantification Calculation; Each road segment is independently checked for safety. When any road segment has a preset physical safety risk triggering condition, the original value of the line risk coefficient is corrected to obtain the final line risk coefficient. S6, One-vote veto security verification; Each road segment undergoes independent safety verification. When any road segment has a preset physical safety risk triggering condition, the original value of the line risk coefficient is corrected to obtain the final line risk coefficient. S7, Level Determination and Strategy Output; The risk level of the line is determined based on the final line risk coefficient, and a safety management strategy is generated and output.

2. The commercial vehicle route risk assessment method based on spatiotemporal coupling of multi-source sensor data as described in claim 1, characterized in that: In step S1, the transportation route to be evaluated is divided into multiple continuous segments according to the characteristics of road infrastructure. Specifically, this includes dividing the route along the trajectory according to road grade change points, administrative division boundaries, start and end points of large structures, and service area locations. After division, the length of a single segment does not exceed 20km, and segments with a length of less than 3km and the same road grade as adjacent segments are merged into adjacent segments.

3. The commercial vehicle route risk assessment method based on spatiotemporal coupling of multi-source sensor data as described in claim 1, characterized in that: In step S3, setting a differentiated time window based on alarm type for deduplication and merging specifically means: for the same vehicle and the same alarm type, if the interval between the trigger timestamp of the later alarm and the timestamp of the previous valid alarm is less than or equal to the corresponding time window value, they are merged into the same valid event; the differentiated time window includes: 5 seconds for DMS driving status alarms, 10 seconds for ADAS emergency deceleration alarms, 5 seconds for ADAS lane departure alarms, 10 seconds for ADAS forward collision warning and close proximity alarms, and overspeed alarms are counted as 1 time for continuous overspeeding states.

4. The commercial vehicle route risk assessment method based on spatiotemporal coupling of multi-source sensor data as described in claim 1, characterized in that: The step S4, which involves associating the deduplicated alarm events with specific road segments using a map matching algorithm, includes: constructing a hidden Markov model; calculating the observation probability based on the vertical projection distance from the GPS trajectory point to the candidate road segment; calculating the inter-segment transition probability based on the difference between the shortest path of the road network between adjacent trajectory points and the GPS spherical distance; and using the Viterbi algorithm to decode and obtain the optimal matching road segment for each GPS trajectory point.

5. The commercial vehicle route risk assessment method based on spatiotemporal coupling of multi-source sensor data as described in claim 1, characterized in that: In step S4, the formula for calculating the normalized index value per 100 kilometers for each alarm type of each road segment is as follows: ； In the formula, This represents the total number of valid events for a specific alarm type on a certain road section after deduplication. This represents the cumulative mileage traveled by all vehicles on this road segment within the statistical period. The unit is "times per 100 kilometers".

6. The commercial vehicle route risk assessment method based on spatiotemporal coupling of multi-source sensor data as described in claim 1, characterized in that: The formula for the linear decay scoring model in step S5 is: ； In the formula: These are the normalized or original values ​​of the sub-indicators. This refers to the sensitivity coefficient of the sub-indicator, which is calibrated based on the correlation coefficient between each sub-indicator and the historical accident rate. A stronger correlation indicates a higher sensitivity coefficient. The larger the value.

7. The commercial vehicle route risk assessment method based on spatiotemporal coupling of multi-source sensor data as described in claim 1, characterized in that: The preset dimensions in step S5 include road segment feature risk dimensions. Traffic flow risk dimension Environmental and climate risk dimensions Operational load risk dimension and historical operational risk dimensions The weights for each dimension are 15%, 25%, 15%, 15%, and 30%, respectively; the original value of the line risk coefficient. The calculation formula is: ； When the number of valid sub-indicators in any dimension is less than 2, that dimension is marked as having insufficient data, and its original weights are proportionally redistributed to other valid dimensions.

8. The commercial vehicle route risk assessment method based on spatiotemporal coupling of multi-source sensor data as described in claim 1, characterized in that: The preset physical security risk triggering conditions in step S6 include at least one of the following: the road segment has a record of fatal traffic accidents in the past 12 months; the road segment has a continuous area with no mobile communication signal coverage for more than 20km. The road section has a historical high-risk incident frequency exceeding 5 times per 100 kilometers; The section has continuous downhill sections exceeding 10km and an average longitudinal slope exceeding 4%; the correction logic is as follows: when any triggering condition is met and the original value of the route risk coefficient is greater than 45, the final route risk coefficient is set to 44 points.

9. The commercial vehicle route risk assessment method based on spatiotemporal coupling of multi-source sensor data as described in claim 1, characterized in that: The safety management strategy in step S7 includes driver qualifications and scheduling requirements, monitoring frequency control instructions, and rest stop layout rules; the monitoring frequency control instructions set different monitoring strategy parameters according to the risk level and sent them to the monitoring and dispatching platform; the rest stop layout rules dynamically adjust the rest interval according to the operational load risk dimension score.

10. The commercial vehicle route risk assessment method based on spatiotemporal coupling of multi-source sensor data as described in claim 1, characterized in that: The static attribute data of each road segment in step S2 includes road geometric feature data, traffic flow feature data, meteorological environment data, historical accident data, and communication signal coverage data; the dynamic sensing data includes DMS alarm data, ADAS alarm data, vehicle speeding alarm data, and GPS trajectory data.

Images