Vehicle event risk assessment method and system

Abstract

The present disclosure relates to vehicle event risk assessment methods, systems, and non-transitory computer-readable storage media. A vehicle-side telematics device receives vehicle bus, inertial, and positioning telemetric streams at different sampling rates, synchronizes the telemetric streams based on timestamps, segments trips, extracts driving event windows, and normalizes by effective driving mileage or effective driving duration to generate a time-series feature tensor containing features, missing mask, and quality metadata, and sends the time-series feature tensor to a server. The server inputs the time-series feature tensor into a competing risk time-series model trained on right-censored samples, obtains conditional hazard rates and uncertainties for multiple vehicle event types over multiple future time intervals, and selects a base or enhanced sampling configuration based on the quality metadata and the uncertainties to control sampling rates, event window lengths, or upload priorities of subsequent telemetric data.

Description

Technical Field

[0001] This invention relates to vehicle telemetry, edge computing, vehicular communication, and machine learning technologies. In particular, it relates to a method where vehicle-side telemetry equipment and a communicating server process multiple vehicle data streams with different sampling rates and clock references. This method utilizes a competing risk time series model to generate conditional hazard rates and uncertainties for multiple vehicle event types within multiple future time intervals, and controls the acquisition and uploading of subsequent vehicle telemetry data based on data quality and prediction uncertainty. The invention also relates to a vehicle event risk assessment system and a non-transitory computer-readable storage medium for implementing the described method. Background Technology

[0002] During vehicle operation, multi-source telemetry data can be obtained from the vehicle bus, inertial measurement unit, positioning interface, and component status interface. Different data sources typically have different sampling rates, different timestamp references, and different data integrity levels. Directly splicing or aggregating these data may cause cross-sensor event misalignment and misinterpret sensor outages as vehicle status or the absence of driving events.

[0003] Mileage, driving time, road type, weather conditions, and traffic conditions can vary significantly between different vehicles and trips. If only the total number of driving events or their duration are used, vehicles with higher driving volume may only show a higher number of events due to higher exposure, thus reducing the comparability between vehicles and between different periods.

[0004] At the end of the historical observation period, some vehicles may not have experienced the target vehicle event, or the vehicles may have left the observation before the event occurred. These right-censored samples simultaneously contain information about "the event not having occurred within the observed period" and "the event potentially occurring after the observation end." If ordinary classification models ignore right-censoring information, they may not consistently represent the event occurrence time, event type, and survival probability.

[0005] Furthermore, continuously acquiring all sensor data channels at a high sampling rate and uploading raw data increases vehicle-side power consumption, storage footprint, and wireless communication bandwidth. Existing risk assessment schemes typically only output risk scores without utilizing data quality and prediction uncertainty to control subsequent physical sensor acquisition, event window retention, and upload queues. Therefore, it is difficult to establish a closed-loop balance between information quality and resource consumption.

[0006] Therefore, there is a need for a technical solution that can reliably synchronize, segment, and normalize exposure of multi-rate vehicle telemetry data, generate time-related risks of multiple event types using right-censored samples, and modify subsequent data acquisition or uploading based on input quality and prediction uncertainty. Summary of the Invention

[0007] In some embodiments, the present invention provides a vehicle event risk assessment method performed by a vehicle-side telemetry device and a server communicating with the vehicle-side telemetry device. The method includes: the vehicle-side telemetry device receiving multiple telemetry streams with different sampling rates from a vehicle bus interface, an inertial measurement unit interface, and a positioning interface, each telemetry stream including a timestamp and a sensor status identifier; the vehicle-side telemetry device estimating the clock deviation between the multiple telemetry streams based on the timestamps and synchronizing the multiple telemetry streams on a common time axis to generate synchronized telemetry data; the vehicle-side telemetry device extracting one or more telemetry data items from the synchronized telemetry data for determining the vehicle's driving state and vehicle motion state, and generating a vehicle driving state and vehicle motion state determination result based on the extracted telemetry data items; the vehicle-side telemetry device dividing the synchronized telemetry data into multiple trips based on the vehicle driving state and vehicle motion state determination result, and extracting a driving event window for each trip; and the vehicle-side telemetry device dividing the number of events in each driving event window by the effective mileage or effective driving distance of the corresponding trip. The exposure normalization feature is generated over a duration, and the exposure normalization feature, a mask indicating missing samples, and quality metadata indicating time synchronization quality are combined into a time-series feature tensor. The vehicle-side telemetry device sends the time-series feature tensor to the server. The server inputs the time-series feature tensor into a competitive risk time-series model trained with training samples marked with right censoring to generate conditional hazard rates and uncertainties for multiple future time intervals and multiple vehicle event types. The server selects a first sampling configuration or a second sampling configuration based on the quality metadata and the uncertainty, wherein the second sampling configuration enables at least one sensor data channel to have a higher sampling rate, a longer event window, or a higher upload priority than the first sampling configuration. The server sends the selected sampling configuration to the vehicle-side telemetry device, causing the vehicle-side telemetry device to collect or upload subsequent telemetry data according to the selected sampling configuration.

[0008] In some implementations, the vehicle driving state includes the ignition state of a fuel-powered vehicle or the high-voltage ready state of an electric vehicle; the vehicle motion state is determined based on at least two of vehicle speed, gear position, and positioning displacement; and a journey ends only after the vehicle driving state is non-ready and the vehicle motion state meets the stationary condition for a predetermined tolerance duration.

[0009] In some implementations, the synchronization process includes: estimating the offset and drift rate of each sensor clock relative to a monotonic clock; resampling the numerical channels to a common time grid; performing hold processing on the discrete state channels; and setting the mask for samples exceeding the maximum interpolation interval.

[0010] In some implementations, the driving event window includes at least one of the following events: rapid acceleration, rapid deceleration, sharp turning, speeding, following too closely, night driving, or abnormal attention, and each event has a pre-event window and a post-event window.

[0011] In some implementations, the exposure normalization features are further context-layered according to road type, weather conditions, day / night conditions, or traffic congestion conditions, so that the same event corresponds to different feature channels in different contexts.

[0012] In some implementations, the quality metadata includes at least three of the following: completeness, timestamp deviation, cross-sensor consistency, data freshness, and effective exposure, and the method prohibits the output of a deterministic risk level for automatic control when the effective exposure is less than a threshold.

[0013] In some implementations, the competition risk timing model includes a behavior sequence encoder, a component status encoder, a fusion layer, a discrete-time hazard rate output header, and an event type output header; the component status encoder receives the remaining lifespan percentage generated based on the component's cumulative operating time, fault codes, maintenance records, and expected lifespan.

[0014] In some implementations, the fusion layer assigns weights to the behavioral sequence representation and the component state representation based on the mask and the quality metadata, thereby reducing the contribution of input channels with high missing rates or low freshness to the fused representation.

[0015] In some implementations, the competitive risk time series model is trained by maximizing the joint likelihood of the event likelihood term containing uncensored samples and the survival probability term containing right-censored samples, and using training, calibration and test sets partitioned chronologically.

[0016] In some implementations, the uncertainty is determined by at least one of the prediction variance among the model set, the variance of random inactivation repeated inference, or the width of the conformal prediction interval.

[0017] In some implementations, selecting the second sampling configuration includes: increasing the sampling rate of the inertial measurement unit in response to the inertial data channel missing rate exceeding a first threshold; shortening the time synchronization message interval in response to the positioning data channel time deviation exceeding a second threshold; and / or retaining and uploading the raw data window around the candidate event in response to the event type uncertainty exceeding a third threshold.

[0018] In some implementations, the server determines the data quality gap of each sensor data channel based on the quality metadata, determines the uncertainty contribution of each sensor data channel to the conditional hazard rate based on the uncertainty, combines the data quality gap and the uncertainty contribution into a channel-level information gap, and adjusts the sampling rate, time synchronization message interval, event window length, cache retention period, or upload priority for sensor data channels with larger channel-level information gaps.

[0019] In some implementations, the method further includes: generating a risk data object, the risk data object including a model version identifier, a feature version identifier, an observation time interval, a time distribution of each vehicle event type, a confidence interval, a quality score, and a selected sampling configuration identifier; and writing the risk data object into an immutable audit log.

[0020] In some implementations, the method further includes: comparing the online feature distribution with the training baseline distribution to generate a drift metric; and when the drift metric exceeds a drift threshold, freezing the automatic policy based on risk output, switching to a predetermined rule model, and triggering model retraining or recalibration.

[0021] In some implementations, the plurality of vehicle event types include collision events, vehicle component failure events, glass damage events, and body damage events.

[0022] In some implementations, the method further includes: determining the priority of security alerts or maintenance appointments based on the risk data object.

[0023] In some embodiments, the present invention provides a vehicle event risk assessment system. The system includes: a vehicle-side telemetry device, comprising a vehicle bus interface, an inertial measurement unit interface, a positioning interface, a wireless communication interface, a first processor, and a first memory; and a server communicating with the vehicle-side telemetry device, the server comprising a second processor and a second memory; wherein the first memory stores instructions for the first processor to perform the following operations: receiving multiple telemetry streams with different sampling rates and each including a timestamp; estimating the clock deviation between the multiple telemetry streams; synchronizing the multiple telemetry streams on a common time axis to generate synchronized telemetry data; extracting one or more telemetry data items from the synchronized telemetry data for determining the vehicle's driving state and vehicle motion state; generating vehicle driving state and vehicle motion state determination results; dividing the synchronized telemetry data into trips; extracting driving event windows for each trip; and generating… The second processor generates a time-series feature tensor containing exposure normalization features, a missing mask, and time synchronization quality metadata. This time-series feature tensor is then sent to the server via the wireless communication interface. Subsequent telemetry data is collected or uploaded according to the sampling configuration received from the server. The second memory stores instructions that cause the second processor to perform the following operations: receive the time-series feature tensor from the vehicle-side telemetry device; input the time-series feature tensor into a competitive risk time-series model trained with training samples marked with right censoring to obtain conditional hazard rates and uncertainties for multiple future time intervals and multiple vehicle event types; select a sampling configuration with different sampling rates, event window lengths, or upload priorities based on the time synchronization quality metadata and the uncertainty; and send the selected sampling configuration to the vehicle-side telemetry device.

[0024] In some implementations, the vehicle-side telemetry device includes a circular cache for storing raw event windows in chronological order when the network is unavailable, and the sampling configuration further specifies the retention period and upload order of the circular cache.

[0025] In some implementations, the server includes a model service, feature storage, a configuration distribution interface, and an audit log; the configuration distribution interface signs the sampled configuration, and the vehicle-side telemetry device applies the sampled configuration only after verifying the signature and version number.

[0026] Through the above scheme, this invention improves the temporal consistency of asynchronous multi-source vehicle data, distinguishing missing data from true zero events, and enhances the comparability of features between different trips or vehicles through exposure normalization. Furthermore, the risk model output is used to modify the sampling rate, synchronization interval, event window, or upload priority of subsequent physical sensors or data transmission paths, thereby improving the quality of risk assessment input while limiting the consumption of vehicle-side computing, storage, and communication resources. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of a vehicle incident risk assessment system according to some embodiments of the present invention.

[0028] Figure 2 This is a flowchart of a vehicle event risk assessment method performed by a vehicle-side telemetry device and a server communicating with it, according to some embodiments of the present invention.

[0029] Figure 3 This is a schematic diagram of the structure of a competition risk time series model according to some embodiments of the present invention.

[0030] Figure 4 This is a schematic diagram of the training process of a model with right-censored samples and time sequence segmentation according to some embodiments of the present invention.

[0031] Figure 5 This is a schematic diagram illustrating the selection and feedback control of sampling configuration based on quality metadata and uncertainty according to some embodiments of the present invention. Detailed Implementation

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments described are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Unless the context explicitly defines it, the terms "comprising," "including," or any other variations thereof used in this specification are intended to cover non-exclusive inclusion; expressions such as "first," "second," etc., are used only to distinguish different technical objects and do not imply that there is necessarily a difference in temporal order, spatial order, or degree of importance between the corresponding technical objects. Unless explicitly stated that a particular embodiment is a working embodiment that has been implemented, the numerical values, thresholds, model outputs, and processing results in this specification are illustrative or anticipatory examples used to illustrate how the corresponding technical solutions can be implemented, and do not imply that the corresponding experiments, training, or deployments have actually been completed or that the listed results have been obtained.

[0033] In this invention, vehicle event risk assessment is performed collaboratively by vehicle-side telemetry equipment and a server communicating with it. The vehicle-side telemetry equipment receives, synchronizes, segments, and preprocesses vehicle telemetry data, and collects or uploads subsequent telemetry data according to the sampling configuration issued by the server. The server runs a competing risk time series model, generates conditional hazard rates and uncertainties for multiple future time intervals and multiple vehicle event types, and forms feedback sampling control based on quality metadata and the uncertainties.

[0034] The technical problems to be solved in the prior art include at least the following: data misalignment caused by different sampling rates and clock references used by different vehicle sensors; sensor disconnection being misinterpreted as zero event quantity; incomparability of features due to differences in effective driving mileage and effective driving time for different vehicles or different trips; and increased vehicle-side power consumption, storage usage and wireless communication bandwidth caused by continuously using high sampling rates and uploading raw data.

[0035] By synchronizing multiple telemetry streams on a common timeline, separating missing masks from numerical features, normalizing driving events according to effective exposure, and selecting different sampling configurations based on the uncertainty of quality metadata and conditional hazard rates, this invention improves the temporal consistency and interpretability of multi-source vehicle data. The model output can be used to modify the sampling rate, event window length, time synchronization message interval, or upload priority of subsequent physical sensor data. Even when all vehicles do not require long-term enhanced sampling, data can be supplemented for channels with insufficient information, thereby improving the quality of subsequent risk assessment inputs while limiting the consumption of vehicle-side computing, storage, and communication resources.

[0036] In this disclosure, "vehicle event" refers to a vehicle-related event that may occur within a future time interval and can be characterized by vehicle operation data, component status data, or event records, including but not limited to collision events, vehicle component failure events, glass damage events, and body damage events. Vehicle event types are the underlying technical output of the competitive risk time series model and can be used for vehicle safety monitoring, maintenance resource scheduling, or manual review.

[0037] In this disclosure, "conditional hazard rate" represents the conditional probability, or the corresponding risk, of a specific vehicle event type occurring first in the current future time interval, provided that no target vehicle event has occurred in the previous future time interval. "Right censoring flag" indicates that no target vehicle event was observed before the observation endpoint of the training samples, but the vehicle event may still occur after the observation endpoint.

[0038] In this disclosure, "quality metadata" refers to data describing whether input data is suitable for risk inference, including but not limited to completeness, timestamp deviation, cross-sensor consistency, data freshness, and effective exposure. "Uncertainty" refers to a predictive reliability index characterized by model parameter differences, inference randomness, or the width of the prediction interval based on calibration data. Calibration error and input distribution drift serve as reliability indicators for risk output calibration gating and drift monitoring, respectively, and are not necessarily equivalent to the aforementioned uncertainty. "Sampling configuration" refers to machine-readable control data executable by vehicle-side telemetry equipment, specifying at least one of the following: sampling rate, event window length, time synchronization message interval, cache retention strategy, or upload priority. "Telemetry data item" refers to a specific signal item or status item in synchronous telemetry data, such as vehicle speed, gear position, positioning displacement, longitudinal acceleration, or ignition status. "Sensor data channel" refers to a data source channel or signal channel controllable by the sampling configuration, which may correspond to a set of data output from a vehicle bus interface, inertial measurement unit interface, or positioning interface, or to one or more specific signal items output from the aforementioned interfaces. The following embodiments can be implemented individually or combined with each other where there is no technical conflict. The parameter ranges, thresholds, model structures, or execution entities described in one embodiment are merely examples and do not constitute a limitation on the scope of the claims.

[0039] General description of vehicle incident risk assessment system and vehicle incident risk assessment method

[0040] Figure 1 This is a schematic diagram of a vehicle incident risk assessment system according to some embodiments of the present invention. Figure 1 As shown, the vehicle incident risk assessment system 100 includes a vehicle-side telemetry device 110, a server 150, and a downstream application 170 capable of receiving risk assessment results. The vehicle-side telemetry device 110 is connected to the server 150 via a communication link. The server 150 is also connected to the downstream application 170 via a communication link. The communication link can include a wireless network or a wired network. The wireless network can be WLAN, 5G, 4G, or LTE networks.

[0041] The vehicle-side telemetry device 110 can receive vehicle bus data, inertial data, and positioning data through the vehicle bus interface 112, inertial measurement unit interface 114, and positioning interface 116, respectively, thereby forming multiple telemetry streams 130. The synchronization, travel, and event preprocessing module 132 can perform time synchronization, travel segmentation, and driving event window extraction on the multiple telemetry streams 130, and generate a temporal feature tensor 134 including a feature plane X, a missing mask M, and quality metadata Q. The wireless communication interface 118 can send the temporal feature tensor 134 and related metadata to the server 150, and receive sampling configurations from the server 150.

[0042] Server 150 may include or run model service 156, feature storage 158, and configuration control interface 160. Model service 156 generates conditional hazard rate output 157 and uncertainty output 159 based on time-series feature tensor 134. The configuration selector in configuration control interface 160 selects subsequent sampling configurations based on quality metadata carried in time-series feature tensor 134 or provided by feature storage 158 and uncertainty output 159, and may further combine conditional hazard rate output 157 to determine target vehicle event type and target future time interval. Quality metadata may be provided to configuration control interface 160 as part of time-series feature tensor 134. Server 150 also generates risk data object 161 and writes it to audit log 162; the audit log 162 may be implemented as an immutable audit log. Risk data object 161 may be provided to downstream applications 170 such as safety alerts, maintenance resource scheduling, or manual review.

[0043] Figure 2 A vehicle event risk assessment method 200, performed by a vehicle-side telemetry device and a server communicating with it according to some embodiments of the present invention, is illustrated. The method may include steps S201 to S207. Steps S201 to S204 are primarily performed by the vehicle-side telemetry device, which completes multi-channel telemetry stream reception, synchronization, trip segmentation, driving event window extraction, exposure normalization, and temporal feature tensor generation. Steps S205 to S207 are performed collaboratively by the vehicle-side telemetry device and the server, wherein the vehicle-side telemetry device sends the temporal feature tensor to the server, the server performs model inference and sampling configuration selection in steps S205 and S206, and sends the selected sampling configuration to the vehicle-side telemetry device in step S207.

[0044] Receive multiple telemetry streams with different sampling rates

[0045] In step S201, the vehicle-side telemetry device receives multiple telemetry streams with different sampling rates from the vehicle bus interface, the inertial measurement unit, and the positioning interface. Each telemetry stream may include a timestamp and sensor status identifiers. The vehicle bus interface may be a CAN interface, CAN FD interface, LIN interface, Ethernet interface, or OBD interface. The inertial measurement unit may output at least one of longitudinal acceleration, lateral acceleration, vertical acceleration, yaw rate, pitch rate, or roll rate. The positioning interface may receive satellite positioning position, velocity, heading, positioning accuracy, or positioning status.

[0046] The sampling rates of each telemetry stream can differ. For example, ignition status, gear position, and fault codes in the vehicle bus can be sampled at a sampling rate of 1 Hz to 10 Hz; vehicle speed, brake pedal opening, and steering wheel angle can be sampled at a sampling rate of 10 Hz to 50 Hz; the inertial measurement unit can be sampled at a sampling rate of 20 Hz to 200 Hz; and the positioning interface can be sampled at a sampling rate of 1 Hz to 10 Hz. Each sample, in addition to the source timestamp, may also include a receive timestamp, channel identifier, serial number, and sensor status identifier. The sensor status identifier can indicate normal, initialization, unlocked, overload, calibration in progress, communication abnormality, or invalid data.

[0047] Generate synchronous telemetry data and generate vehicle driving status and vehicle motion status determination results.

[0048] In step S202, the vehicle-side telemetry device estimates the clock deviation between the multiple telemetry streams based on the timestamps, and performs synchronization processing on the multiple telemetry streams on a common time axis to generate synchronized telemetry data. The common time axis can be formed by the monotonic clock of the vehicle-side telemetry device, or by a monotonic clock corrected for positioning and timing. Using a monotonic clock can reduce the impact of manual adjustment of system time, network timing jumps, or daylight saving time changes on telemetry synchronization.

[0049] For the i-th telemetry stream, its offset a_i and drift rate b_i relative to the common time axis can be estimated within the synchronization window, and the source timestamp t_i can be converted to the corrected timestamp t_i′ according to the correction relation t_i′ = a_i + b_i·t_i. The offset and drift rate can be estimated through time synchronization messages, cross-sensor events with common physical meaning, sequence number continuity, or received timestamps.

[0050] In one specific implementation, multiple time-corresponding points (t_i,n, t_ref,n) are collected within a sliding synchronization window of length W. a_i and b_i are then solved using least-squares fitting or robust regression with outlier suppression to minimize Σ_n[t_ref,n-(a_i+b_i·t_i,n)]². Vehicle-side telemetry equipment can update a_i and b_i at predetermined intervals. When the fitting residual, the change in drift rate, or the number of valid corresponding points does not meet the conditions, the corresponding synchronization quality is marked as low quality, and the generation of a deterministic risk level is avoided by relying solely on this channel.

[0051] In step S202, the vehicle-side telemetry device further extracts one or more telemetry data items from the synchronous telemetry data to determine the vehicle's driving state and vehicle motion state, and generates vehicle driving state and vehicle motion state determination results based on the extracted telemetry data items. The telemetry data items may be specific signal items or status items in the synchronous telemetry data, rather than the entire telemetry data. The vehicle driving state may be determined based on at least one of the following in the vehicle bus telemetry stream: ignition state, engine running state, high-voltage system ready state, drive system ready state, gear position, or drive motor torque request. The vehicle motion state may be determined based on at least two of the following: vehicle speed and gear position in the vehicle bus telemetry stream, and positioning displacement in the positioning telemetry stream.

[0052] In some implementations, the vehicle-side telemetry equipment can also utilize longitudinal acceleration, lateral acceleration, or angular velocity from the inertial measurement unit telemetry stream to verify the vehicle's motion state, or generate a degraded vehicle motion state determination result when vehicle speed or positioning signals are unavailable. Therefore, the vehicle driving state and vehicle motion state determination results are not additional data sources independent of the multiple telemetry streams, but rather intermediate processing results determined by the vehicle-side telemetry equipment based on relevant data from time-synchronized vehicle bus telemetry streams, inertial measurement unit telemetry streams, and positioning telemetry streams.

[0053] Divide the trip and extract the driving event window

[0054] In step S203, the vehicle-side telemetry device divides the synchronous telemetry data into multiple trips based on the vehicle's driving state and motion state determination results, and extracts a driving event window for each trip. A trip represents the continuous operating interval of the vehicle from when it is ready to drive and begins to move until it is not ready to drive and remains stationary. Overlapping data for a predetermined time can be retained at the trip boundaries to avoid truncating driving events located near the boundaries.

[0055] The driving event window is built around the candidate driving event and includes at least a pre-event window describing the state before the candidate driving event and a post-event window describing the restored state or consequences after the candidate driving event. The lengths of the pre-event and post-event windows can be fixed or adjusted according to the event type, speed, or uncertainty.

[0056] Generate temporal feature tensors

[0057] In step S204, the vehicle-side telemetry device determines the event quantity in each driving event window. When forming trip-level or observation time interval-level features, the event quantities of multiple driving event windows belonging to the same event type or the same context within the same trip can be aggregated, and the event quantity in a single driving event window or the aggregated event quantity can be divided by the effective mileage or effective driving time of the corresponding trip or context to generate exposure normalization features. The event quantity can be the number of events, the cumulative duration of events, the integral exceeding a threshold, the sum of event severity, or the quantile of event severity. The effective mileage is the mileage after excluding positioning failures, towing transport, idling without driving, or invalid data intervals. The effective driving time is the driving time after excluding invalid data intervals, long periods of stationary driving, or intervals that do not meet driving state conditions.

[0058] For example, the mileage-normalized event rate of the e-th event in the r-th trip can be expressed as R_e,r = N_e,r / max(D_r, ε), where N_e,r is the single event quantity of the e-th event in the r-th trip or the total event quantity obtained by aggregating the event quantities of multiple similar driving event windows, D_r is the effective driving mileage, and ε is the minimum positive number to prevent division by zero. The duration-normalized feature can be expressed as P_e,r = L_e,r / max(T_r, ε_t), where L_e,r is the cumulative duration of the event, T_r is the effective driving time, and ε_t is the minimum effective duration.

[0059] The vehicle-side telemetry equipment combines the exposure normalization features, the mask indicating missing samples, and the quality metadata indicating time synchronization quality into a time-series feature tensor. This time-series feature tensor can be organized in the form of "journey or time interval × feature channel × data category," where the data category includes at least numerical features, the missing mask, and quality metadata. The missing mask and the masked data are stored as different channels, enabling the model to distinguish between "zero event quantity" and "no valid samples obtained."

[0060] As a non-restrictive implementation, the numerical features of the most recent L runs form a matrix X ∈ R^(L×F), the corresponding missing mask forms a matrix M ∈ {0,1}^(L×F), and the run-level quality metadata forms a matrix Q ∈ R^(L×G). X, M, and Q are then concatenated along the feature dimension to form the temporal feature tensor Z. When standardizing the numerical features, the mean, standard deviation, quantiles, or cutoff ranges determined by the training set and managed through feature version identifiers are used. For missing locations, zero values, the training set mean, or predetermined placeholder values ​​can be filled in, but since the model also receives M, the placeholder values ​​are not interpreted as true measurements.

[0061] Generation condition hazard rate and uncertainty

[0062] In step S205, the vehicle-side telemetry device sends the time-series feature tensor to the server. The server inputs the time-series feature tensor into a competing risk time-series model trained with training samples marked with right censoring, to generate conditional hazard rates and uncertainties for multiple future time intervals and multiple vehicle event types. The multiple future time intervals can be continuous and non-overlapping discrete time intervals, such as 0 to 7 days, 8 to 30 days, 31 to 90 days, and 91 to 180 days; or they can be daily, weekly, or monthly time intervals of the same length.

[0063] For each future time interval, the competitive risk time series model simultaneously outputs conditional hazard rates for multiple vehicle event types, creating a competitive relationship between different vehicle event types for the "first occurring vehicle event." Based on the conditional hazard rates for each future time interval, the event time distribution, survival probability, and cumulative occurrence probability for each vehicle event type can be further obtained. The uncertainty is used to characterize the reliability of the conditional hazard rate being affected by model parameters, training samples, input quality, or distribution shift.

[0064] Select the first sampling configuration or the second sampling configuration.

[0065] In step S206, the server selects either a first sampling configuration or a second sampling configuration based on the quality metadata and the uncertainty. The first sampling configuration can be a basic sampling configuration, used to control regular data acquisition and uploading when the quality metadata meets the quality threshold and the uncertainty does not exceed the uncertainty threshold; the second sampling configuration can be an enhanced sampling configuration, used to supplement information when the quality metadata does not meet the quality threshold or the uncertainty exceeds the uncertainty threshold.

[0066] The second sampling configuration allows at least one sensor data channel to have a higher sampling rate, a longer event window, or a higher upload priority than the first sampling configuration. For example, the first sampling configuration may specify that the inertial measurement unit (IMU) acquires data at 20 Hz and uploads only aggregated features, while the second sampling configuration may specify that the IMU acquires data at 50 Hz, extends the pre-event window and post-event window to 30 seconds and 60 seconds respectively, and prioritizes uploading the raw inertial data corresponding to candidate events.

[0067] Send and apply the selected sampling configuration

[0068] In step S207, the server sends the selected sampling configuration to the vehicle-side telemetry device, enabling the device to collect or upload subsequent telemetry data according to the selected configuration. The sampling configuration may include a configuration identifier, version number, effective time, expiration time, applicable channel, sampling rate, event window length, upload priority, maximum duration, maximum data volume, and recovery conditions. After applying the new sampling configuration, the vehicle-side telemetry device can return configuration confirmation information to the server.

[0069] The vehicle event risk assessment method disclosed herein establishes a closed loop between the vehicle-side telemetry equipment and the server: the vehicle-side telemetry equipment provides a time-series feature tensor that has been synchronized, segmented, and normalized for exposure; the server generates risk results and uncertainties based on a competing risk time-series model; and the server then uses quality metadata and uncertainties to adjust the acquisition or uploading of subsequent telemetry data. Therefore, it is not necessary to maintain high sampling rates and high upload bandwidth for all vehicles for extended periods to supplement information for vehicles and sensor data channels with insufficient data quality or high prediction uncertainty.

[0070] Travel division based on vehicle driving state and vehicle motion state

[0071] In the vehicle event risk assessment method disclosed herein, the vehicle driving state includes the ignition state of a fuel-powered vehicle or the high-voltage ready state of an electric vehicle. For fuel-powered vehicles, the ignition state can be jointly determined by the ignition switch state, engine speed, and engine operating state in the vehicle bus; for electric vehicles, the high-voltage ready state can be jointly determined by the high-voltage system ready signal, the drive system ready signal, the gear position, and the drive motor torque request.

[0072] In the vehicle incident risk assessment method disclosed herein, the vehicle motion state includes at least two of the following: vehicle speed, gear position, and positioning displacement. Using at least two vehicle motion states for cross-judgment can reduce erroneous travel boundaries caused by single speed signal lag, gear signal delay, or positioning drift. For example, when the vehicle speed is greater than 1 km / h and the positioning displacement is greater than 5 m within 10 seconds, the vehicle is determined to be in motion; or, when the gear is in forward or reverse and the vehicle speed is greater than 0.5 km / h, the vehicle is determined to be in motion.

[0073] A trip can begin when the vehicle's driving state becomes ready and the vehicle's motion state meets the start-up conditions. A trip ends only after the vehicle's driving state is not ready and the vehicle's motion state meets the stationary condition for a predetermined tolerance duration. The stationary condition can be determined by at least two of the following: vehicle speed is below a vehicle speed threshold, positioning displacement is below a displacement threshold, and the gear is in a non-driving gear. The predetermined tolerance duration can be 30 seconds, 60 seconds, 120 seconds, or other durations set according to the vehicle type.

[0074] For example, when a gasoline-powered vehicle temporarily shuts off the engine using the start-stop system, the ignition state remains ready, and therefore the trip does not end; similarly, when an electric vehicle is briefly stopped, the high-voltage ready state remains ready, and therefore the trip does not end. Even if the vehicle's driving state briefly becomes inactive, the trip will not end as long as the vehicle's motion state does not fall below the stationary threshold for a predetermined tolerance period.

[0075] Synchronization processing based on a monotonic clock

[0076] In the vehicle incident risk assessment method disclosed herein, the synchronization processing based on a monotonic clock may include: estimating the offset and drift rate of each sensor clock based on a monotonic clock; resampling the numerical channels to a common time grid; performing hold processing on the discrete state channels; and setting the mask for samples exceeding the maximum interpolation interval.

[0077] The time interval of the common time grid can be determined according to the required time resolution of the model. For example, the common time grid can be 100 ms. For numerical channels such as vehicle speed, acceleration, and steering wheel angle, linear interpolation, spline interpolation, nearest neighbor interpolation, or physical constraint-based interpolation can be used to resample to the common time grid. For discrete state channels such as ignition status, gear position, and fault code status, zero-order hold processing can be performed, that is, the previous valid state is held until the next valid discrete state sample is obtained.

[0078] To avoid unreliable filling of long-term data interruptions, maximum interpolation intervals are set for different channels. For example, the maximum interpolation interval for the inertial measurement unit can be 300 ms, the maximum interpolation interval for the vehicle bus numerical channel can be 1 second, and the maximum interpolation interval for the positioning data channel can be 3 seconds. When the time interval between a target sample on a common time grid and an adjacent valid source sample exceeds the corresponding maximum interpolation interval, no inferred value is generated, or the corresponding value is set to a predetermined placeholder value, and the corresponding position of the missing mask is set to the value indicating the missing sample.

[0079] Time synchronization quality can be determined based on timestamp discrepancies before and after correction, offset estimation residuals, drift rate stability, sequence number continuity, and cross-sensor event alignment. For example, the time of rapid brake pedal increase can be compared with the time of longitudinal deceleration; if the corrected time difference between the two is less than an allowable threshold, the cross-sensor consistency score is improved.

[0080] Types and extraction of driving event windows

[0081] In the vehicle event risk assessment method disclosed herein, the driving event window includes at least one event selected from rapid acceleration, rapid deceleration, sharp steering, speeding, close following, nighttime driving, or abnormal attention, and each event has a pre-event window and a post-event window. Event detection can be triggered by a fixed threshold, a vehicle model adaptive threshold, a driver baseline threshold, or a combination of multiple signal rules. For nighttime driving, the event can refer to the vehicle entering a nighttime driving state, the continuous nighttime driving duration reaching a predetermined threshold, or a representative trigger point within a nighttime driving interval; the pre-event window and the post-event window can be established around the entry point, the threshold-reaching point, or the representative trigger point, respectively. The nighttime state can also be used independently as a contextual hierarchical condition.

[0082] Rapid acceleration events can be determined based on longitudinal acceleration exceeding a first acceleration threshold and lasting for more than a first duration; rapid deceleration events can be determined based on longitudinal acceleration falling below a second acceleration threshold and lasting for more than a second duration; sharp steering events can be determined based on at least two of yaw rate, lateral acceleration, and steering wheel angular velocity; speeding events can be determined based on the difference between vehicle speed and road speed limit and its duration; following too closely events can be determined based on forward collision warning information, time distance, or target distance obtained from the vehicle bus interface; and attention deficit events can be determined based on the driver monitoring system status or steering wheel micro-operation characteristics obtained from the vehicle bus interface.

[0083] Each event has a pre-event window to preserve the speed, acceleration, gear, road conditions, or driving actions before the event is triggered, and a post-event window to preserve the recovery process, subsequent actions, or changes in vehicle status after the event. For example, a sudden deceleration event can have a 20-second pre-event window and a 30-second post-event window; an attention deficit event can have a 60-second pre-event window and a 60-second post-event window. Overlapping event windows of the same type can be merged, or representative events can be selected based on event severity.

[0084] In the vehicle incident risk assessment method disclosed herein, the exposure normalization feature is further context-layered according to road type, weather conditions, day / night conditions, or traffic congestion conditions, so that the same event corresponds to different feature channels in different contexts. Road type may include highways, urban expressways, urban ordinary roads, and other roads; weather conditions may include sunny, rainy, snowy, foggy, or low visibility; day / night conditions may be determined based on local sunrise and sunset times, ambient light conditions, or preset time intervals; traffic congestion conditions may be determined based on average speed, speed fluctuations, number of stops, or external traffic conditions.

[0085] For example, sudden deceleration events form different characteristic channels such as "sudden deceleration rate on highways", "sudden deceleration rate on ordinary urban roads", "sudden deceleration rate in rainy weather", and "sudden deceleration rate at night". Even if the total number of sudden deceleration events is the same for two trips, they can form different temporal characteristic tensors due to differences in road type, weather conditions, day / night conditions, or traffic congestion conditions.

[0086] The event volume after context layering still uses the effective mileage or effective driving time in the corresponding context as the denominator, rather than the total mileage or total time of the entire trip. For example, the number of sudden decelerations during nighttime driving divided by the effective mileage during nighttime driving forms the nighttime sudden deceleration rate; the cumulative duration of close following in congested conditions divided by the effective driving time in congested conditions forms the percentage of close following duration in congested conditions.

[0087] When the effective exposure of a context is less than the minimum context exposure, the context feature can be marked as low reliability and the corresponding effective exposure can be recorded in the quality metadata, instead of directly interpreting the context feature as low risk.

[0088] Quality metadata and risk output gating

[0089] In the vehicle incident risk assessment method disclosed herein, the quality metadata includes at least three of the following: completeness, timestamp bias, cross-sensor consistency, data freshness, and effective exposure. Completeness can be determined by the ratio of the actual number of valid samples obtained to the theoretically required number of samples; timestamp bias can be determined by the residual bias or bias quantile after synchronization correction; cross-sensor consistency can be determined by the degree of consistency between physically related channels; data freshness can be determined by the time difference between the most recent valid sample time and the risk assessment time; effective exposure may include effective mileage, effective driving time, or effective exposure in a specific context.

[0090] Quality metadata can be calculated separately for channels, travel distances, observation time intervals, or the entire time-series feature tensor. For example, the inertial data channel integrity is 0.92, the 95th percentile of the timestamp deviation of the positioning data channel is 1.8 seconds, the cross-sensor consistency between the vehicle bus and the inertial measurement unit is 0.86, the data freshness is 12 hours, and the effective driving range is 180 km.

[0091] The server can generate a quality score Q_score based on at least three quality metadata items. For example, Q_score = w_cq_c + w_tq_t + w_sq_s + w_fq_f + w_eq_e, where q_c, q_t, q_s, q_f, and q_e represent the normalized scores for completeness, time synchronization quality, cross-sensor consistency, data freshness, and effective exposure, respectively, and w_c, w_t, w_s, w_f, and w_e are the corresponding weights. Alternatively, a single quality score can be avoided, and gating can be applied to each quality metadata item separately. The quality score Q_score is distinct from the quality metadata matrix Q in the time-series feature tensor.

[0092] When the effective exposure is less than a threshold, the method prohibits the output of deterministic risk levels for automatic control. The "deterministic risk level" refers to a discrete level used directly to trigger automatic control without accompanying quality limitations or uncertainty descriptions, such as "low risk," "medium risk," or "high risk." When the effective exposure is insufficient, the server can still retain conditional hazard rates, event time distributions, confidence intervals, and quality metadata for manual review, continued data collection, or non-automatic prompts, but will not send the deterministic risk level to the automatic policy interface.

[0093] Effective exposure thresholds can be set according to vehicle type, risk assessment timeframe, or vehicle event type. For example, for a 30-day collision event risk assessment, the effective mileage threshold could be 100 km; for vehicle component failure events, the corresponding component status data could be required to cover at least 20 hours of effective driving time. The structure and quality perception fusion of competing risk time-series models are also discussed.

[0094] The structure and fusion method of the competition risk time series model will be further explained below. (Refer to...) Figure 3 The competitive risk time series model includes a behavior sequence encoder, a component state encoder, a fusion layer, a discrete-time hazard rate output header, and an event type output header.

[0095] like Figure 3 As shown, the competition risk timing model 300 receives a behavioral timing feature tensor 301 and component status data 303. The behavioral timing feature tensor 301 may be generated by a vehicle-side telemetry device and includes at least a numerical feature plane X, a missing mask plane M, and a quality metadata plane Q aligned to the same time index or travel index. A behavioral sequence encoder 302 encodes the behavioral timing feature tensor 301, and a component status encoder 304 encodes the component status data 303. In some embodiments, the component status data 303 may be generated from cumulative operating time and fault codes obtained from the vehicle bus or diagnostic interface, maintenance records stored on a server, and vehicle model or component life data, and may be used as a server-side input distinct from the behavioral timing feature tensor 301. In other embodiments, at least a portion of the component status data 303 may also be used as a feature channel of the timing feature tensor. The mask and quality gating signal 305 are not merely used as ordinary numerical feature inputs, but are used to adjust the contribution of each encoded representation or feature channel in the quality-aware fusion layer 306. The fusion representation of the quality-aware fusion layer 306 is provided to the discrete-time hazard rate output head 308 and the event type output head 310, respectively, to generate conditional hazard rate, event time distribution, survival probability and confidence interval output 312.

[0096] Behavior sequence encoder

[0097] The behavior sequence encoder receives normalized exposure features, missing masks, and quality metadata related to driving behavior from a temporal feature tensor and generates a behavior sequence representation. The behavior sequence encoder can employ one-dimensional convolutional networks, temporal convolutional networks, gated recurrent networks, Transformers, or state-space models. It can process input at the trip, day, week, or sliding time window level, allowing earlier and more recent trips to have a distinguishable impact on risk assessment.

[0098] In one specific, but not limiting, implementation, the behavior sequence encoder receives temporal feature tensors from the most recent 30 runs. Each run is first mapped to a d-dimensional vector via linear projection, and then encoded using a temporal convolutional network comprising three residual temporal convolutional blocks. The dilation coefficients of the three residual temporal convolutional blocks can be 1, 2, and 4, respectively. Each convolutional block can include convolution, normalization, nonlinear activation, and random deactivation. Missing data masks and quality metadata can be used as additional channel inputs or for gating intermediate representations after each convolutional block. This specific architecture illustrates an encoding method that can be implemented by those skilled in the art, but does not preclude the use of other temporal encoders capable of generating behavior sequence representations. In this implementation, the input vector for each run simultaneously retains numerical features, the corresponding missing data mask, and the corresponding quality metadata, enabling the case where the same event quantity is zero to be distinguished from the case where no valid samples were obtained at the model input level.

[0099] Component status encoder

[0100] The component status encoder receives the remaining lifespan ratio generated based on the component's cumulative operating time, fault codes, maintenance records, and expected lifespan. For the j-th component, the remaining lifespan ratio r_j = max(0, 1 - H_j / H_j,exp) can be generated based on the cumulative operating time H_j and expected lifespan H_j,exp. When maintenance records exist, the cumulative operating time can be updated based on the maintenance type, maintenance time, or replacement record; fault codes can form component status characteristics through fault code type, frequency of occurrence, duration, and clearing status.

[0101] The component status encoder can generate embedded representations according to component categories and can adapt to different vehicle models with different component sets. For components lacking expected lifespans, the vehicle model baseline lifespan, statistical lifespan of similar components, or a missing mask can be used to indicate that the item is unavailable.

[0102] Fusion layer

[0103] The fusion layer assigns weights to the behavioral sequence representation and the component state representation based on the mask and the quality metadata, reducing the contribution of input channels with high missing rates or low freshness to the fused representation. As an example, the fused representation z can be expressed as z = α_bz_b + α_pz_p, where z_b is the behavioral sequence representation, z_p is the component state representation, and α_b and α_p are normalized weights generated based on the mask and quality metadata.

[0104] For example, the behavior sequence representation z_b, the component state representation z_p, the channel missing rate vector m_q and the quality metadata vector q can be input into a gating network to obtain gating scores g_b and g_p, and α_b and α_p are obtained through softmax. The gating network can be constrained such that when the missing rate of the corresponding input increases or the data freshness decreases, the corresponding weight does not increase. This constraint can be implemented through a monotonic network, a weight sign constraint, a training penalty term, or rule correction after deployment.

[0105] When the missing rate of the inertial data channel is high, the fusion layer reduces the weights corresponding to the rapid acceleration, rapid deceleration and rapid steering features that rely on the inertial measurement unit; when the freshness of maintenance records is low, the fusion layer reduces the weight of the component state representation; when the timestamp deviation of the positioning interface is large, the fusion layer reduces the weights of features that rely on positioning displacement, road type or overspeed judgment.

[0106] Discrete-time hazard output head and event type output head

[0107] The discrete-time hazard output head is configured to generate an overall hazard rate for event occurrence conditions for each future time interval, and the event type output head is configured to generate a type distribution of each vehicle event type in the corresponding future time interval. The two can jointly generate a conditional hazard rate h(k,m) for multiple future time intervals and multiple vehicle event types, where k represents the vehicle event type and m represents the future time interval.

[0108] In one example, the discrete-time hazard output head outputs an overall conditional hazard rate q_m for the time interval m, and the event type output head outputs a type probability π(k,m) under the condition that an event occurs, and h(k,m)=q_m·π(k,m), where 0≤q_m≤1 and Σ_kπ(k,m)=1. In an alternative implementation that does not require the aforementioned two independent output heads, a joint output structure may output logits for K event classes and one "no event in this time interval" class for each time interval, and normalization is performed via softmax such that Σ_kh(k,m)≤1. Said alternative implementation does not alter the explicit disclosure of the implementation employing a discrete-time hazard output head and an event type output head. The above normalization method can prevent the sum of conditional hazard rates of multiple competing event types from exceeding 1.

[0109] In one example, the probability that event type k occurs first in future time interval m is P(T=m,J=k)=h(k,m)·S(m-1), where S(m-1)=∏_{s<m}[1-Σ_kh(k,s)] represents the survival probability that no target vehicle event has occurred before future time interval m. By accumulating the conditional hazard rates for each future time interval, the time distribution and cumulative occurrence probability of each vehicle event type can be obtained.

[0110] Training, temporal partitioning, and uncertainty with right-censored samples

[0111] The following section further explains the training and uncertainty determination of the competition risk time series model. (Refer to...) Figure 4 Training samples with right censoring can be used and the training set, calibration set, and test set can be divided according to time sequence.

[0112] like Figure 4 As shown, the model training process 400 with right-censored samples takes historical time-series feature tensors and event data sources 402 as inputs. The event label or right-censored label generation module 404 generates vehicle event types and event occurrence time intervals for each training sample, or generates right-censoring identifiers and censoring time intervals. The time-sequence partitioning module 406 divides the samples into a training set 407, a calibration set 409, and a test set 411. The joint likelihood training module 408 trains the model using the uncensored event likelihood term and the right-censored survival probability term in the training set 407. The calibration module 410 determines calibration parameters and confidence interval parameters using the calibration set 409. Finally, the testing module 412 evaluates the model using the test set 411. Through this process, a versioned competition risk time-series model 414 is obtained. The versioning information may include model identifiers, feature identifiers, label definitions, and future time interval definitions.

[0113] Training samples and right censoring markers

[0114] For each historical index time, an observation time interval preceding the historical index time and a prediction time interval following the historical index time are constructed. Data within the observation time interval is used to generate a temporal feature tensor; the first target vehicle event occurring within the prediction time interval is used to determine the event type and the event occurrence time interval.

[0115] If the k-th type of vehicle event occurs first within the prediction time interval, then the training sample is an uncensored sample, and the vehicle event type k and the corresponding event occurrence time interval m are recorded. If the target vehicle event does not occur before the observable endpoint, or the vehicle exits the observation and data recording terminates, making it impossible to continue confirming the event, then the training sample includes a right censoring flag and a censoring time interval c.

[0116] Vehicle event tags can be generated from verified event sources. The event timing of a collision event can be preferentially determined from event data loggers, restraint system trigger records, collision sensors, or verified accident records; vehicle component failure events can be determined from the moment a diagnostic fault code first reaches a predetermined severity, a component failure record, or a repair confirmation moment; glass damage events and body damage events can be determined from authorized repair records, image verification results, or vehicle event records. The data sources, priorities, and conflict resolution rules used to generate tags are managed by the tag version identifier.

[0117] When a plurality of vehicle events are recorded within the same future time interval, the vehicle event with the earliest timestamp may be selected as a competing risk target; when a plurality of events have indistinguishable identical timestamps, one target may be selected according to a predetermined priority, marked as a composite event type, or excluded from samples used for single initial event training. The same rule is applied to the training set, calibration set, test set and online inference result interpretation, so as to avoid inconsistent label definitions.

[0118] When constructing training samples, telemetry data, maintenance records, vehicle event records or maintenance confirmation results after the historical index moment are not written into input features, thereby avoiding label leakage. For a plurality of historical index moments of the same vehicle, samples may be extracted according to a minimum interval, or weights may be set for related samples of the same vehicle in the training loss.

[0119] Joint Likelihood Training

[0120] The competing risk time-series model is trained by maximizing the joint likelihood comprising an event likelihood term for uncensored samples and a survival probability term for right-censored samples. For an uncensored sample in which a k-th type vehicle event occurs in a future time interval m, the event likelihood term thereof is P(T=m,J=k); for a right-censored sample whose observation ends at a censoring time interval c, the survival probability term thereof is S(c).

[0121] For an n-th training sample, let δ_n=1 represent uncensored, and δ_n=0 represent right-censored. The likelihood thereof can be expressed as L_n=[h(k_n,m_n)·∏_(s<m_n)(1−Σ_jh(j,s))]^(δ_n)·[∏_(s≤c_n)(1−Σ_jh(j,s))]^(1−δ_n). Model training can minimize −Σ_nlogL_n, and mini-batch gradient descent, Adam or other numerical optimization methods can be used to update model parameters. To avoid log(0), the conditional hazard rate can be limited within a predetermined open interval or a numerical stability constant can be added to the logarithmic operation.

[0122] Equivalently, the joint negative log-likelihood can be minimized: the uncensored sample corresponds to −logP(T=m,J=k), and the right-censored sample corresponds to −logS(c). The training objective may further include an auxiliary loss for improving event type distinction, time sorting or probability calibration, but the auxiliary loss does not replace the event likelihood term and the survival probability term.

[0123] Dividing the training set, calibration set and test set according to chronological order

[0124] Training samples are divided into training, calibration, and test sets according to time sequence. For example, samples from the earliest time period are used as the training set, subsequent time periods as the calibration set, and the most recent time period as the test set. Information from time periods after the same historical index time must not be included in feature construction or model selection for earlier time periods.

[0125] To reduce the leakage of information about the same vehicle across sets, samples of the same vehicle can be ensured to appear in only one set, or isolated time intervals can be set at time division boundaries. The training process can perform early stopping based on independent validation subsets in the training set and save the model parameters that produce the lowest negative log-likelihood or the lowest overall Brier score. The calibration set is used for probabilistic calibration, uncertainty thresholding, or determining conformal prediction parameters after the model parameters and model structure are determined, thereby avoiding the simultaneous use of the same calibration data for model selection and final calibration.

[0126] The calibration set is used to determine probabilistic calibration parameters, uncertainty thresholds, quality thresholds, or conformal prediction quantiles; the test set is used to evaluate time-dependent AUC, composite Brier score, negative log-likelihood, expected calibration error, prediction interval coverage, and recall rate for each vehicle event type. Separating the test set by time allows for simulation of scenarios involving newer vehicles and road conditions after model deployment.

[0127] Uncertainty of conditional hazard rate

[0128] The uncertainty is determined by at least one of the prediction variance among the model set, the variance of random inactivation repeated inference, or the width of the conformal prediction interval.

[0129] When using model ensembles, multiple competing risk time series models can be trained using different random initializations, different training subsets, or different model structures, and the variance of the predictions of multiple models for the same conditional hazard rate can be calculated. When using random deactivation and repeated inference, random deactivation is kept enabled during the inference phase, multiple inferences are performed on the same time series feature tensor, and the uncertainty is determined based on the variance of the multiple conditional hazard rates. When using conformal prediction, inconsistency scores can be obtained on the calibration set, and conformal prediction intervals for event time distributions or cumulative occurrence probabilities can be generated based on the target coverage, with the interval width representing the uncertainty.

[0130] In a conformal prediction implementation, a non-consistency score s_n = 1 - F_hat(k_n, m_n) is calculated for the calibration samples, where F_hat(k_n, m_n) represents the model's cumulative probability of occurrence for the actual event type and the time interval no later than the actual event. For right-censored calibration samples, a censoring adjustment score compatible with their observable interval can be used. A threshold is determined based on the (1-α) empirical quantile of the calibration score, and a prediction set or cumulative probability interval covering the target event type / time interval is formed accordingly. The server can use the size of the prediction set or the width of the interval as the event type uncertainty.

[0131] Uncertainty can be calculated separately for different future time intervals and different vehicle event types. The maximum uncertainty, weighted average uncertainty, or uncertainty of key vehicle event types can also be used as the basis for sampling configuration selection.

[0132] Feedback sampling based on quality metadata and uncertainty

[0133] The following section provides further explanation of the selection and restoration of the second sampling configuration. (Refer to...) Figure 5 The server can perform configuration selection, configuration distribution, and recovery control based on quality metadata and uncertainty.

[0134] like Figure 5As shown, the sampling control process 500 for quality and uncertainty gating inputs quality metadata 502, prediction uncertainty 504, and current sampling configuration and vehicle-side resource constraints 506 into the sampling configuration selection function 508. The sampling configuration selection function 508 first determines the data quality gap of each sensor data channel based on the quality metadata 502, and then determines the uncertainty contribution of each sensor data channel to the target vehicle event type and the target future time interval based on the prediction uncertainty 504. Finally, it combines the data quality gap and uncertainty contribution into a channel-level information gap. The sampling configuration selection function 508 selects either a first sampling configuration or a basic sampling configuration 510, or a second sampling configuration or an enhanced sampling configuration 512, based on the channel-level information gap, threshold, hysteresis condition, maximum duration, and resource boundaries. Both the first sampling configuration or basic sampling configuration 510 and the second sampling configuration or enhanced sampling configuration 512 may include control parameters executed by the vehicle-side telemetry equipment. These control parameters include at least one of the following: inertial measurement unit sampling rate control parameter 514, time synchronization message interval control parameter 516, raw event window retention and upload parameter 518, and circular buffer and upload queue control parameter 520. The first sampling configuration or basic sampling configuration 510 may specify baseline values, default values, or restored values ​​for the aforementioned control parameters. The second sampling configuration or enhanced sampling configuration 512, relative to the first sampling configuration or basic sampling configuration 510, enhances at least one of the aforementioned control parameters to supplement vehicle-side telemetry data related to data quality gaps or prediction uncertainties. For example, the enhanced sampling configuration 512 may modify the inertial measurement unit sampling rate control parameter 514, the time synchronization message interval control parameter 516, the raw event window retention and upload parameter 518, and the circular buffer and upload queue control parameter 520. The basic sampling configuration 510 may restore the corresponding parameters to their baseline values. The adjusted parameters are used to collect subsequent telemetry data and to feed back the newly generated quality metadata 502 and prediction uncertainty 504 to the next configuration selection.

[0135] The server first compares the quality metadata with the corresponding quality thresholds, and then compares the uncertainty with the corresponding uncertainty thresholds. If all quality metadata meets the quality thresholds and the uncertainty does not exceed the uncertainty threshold, the first sampling configuration is selected; if at least one key quality metadata does not meet the quality threshold or at least one key uncertainty exceeds the uncertainty threshold, the second sampling configuration is selected.

[0136] Sampling configuration selection can be implemented using a deterministic configuration selection function. This function receives quality metadata organized by sensor data channels, uncertainties organized by vehicle event type and future time intervals, the current sampling configuration, and resource constraints, and outputs the next sampling configuration. To ensure that configuration changes correspond to the information to be supplemented, a mapping of "trigger condition—controlled sensor data channel—configuration parameters" can be maintained. For example, the inertial data channel missing rate can be mapped to the inertial measurement unit sampling rate, the positioning data channel timestamp deviation can be mapped to the time synchronization message interval, and the event type uncertainty can be mapped to the original event window retention and uploading.

[0137] Selecting the second sampling configuration includes at least one of the following operations: increasing the sampling rate of the inertial measurement unit in response to the inertial data channel missing rate exceeding a first threshold; shortening the time synchronization message interval in response to the positioning data channel time deviation exceeding a second threshold; or retaining and uploading the original data window around the candidate event in response to the event type uncertainty exceeding a third threshold.

[0138] As a first example, the initial threshold is 10%. When the inertial data channel missing rate reaches 18%, the server increases the inertial measurement unit sampling rate from 20 Hz to 50 Hz and raises the upload priority of the inertial data channel from normal priority to high priority. Increasing the sampling rate increases the number of effective samples available for identifying rapid acceleration, deceleration, and sharp turns in subsequent journeys.

[0139] As a second example, the second threshold is 1 second. When the 95th percentile of the positioning data channel time deviation is 1.8 seconds, the server will shorten the time synchronization message interval from 60 seconds to 10 seconds, or require the vehicle-side telemetry equipment to add time synchronization messages at the beginning and end of each trip to improve the alignment between positioning data and vehicle bus data and inertial data.

[0140] As a third example, the third threshold is 0.08. When the event type uncertainty between a collision event and a vehicle damage event is 0.13, the server causes the vehicle-side telemetry device to retain and upload the raw data window of 30 seconds before and 60 seconds after the candidate deceleration or impact event, instead of just uploading the aggregated exposure normalized features.

[0141] The second sampling configuration allows setting the maximum duration, maximum number of trips, maximum daily data volume, and recovery conditions to avoid prolonged consumption of vehicle computing resources, storage space, or communication bandwidth by enhanced sampling. When a predetermined number of consecutive trips meet the quality threshold and the uncertainty does not exceed the uncertainty threshold, the server selects the first sampling configuration, enabling the vehicle-side telemetry equipment to restore the basic sampling rate, basic event window, and basic upload priority.

[0142] To avoid frequent switching near the threshold, the trigger threshold for entering the second sampling configuration can be higher than the recovery threshold for restoring the first sampling configuration, or the trigger condition can be required to be met continuously for a first predetermined number of times, and the recovery condition can be required to be met continuously for a second predetermined number of times. The vehicle-side telemetry equipment can also be set with a minimum configuration retention time, and automatically restore the first sampling configuration after the second sampling configuration reaches the maximum duration or maximum data volume, unless the server sends a new configuration that has been verified by version.

[0143] After the second sampling configuration is applied, the vehicle-side telemetry equipment actually alters at least one vehicle sensor data stream or its data processing path, such as changing the sensor sampling frequency, synchronization message frequency, ring buffer retention time, or wireless transmission queue priority. Therefore, the risk inference results are not merely used as reporting information, but are used to control the acquisition and transmission of subsequent vehicle telemetry data.

[0144] The server can enhance sampling only for channels that cause insufficient quality or increased uncertainty, while keeping other channels unchanged. For example, when there is a large time skew in the positioning data channel, the positioning synchronization message interval can be shortened without increasing the inertial measurement unit sampling rate. This ensures that the adjustment of the sampling configuration corresponds to the information to be supplemented.

[0145] Risk data objects and immutable audit logs

[0146] The following section provides further explanation of the structured record of the risk assessment results.

[0147] The server generates a risk data object. The risk data object includes a model version identifier, a feature version identifier, an observation time interval, the time distribution of each vehicle event type, a confidence interval, a quality score, and the selected sampling configuration identifier.

[0148] Risk data objects can be stored as binary objects, relational database records, key-value records, or serialized messages with predefined field definitions. In addition to the listed fields, risk data objects may include inference timestamps, future time interval definitions, vehicle event type definition versions, quality gating status, and uncertainty calculation method identifiers, enabling the receiving system to interpret historical risk outputs without relying on the current model service state.

[0149] The model version identifier identifies the competing risk time series model and its parameter versions that generated the current risk assessment results; the feature version identifier identifies the feature processing rules used for time synchronization, trip segmentation, driving event detection, exposure normalization, and context hierarchical layering; the observation time interval identifies the range of historical telemetry data used in this risk assessment; the time distribution of each vehicle event type may include the conditional hazard rate, cumulative probability of occurrence, or survival probability for each future time interval; the confidence interval may be generated by model set, random inactivation repeated inference, or conformal prediction; the quality score reflects quality metadata; and the sampling configuration identifier identifies the first or second sampling configuration selected and issued by the server.

[0150] The server writes the risk data object to an immutable audit log. The immutable audit log can be implemented using an append-only log store that prohibits overwriting, an object store with a write-after-retention policy, a hash chain-based log, or a database with access control and version retention mechanisms.

[0151] In a hash chain-based implementation, each audit log entry may include a digest value of the current risk data object, a digest value of the previous log entry, the write time, and the write subject identifier. Modification to any historical risk data object will result in an inconsistency in the digest chain, which can then be identified by the audit program.

[0152] For example, a first summary value can be calculated for the risk data object after it has been normalized according to a fixed field order. Then, the first summary value, the summary value of the previous log record, and the monotonically increasing sequence number can be used to calculate the summary value of the current log record. Write operations can use the inference request identifier as an idempotent key to avoid semantically duplicated log records caused by network retries. Access permissions for immutable audit logs can be separated from the access permissions of the model inference service and downstream application services.

[0153] Risk data objects may also include desensitized values ​​of vehicle-side telemetry device identifiers, inference request identifiers, quality gating results, uncertainty threshold versions, and sampling configuration application confirmation information. However, these additional fields do not affect the fact that the risk data object must at least include model version identifiers, feature version identifiers, observation time intervals, time distributions of each vehicle event type, confidence intervals, quality scores, and the selected sampling configuration identifier. (Drift monitoring, automatic policy freezing, and safety degradation are also included.)

[0154] The server can compare the online feature distribution with the training baseline distribution to generate a drift metric. When the drift metric exceeds a drift threshold, it freezes the automatic policy based on the risk output, switches to a predetermined rule model, and triggers model retraining or recalibration. This drift monitoring and security degradation process can serve as a reliability gating mechanism before the risk output enters the downstream automatic policy.

[0155] The online feature distribution can be generated based on exposure-normalized features, missing masks, quality metadata, vehicle type proportions, or sensor status from the most recently predetermined number of vehicles, trips, or observation time intervals; the training baseline distribution can be generated based on the same features in the training set. Drift metrics can be at least one of the following: overall stability index, Kullback-Leibler divergence, Jensen-Shannon divergence, Wasserstein distance, maximum mean difference, or classifier discriminancy.

[0156] For example, for a feature divided into B intervals, the overall stability index PSI can be calculated based on the training baseline proportion p_b and the online proportion q_b: PSI = Σ_b(q_b-p_b)·ln(q_b / p_b). When p_b or q_b is zero, a predetermined smoothing constant can be used. The drift threshold can be determined on a calibration set or historical stable operating intervals, and a safety degradation can be triggered only after the drift metric exceeds the threshold for two or more consecutive monitoring windows.

[0157] Drift metrics can be calculated for individual features, feature groups, or the entire time-series feature tensor. For example, the distribution of nighttime deceleration rate, the distribution of missing location data channels, and the distribution of remaining lifetime proportion can be monitored separately. For multiple drift metrics, the overall drift determination can be formed by using the maximum value, weighted sum, or the number of features that meet the threshold.

[0158] When the drift metric exceeds the drift threshold, the server freezes the automatic policy based on risk output, switches to a predetermined rule model, and triggers model retraining or recalibration. Freezing the automatic policy based on risk output means preventing the risk output of the current competing risk time series model from directly triggering automatic control, automatic maintenance of resource allocation, or other automatic policy processing, but the risk output can be retained for manual review, offline analysis, or model diagnosis.

[0159] Predefined rule models can generate alternative results based on fixed vehicle safety thresholds, manufacturer maintenance rules, fault code severity, or conservative driving behavior rules. For example, when a critical fault code in the braking system is active, the predefined rule model directly generates a maintenance prompt without relying on the conditional hazard rate of a competing risk time series model.

[0160] Model retraining can utilize a training set supplemented with samples from new time periods, and continues to use training samples with right censoring and chronologically partitioned training, calibration, and test sets. Recalibration keeps model parameters unchanged, updating only temperature parameters, conformal prediction quantiles, or risk thresholds using the newer calibration set. Candidate models can undergo offline testing, shadow inference, calibration checks, and version approval before re-enabling automatic strategies.

[0161] Application of vehicle incident types and risk outcomes

[0162] The following section provides further explanation of the types of vehicle incidents and their application to risk outcomes.

[0163] The various vehicle event types include collision events, vehicle component failure events, glass damage events, and body damage events. Collision events can be determined based on impact acceleration, sudden velocity changes, restraint system status, or collision records; vehicle component failure events can be determined based on fault codes, remaining life percentage, abnormal temperatures, or abnormal operating conditions; glass damage events and body damage events can be determined based on maintenance records, event logs, or authorized vehicle status information.

[0164] The server can determine the priority of safety alerts, maintenance appointments, or manual review based on the risk data objects. Safety alerts may include suggestions to reduce speed, check driving conditions, or perform vehicle inspections; maintenance appointment priorities can be determined based on the conditional hazard rate and confidence interval of vehicle component failure events in different future time intervals; manual review priorities can be determined based on a combination of conditional hazard rate, uncertainty, and quality score.

[0165] Specific examples of anticipatory closed-loop risk assessment

[0166] The following is a non-limiting, anticipatory example illustrating how the foregoing embodiments can be combined and implemented. Unless otherwise expressly stated, the numerical values ​​and outputs in this embodiment are illustrative and not statements of completed vehicle testing or model training results.

[0167] Telemetry stream reception and synchronization

[0168] The vehicle-side telemetry equipment receives vehicle speed, brake pedal opening, gear position, and ignition status at 10 Hz from the vehicle bus interface; longitudinal acceleration, lateral acceleration, and yaw rate at 100 Hz from the inertial measurement unit; and position, positioning speed, and positioning status at 1 Hz from the positioning interface. Each sample includes a source timestamp and sensor status identifier.

[0169] The vehicle-side telemetry equipment uses a monotonic clock as a reference, estimating the vehicle bus clock offset as 35 ms with a drift rate of 1.000002, the inertial measurement unit clock offset as -18 ms with a drift rate of 0.999998, and the positioning interface clock offset as 420 ms with a drift rate of 1.000010. After calibration, the numerical channels are resampled to a common time grid of 100 ms, and hold processing is performed on ignition status and gear position. The positioning interface exceeds the maximum interpolation interval of 3 seconds within a 6-second interruption interval; therefore, positioning samples in this interval are set with a corresponding missing mask.

[0170] Trip segmentation, event window, and exposure normalization

[0171] This vehicle is a gasoline-powered vehicle, and its driving status is determined by ignition status and engine running status. Vehicle motion is determined by three factors: vehicle speed, gear position, and positioning displacement. Once the vehicle's ignition status changes to non-ready, if the vehicle speed is less than 0.5 km / h and the positioning displacement is less than 3 m for 60 seconds, the vehicle-side telemetry equipment terminates the current trip. The effective mileage of this trip is 42.6 km, and the effective driving time is 1.15 hours.

[0172] During this trip, 3 rapid acceleration events, 5 rapid deceleration events, 2 sharp turning events, 1 speeding event, and 4 over-the-horse following events were detected. A 20-second pre-event window and a 30-second post-event window were established for each event. After normalization based on effective mileage, the rapid deceleration rate was 5 / 42.6 = 0.117 times / km, and the sharp turning rate was 2 / 42.6 = 0.047 times / km.

[0173] The trip included 30.0 km of ordinary urban roads and 12.6 km of highways, with 18.0 km traveling at night. There were 4 sudden decelerations on ordinary urban roads and 1 sudden deceleration on the highway, resulting in a sudden deceleration rate of 0.133 times / km on ordinary urban roads and 0.079 times / km on the highway. There were 3 sudden decelerations at night, resulting in a sudden deceleration rate of 0.167 times / km at night. These contextual hierarchical features were written into different feature channels.

[0174] Quality metadata and temporal feature tensors

[0175] Based on this trip and the aforementioned observation time interval, the inertial data channel integrity is 0.94, the positioning data channel integrity is 0.88, the 95th percentile of the positioning data channel timestamp deviation is 1.4 seconds, the cross-sensor consistency is 0.90, the data freshness is 2 hours, and the effective driving distance is 186 km. The vehicle-side telemetry equipment combines the exposure normalization features, the missing mask, and the aforementioned quality metadata into a temporal feature tensor and sends it to the server.

[0176] Model inference and uncertainty

[0177] The server's behavior sequence encoder encodes the temporal feature tensors of the most recent 30 trips. The component status encoder receives the remaining life ratios of the braking system, tire system, and power system. The braking system has accumulated 1600 hours of operation time, an expected lifespan of 2400 hours, and a remaining life ratio of 0.333. The tire maintenance record shows that the accumulated operation time since the most recent replacement is 300 hours, so the tire remaining life ratio is generated based on the updated accumulated operation time.

[0178] In this prospective example, the behavior sequence encoder employs three residual temporal convolutional blocks, the component state encoder uses a two-layer fully connected network, and the fusion layer generates gating weights based on the integrity of the positioning data channel, the integrity of the inertial data channel, and the freshness of the maintenance records. The output generates an overall conditional hazard rate and four event type probabilities for each future time interval, and obtains the conditional hazard rate for each vehicle event type using h(k,m)=q_m·π(k,m).

[0179] The fusion layer reduces the weight of location-related features based on missing masks and quality metadata. Discrete-time hazard rate output headers and event type output headers generate conditional hazard rates for collision events, vehicle component failure events, glass damage events, and body damage events in three future time intervals: 0 to 7 days, 8 to 30 days, and 31 to 90 days.

[0180] A model ensemble of five models was used. For vehicle component failure events occurring within 8 to 30 days, the mean conditional hazard rate output by the five models was 0.086, with an uncertainty of 0.012 corresponding to the prediction variance. For collision events occurring within 8 to 30 days, the mean conditional hazard rate was 0.041, with an uncertainty of 0.006. For glass damage events and vehicle body damage events, the mean conditional hazard rates were 0.018 and 0.027, respectively. The server further generated the event time distribution and confidence interval for each vehicle event type based on the conditional hazard rates.

[0181] The server can also calibrate the cumulative occurrence probability using a calibration set divided chronologically, and generate confidence intervals or prediction intervals based on the empirical quantiles predicted by the model set on the calibration set. The interval width, together with the model set variance, forms the event type uncertainty, which is compared with a predetermined third threshold. Sampling configuration selection and distribution.

[0182] The server sets a second threshold for the time deviation of the positioning data channel to 1 second. Since the 95th percentile of the timestamp deviation of the current positioning data channel is 1.4 seconds, the server selects the second sampling configuration based on the quality metadata. The second sampling configuration shortens the positioning time synchronization message interval from 60 seconds to 10 seconds and increases the priority of the positioning data channel upload to high priority.

[0183] Since the integrity of the inertial data channel did not fall below the first threshold, the sampling rate of the inertial measurement unit remained at 100 samples per second.

[0184] The server sends the selected second sampling configuration to the vehicle-side telemetry device. After verifying the configuration identifier, version number, and effective time, the vehicle-side telemetry device applies the second sampling configuration and collects or uploads subsequent positioning telemetry data according to the shortened time synchronization message interval. After the time deviation of the positioning data channel for three consecutive trips is less than 1 second, the server selects the first sampling configuration, restoring the time synchronization message interval to 60 seconds.

[0185] In the first sampling configuration, the vehicle-side telemetry device can upload only the exposure normalization feature, missing mask, and quality metadata. In the second sampling configuration of this prospective example, the vehicle-side telemetry device can increase the upload priority of only the positioning time synchronization information and positioning data channel data, without continuously uploading the entire raw telemetry stream. If the event type uncertainty exceeds a corresponding threshold, the second sampling configuration can further increase the upload priority of the raw data window related to the candidate event. Thus, the vehicle-side telemetry device can supplement information for positioning time deviations while avoiding irrelevant channels from occupying wireless communication bandwidth for extended periods.

[0186] Risk data objects, auditing and application

[0187] The server generates a risk data object, which includes a model version identifier "CRTM-3.2", a feature version identifier "FT-2.5", the observation time interval of the most recent 90 days, the time distribution of four types of vehicle events, the confidence interval, a quality score of 0.82, and a second sampling configuration identifier "SC-ENH-17", and writes the risk data object to an immutable audit log.

[0188] Based on the risk data object, the server determines that the braking system maintenance appointment has a high priority and generates a vehicle safety alert. The risk data object can also be provided to a manual review interface to display the vehicle event type, future time interval, conditional hazard rate, confidence interval, quality score, and sampling configuration identifier.

[0189] Those skilled in the art, upon reading this specification, can make various modifications, substitutions, and combinations to the disclosed embodiments without departing from the spirit and scope of the appended claims. Therefore, the scope of protection of this invention should be determined by the claims and their equivalents.

Claims

1. A method for assessing vehicle event risk performed by a vehicle-side telemetry device and a server communicating with the vehicle-side telemetry device, the method comprising: The vehicle-side telemetry equipment receives multiple telemetry streams with different sampling rates from the vehicle bus interface, inertial measurement unit interface, and positioning interface. Each telemetry stream includes a timestamp and a sensor status identifier. The vehicle-side telemetry device estimates the clock deviation between the multiple telemetry streams based on the timestamp, and performs synchronization processing on the multiple telemetry streams on a common time axis to generate synchronized telemetry data; The vehicle-side telemetry device extracts one or more telemetry data items from the synchronous telemetry data to determine the vehicle driving state and vehicle motion state, and generates a vehicle driving state and vehicle motion state determination result based on the extracted telemetry data items. The vehicle-side telemetry device divides the synchronous telemetry data into multiple trips based on the vehicle driving state and vehicle motion state determination results, and extracts driving event windows for each trip. The vehicle-side telemetry device divides the number of events in each driving event window by the effective mileage or effective driving time of the corresponding trip to generate exposure normalization features. The exposure normalization features, the mask indicating missing samples, and the quality metadata indicating time synchronization quality are combined into a time-series feature tensor. The vehicle-side telemetry device sends the time-series feature tensor to the server; The server inputs the temporal feature tensor into a competitive risk time series model trained with training samples marked with right censoring, to generate conditional hazard rates and uncertainties for multiple future time intervals and multiple vehicle event types. The server selects a first sampling configuration or a second sampling configuration based on the quality metadata and the uncertainty, wherein the second sampling configuration enables at least one sensor data channel to have a higher sampling rate, a longer event window, or a higher upload priority than the first sampling configuration. as well as The server sends the selected sampling configuration to the vehicle-side telemetry device, enabling the vehicle-side telemetry device to collect or upload subsequent telemetry data according to the selected sampling configuration.

2. The method according to claim 1, wherein, The synchronization process includes: estimating the offset and drift rate of each sensor clock based on a monotonic clock; resampling the numerical channels to a common time grid; performing hold processing on the discrete state channels; and setting the mask for samples that exceed the maximum interpolation interval.

3. The method according to claim 1, wherein, The exposure normalization feature is further layered according to road type, weather condition, day / night condition, or traffic congestion condition, so that the same event corresponds to different feature channels in different contexts.

4. The method according to claim 1, wherein, The quality metadata includes at least three of the following: completeness, timestamp deviation, cross-sensor consistency, data freshness, and effective exposure. The method prohibits the output of a deterministic risk level for automatic control when the effective exposure is less than a threshold.

5. The method according to claim 1, wherein, The competitive risk time series model includes a behavior sequence encoder, a component status encoder, a fusion layer, a discrete-time hazard rate output header, and an event type output header; the component status encoder receives the remaining lifespan ratio generated based on the component's cumulative working time, fault codes, maintenance records, and expected lifespan.

6. The method according to claim 5, wherein, The fusion layer assigns weights to the behavioral sequence representation and the component state representation based on the mask and the quality metadata, thereby reducing the contribution of input channels with high missing rates or low freshness to the fused representation.

7. The method according to claim 1, wherein, Selecting the second sampling configuration includes: increasing the sampling rate of the inertial measurement unit in response to the inertial data channel missing rate exceeding a first threshold; shortening the time synchronization message interval in response to the positioning data channel time deviation exceeding a second threshold; and / or retaining and uploading the original data window around the candidate event in response to the event type uncertainty exceeding a third threshold.

8. The method according to claim 1, wherein, The server determines the data quality gap of each sensor data channel based on the quality metadata, determines the uncertainty contribution of each sensor data channel to the conditional hazard rate based on the uncertainty, combines the data quality gap and the uncertainty contribution into a channel-level information gap, and changes the sampling rate, time synchronization message interval, event window length, cache retention period or upload priority for sensor data channels with large channel-level information gaps.

9. A vehicle incident risk assessment system, comprising: The vehicle-side telemetry equipment includes a vehicle bus interface, an inertial measurement unit interface, a positioning interface, a wireless communication interface, a first processor, and a first memory. The server communicates with the vehicle-side telemetry equipment, the server including a second processor and a second memory; wherein the first memory stores instructions for the first processor to perform the following operations: receiving multiple telemetry streams with different sampling rates and each including a timestamp; estimating the clock offset between the multiple telemetry streams; synchronizing the multiple telemetry streams on a common time axis to generate synchronized telemetry data; extracting one or more telemetry data items from the synchronized telemetry data for determining the vehicle's driving state and vehicle motion state; generating vehicle driving state and vehicle motion state determination results; dividing the synchronized telemetry data into trips; extracting driving event windows for each trip; and generating a time series containing exposure normalization features, missing masks, and time synchronization quality metadata. The feature tensor is transmitted to the server via the wireless communication interface, and subsequent telemetry data is collected or uploaded according to the sampling configuration received from the server; and the second memory stores instructions for the second processor to perform the following operations: receive the time-series feature tensor from the vehicle-side telemetry device, input the time-series feature tensor into a competitive risk time-series model trained with training samples with right censoring to obtain conditional hazard rates and uncertainties for multiple future time intervals and multiple vehicle event types, select a sampling configuration with different sampling rates, event window lengths or upload priorities based on the time synchronization quality metadata and the uncertainty, and send the selected sampling configuration to the vehicle-side telemetry device.

10. The system according to claim 9, wherein, The vehicle-side telemetry device includes a circular cache for storing raw event windows in chronological order when the network is unavailable, and the sampling configuration further specifies the retention period and upload order of the circular cache.

Images