Vehicle risk prediction method and vehicle

Abstract

This application provides a vehicle risk prediction method and vehicle, relating to the field of vehicle driving safety technology. The method includes: collecting vehicle driving data; determining the existence of high-risk events based on the vehicle driving data and generating event signals; predicting multiple frames of future vehicle posture data using a pre-built vehicle posture prediction model based on the vehicle driving data and event signals; obtaining a risk level using a pre-built risk level classification model based on the vehicle driving data, event signals, and multiple frames of future vehicle posture data; and determining target risk decision data based on the event signals, multiple frames of future vehicle posture data, and risk levels. The fusion of event signals, multiple frames of future vehicle posture data, and risk levels provides a comprehensive and reliable basis for risk decision-making, which is beneficial for improving the stability and consistency of target risk decision data in complex scenarios, thereby improving the accuracy and timeliness of vehicle risk prediction.

Description

Technical Field

[0001] This application relates to the field of vehicle driving safety technology, and in particular to a vehicle risk prediction method and a vehicle. Background Technology

[0002] In environments with extreme slopes or irregular terrain, vehicles are prone to rollover risks due to factors such as weight shifts and drastic changes in vehicle posture. However, existing technologies mostly employ fixed threshold judgment strategies or single models for rollover risk identification, which cannot reflect the nonlinear evolution of vehicle posture in complex terrain in real time. This leads to delays or misjudgments in rollover identification, affecting vehicle driving safety. Summary of the Invention

[0003] In view of this, the purpose of this application is to propose a vehicle risk prediction method and a vehicle to solve the problem of delay or misjudgment caused by using a fixed threshold judgment strategy or a single model for rollover risk identification.

[0004] To achieve the above objectives, the first aspect of this application provides a vehicle risk prediction method, comprising:

[0005] Collect vehicle driving data;

[0006] Based on the vehicle driving data, determine whether a high-risk event exists and generate an event signal;

[0007] Based on the vehicle driving data and the event signal, multiple frames of future vehicle posture data are predicted by a pre-built vehicle posture prediction model.

[0008] Based on the vehicle driving data, the event signal, and the multi-frame vehicle future attitude data, the risk level is obtained through a pre-built risk level classification model.

[0009] Based on the event signal, the multi-frame vehicle future attitude data, and the risk level, target risk decision data is determined.

[0010] Optionally, determining whether a high-risk event exists based on the vehicle driving data and generating an event signal includes:

[0011] Based on the vehicle driving data, the values ​​corresponding to each preset indicator are calculated.

[0012] Based on the values ​​corresponding to each preset indicator and the preset event trigger threshold, it is determined whether a high-risk event exists and an event signal is generated.

[0013] Optionally, the step of predicting multiple frames of future vehicle attitude data based on the vehicle driving data and the event signal using a pre-built vehicle attitude prediction model includes:

[0014] In response to the fact that the event signal does not include a high-risk trigger signal, based on the vehicle driving data, the vehicle body posture prediction model uses a first preset inference mechanism to predict multiple frames of future vehicle posture data.

[0015] In response to the event signal, including the high-risk trigger signal, based on the vehicle driving data and the high-risk trigger signal, the vehicle attitude prediction model uses a second preset inference mechanism to predict multiple frames of future vehicle attitude data.

[0016] Optionally, the method for constructing the vehicle body attitude prediction model includes:

[0017] Collect historical vehicle time-series data;

[0018] Based on the historical vehicle time series data, pre-constructed high-risk event samples, and predefined physical prior knowledge, an initial dynamic Bayesian network is constructed.

[0019] Based on the historical vehicle time series data, the initial dynamic Bayesian network is subjected to parameter learning and structural optimization to obtain the vehicle posture prediction model.

[0020] Optionally, the method for constructing the high-risk event samples includes:

[0021] Based on the historical vehicle time series data, calculate the historical values ​​corresponding to each preset indicator;

[0022] Based on the historical data and preset event trigger thresholds, determine whether a high-risk event exists;

[0023] In response to the existence of a high-risk event, data related to a preset time period is extracted from the historical change data based on the trigger time of the high-risk event to form event window data corresponding to the high-risk time.

[0024] The event window data is labeled to form event tags;

[0025] Calculate the sparsity score based on the event window data;

[0026] The event window data, the event labels, and the sparsity scores are used as samples of high-risk events.

[0027] Optionally, the step of obtaining the risk level based on the vehicle driving data, the event signal, and the multi-frame vehicle future attitude data through a pre-built risk level classification model includes:

[0028] The current operating condition of the vehicle is determined based on the vehicle driving data.

[0029] The target risk level classification model is selected from multiple pre-built risk level classification models based on the current operating condition of the vehicle.

[0030] Based on the vehicle driving data, the event signal, and the multi-frame vehicle future attitude data, the risk level is obtained through the target risk level classification model.

[0031] Optionally, determining the target risk decision data based on the event signal, the multi-frame vehicle future attitude data, and the risk level includes:

[0032] Determine whether the event signal, the multi-frame vehicle future attitude data, and the risk level meet the preset high-risk conditions;

[0033] In response to the fulfillment of preset high-risk conditions, preset risk decision data corresponding to the preset high-risk conditions is used as the target risk decision data.

[0034] Optional, also includes:

[0035] In response to the failure to meet the preset high-risk conditions, the weight coefficients corresponding to the event signal, the multi-frame vehicle future attitude data, and the risk level are calculated respectively.

[0036] The target risk decision data is determined based on the weighting coefficients, the event signals, the multi-frame vehicle future attitude data, and the risk level.

[0037] Optional, also includes:

[0038] The target risk decision data is sent to the relevant controllers of the vehicle so that the relevant controllers can execute the control instructions carried in the target risk decision data.

[0039] Based on the same inventive concept, a second aspect of this application also provides a vehicle risk prediction device, comprising:

[0040] The data acquisition module is configured to collect vehicle driving data;

[0041] The determination module is configured to determine whether a high-risk event exists based on the vehicle driving data, and to generate an event signal;

[0042] The prediction module is configured to predict multiple frames of future vehicle posture data based on the vehicle driving data and the event signal using a pre-built vehicle posture prediction model.

[0043] The classification module is configured to obtain the risk level based on the vehicle driving data, the event signal, and the multi-frame vehicle future attitude data through a pre-built risk level classification model.

[0044] The decision-making module is configured to determine target risk decision data based on the event signal, the multi-frame vehicle future attitude data, and the risk level.

[0045] Based on the same inventive concept, a third aspect of this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.

[0046] Based on the same inventive concept, a fourth aspect of this application also provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the method described above.

[0047] Based on the same inventive concept, a fifth aspect of this application also provides a vehicle, said vehicle comprising:

[0048] Memory, used to store executable program code;

[0049] A processor is configured to call and run the executable program code from the memory, causing the vehicle to perform the method described above.

[0050] As described above, the vehicle risk prediction method and vehicle provided in this application include: collecting vehicle driving data, determining whether a high-risk event exists based on the vehicle driving data, and generating an event signal. High-risk events are sudden events; identifying sudden events can improve the timeliness of subsequent risk prediction. Based on the vehicle driving data and the event signal, a pre-built vehicle posture prediction model predicts multiple frames of future vehicle posture data. These multiple frames of future vehicle posture data can predict the future driving state of the vehicle, which is beneficial for early identification of vehicle risks. If a high-risk event exists, the event signal input to the vehicle posture prediction model can drive the model to predict multiple frames of future vehicle posture data with higher risk, improving prediction accuracy. Based on the vehicle driving data, the event signal, and the multiple frames of future vehicle posture data, a risk level is obtained through a pre-built risk level classification model. The risk level can provide a fast and clear basis for subsequent risk decisions. Based on the event signal, the multiple frames of future vehicle posture data, and the risk level, target risk decision data is determined. By fusing event signals, multi-frame vehicle future attitude data, and risk levels, a comprehensive and reliable basis can be provided for risk decision-making. This is beneficial for improving the stability and consistency of target risk decision-making data in complex scenarios, thereby enhancing the accuracy and timeliness of vehicle risk prediction. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart illustrating the vehicle risk prediction method according to an embodiment of this application;

[0053] Figure 2 This is a schematic diagram of the vehicle risk prediction device according to an embodiment of this application;

[0054] Figure 3 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0056] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0057] As described in the background section, in environments with extreme slopes or irregular terrain, vehicles are prone to rollover risks due to factors such as weight shifts and drastic changes in vehicle posture. Existing technologies mostly employ fixed threshold judgment strategies or single models for risk identification, which cannot reflect the nonlinear evolution of vehicle posture in complex terrain in real time, leading to delays or misjudgments in rollover identification.

[0058] Specifically, fixed threshold strategies (such as "warning when roll angle > 30°") are static judgment strategies that cannot adapt to the dynamic changes in a vehicle. For example, if a vehicle travels at high speed over a raised surface with all four wheels briefly off the ground, the roll angle may be zero (not triggering an alarm), but the vehicle is already in extreme danger. The fixed threshold becomes completely ineffective, resulting in a judgment delay problem. Conversely, in slow, extreme off-road driving, the roll angle may exceed the threshold for an extended period while the vehicle remains controllable, leading to frequent false alarms.

[0059] Vehicle attitude changes are nonlinear, involving inertia, oscillations, and coupling effects (such as the coupling between steering and roll). A single model (such as a simple classifier or regression model) lacks sufficient expressive power and temporal context to capture this complex dynamic behavior, and can only make instantaneous and one-sided judgments.

[0060] Furthermore, issues such as data asynchrony and information decoupling exist among multi-source sensors. When sensor data from different time stamps is input into the model, it can lead to cognitive dissonance, affecting the accuracy of risk assessment. Simultaneously, the models generally only focus on current data, failing to integrate historical data, unexpected events, and future trends. This lack of ability to predict risks in advance results in delayed risk assessments, severely impacting vehicle driving safety.

[0061] In view of this, this application proposes a vehicle risk prediction method that can integrate sudden events, future vehicle trends and risk levels, providing a comprehensive and reliable basis for risk decision-making. This method is conducive to improving the stability and consistency of target risk decision-making data in complex scenarios, thereby improving the accuracy and timeliness of vehicle risk prediction.

[0062] The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0063] The embodiments of this application provide a vehicle risk prediction method, with reference to Figure 1 This includes the following steps:

[0064] Step 102: Collect vehicle driving data.

[0065] Specifically, vehicle driving data includes data collected from multiple sensors, which can be acquired through an onboard multi-source sensor acquisition platform. The construction of this platform relies on the high integration of signals from multiple vehicle sensors, which collectively form the basis for describing the vehicle's state in complex slope environments. These multiple sensor signals can cover various complex slope and attitude change scenarios. Throughout the acquisition process, all sensors must meet automotive-grade accuracy and sampling stability requirements, ensuring data continuity and integrity throughout vehicle operation to provide robust data support for subsequent processing. These multiple vehicle sensors include an onboard inertial measurement unit (IMU), wheel speed sensors, high-precision gyroscopes, vehicle attitude angle sensors, a Global Positioning System (GPS) module, and slope angle sensors, among others.

[0066] The IMU provides six-axis raw signals, including three-axis linear acceleration and three-axis angular velocity, to reflect the vehicle's real-time dynamic behavior in three-dimensional space, including acceleration, braking, steering, and tilt characteristics. A high-precision gyroscope provides roll and yaw angle estimates through angular velocity integration, used to analyze the vehicle's attitude trends and directional changes. Wheel speed sensors provide linear velocity information for each wheel; by comparing differences, it can determine if the vehicle is experiencing potential signs of attitude loss of control, such as slippage, misalignment, or steering torque imbalance. The vehicle attitude angle sensor provides accurate, real-time measurements of the vehicle's "body posture" in three-dimensional space, serving as the most direct reference data for sensing vehicle stability and rollover risk. The GPS module combines vehicle location information with map-marked altitude data to derive the rate of elevation change at the current location and estimate ground slope changes accordingly. The slope angle sensor directly measures the tilt angle between the front and rear axles of the vehicle, serving as a real-time redundant check on the GPS-estimated slope, enhancing the robustness of slope perception.

[0067] To avoid the impact of abnormal sensor signals on subsequent model training, the vehicle-mounted multi-source sensor acquisition platform is equipped with a sliding window and statistical detection method to diagnose and label the stability, drift amplitude, and frequency of abrupt changes of sensor signals in real time. Data frames with abnormal markings are automatically recorded as low-confidence data and selectively removed or repaired in subsequent steps.

[0068] All sensor signals are transmitted to the vehicle-mounted central data synchronization module via the controller area network bus interface or a high-bandwidth Ethernet link. Before reaching the central data synchronization module, each channel's data is individually timestamped to ensure time consistency. After transmission to the central data synchronization module, all data channels undergo time-series reconstruction using interpolation algorithms to eliminate time drift caused by communication delays or inconsistent sampling frequencies, ensuring that sensor data can be aligned at the millisecond level. Finally, the data output by the central data synchronization module is stored in a unified structured format, including time-series vehicle dynamics information, terrain slope data, sensor confidence indicators, and time synchronization markers. The structured data is stored in a high-performance local data recording system and supports batch uploads to a data cloud platform for centralized processing and model training.

[0069] Each frame of data constitutes a data sample, and each data sample encompasses data from all data transmission channels. Simultaneously, both the raw sensor data and the filtered sensor data are retained for subsequent comparative analysis.

[0070] Step 104: Determine whether a high-risk event exists based on the vehicle driving data, and generate an event signal.

[0071] Specifically, vehicle driving data includes multiple frames of continuous data. Analysis of these frames identifies high-risk events. High-risk events are those with a low incidence but a decisive impact, severely affecting vehicle driving safety. They are characterized by their suddenness, often accompanied by rapid changes in gradient, vehicle posture, and nonlinear load distribution disturbances within a short period. To identify high-risk events, it is necessary to filter data from the multiple frames of continuous data that exhibit significant characteristic changes, high rates of change, and high deviations from historical averages. Based on this data, corresponding high-risk events can be analyzed, and event signals can be generated. These event signals provide crucial information for subsequent risk prediction in models.

[0072] Step 106: Based on the vehicle driving data and the event signal, predict multiple frames of future vehicle attitude data using a pre-built vehicle attitude prediction model.

[0073] Specifically, the vehicle attitude prediction model is pre-built and trained, capable of predicting future vehicle attitude data across multiple frames. This multi-frame prediction data reflects the evolution trend of the vehicle attitude. In this embodiment, the vehicle attitude prediction model is a Dynamic Bayesian Network (DBN). A DBN is a probabilistic graphical model used for modeling and analyzing time-series data. Its core function is to correlate variables through time steps, enabling causal relationships and probabilistic inference about dynamic systems. A DBN typically consists of multiple layers, each containing multiple nodes, each representing a latent variable or feature. In the time-series data contained in the vehicle driving data, the input at each time step activates the corresponding node in the network.

[0074] The vehicle's future attitude data includes the state probability distribution of vehicle body attitude variables. During the prediction process, vehicle driving data and event signals are mapped to corresponding nodes in the DBN. Based on the conditional probability inference mechanism, and combining the currently mapped node (observed state) and the parent node, forward inference is performed on unobserved nodes, outputting the state probability distribution of vehicle body attitude variables over several future time frames. For example, these future time frames can cover multiple time points such as 100ms, 200ms, and 300ms, providing a prediction timeline. For each time point, the DBN outputs the probability distribution of multiple vehicle body attitude variables. Vehicle body attitude variables can include roll angle, lateral acceleration, and slope angle. Taking roll angle as an example, at the future time t+300ms, the state probability distribution output by the DBN can be in the form of: {P(roll angle ∈ "safe") = 10%, P(roll angle ∈ "slight instability") = 25%, P(roll angle ∈ "moderate imbalance") = 40%, P(roll angle ∈ "severe rollover boundary") = 20%}. The probability distribution of the roll angle shows that at the future time t+300ms, the most likely state of the roll angle is "moderate imbalance", and the corresponding probability of 40% is the maximum confidence state prediction value.

[0075] Simultaneously, the vehicle's future attitude data can also include event chain path numbers. The path corresponding to the event chain path number can explain the reason for the probability distribution of the vehicle attitude variables output by the DBN, that is, the inference path corresponding to the probability distribution of the vehicle attitude variables in the model. For example, if the path corresponding to the event chain path number is "slope + steering", then it can be determined that the probability distribution of the vehicle attitude variables output by the current model is caused by slope and steering. The vehicle's future attitude data can also include initial risk level labels. The initial risk level labels are the risk levels obtained by the DBN model after integrating various vehicle attitude variables, reflecting the degree of future vehicle rollover risk.

[0076] Step 108: Based on the vehicle driving data, the event signal, and the multi-frame vehicle future attitude data, obtain the risk level through a pre-built risk level classification model.

[0077] Specifically, the risk level classification model in this embodiment is the LightGBM (Light Gradient Boosting Machine) model. The LightGBM model employs histogram-based decision tree algorithms and one-sided gradient sampling techniques, resulting in fast training and extremely fast prediction and inference speeds (milliseconds), meeting the real-time requirements of the vehicle controller. The gradient boosting framework itself has strong expressive power and can effectively capture complex nonlinear relationships between features. The LightGBM model is insensitive to noise and outliers in the data and supports parallel training.

[0078] The input data for the risk level classification model includes vehicle driving data, event signals, and multi-frame vehicle future attitude data. The input data dimension is generally between 20 and 40, varying slightly depending on the vehicle sensor configuration and data channel definition. The output data of the risk level classification model includes the risk level and its corresponding confidence level. Risk levels can include safe, slightly unstable, moderately unbalanced, and severe rollover boundary. The risk level classification model can classify the risk level of the current frame data, providing a fast and clear basis for judgment for the vehicle control system.

[0079] Step 110: Determine target risk decision data based on the event signal, the multi-frame vehicle future attitude data, and the risk level.

[0080] Specifically, by fusing event signals, multi-frame vehicle future attitude data, and risk levels, target risk decision data is formed, including target risk assessment and control commands. Through the fusion process, information complementarity and dynamic weight adjustment between models are achieved, thereby improving prediction stability and output consistency in complex scenarios. The target risk level can be set to four levels: low risk, medium risk, high risk, and extremely high risk. Control commands also include four different priority levels: low-priority control commands (no active intervention), medium-priority control commands (warning), high-priority control commands (light braking), and extremely high-priority control commands (full braking and stability control).

[0081] Based on steps 102 to 110 above, this embodiment provides a vehicle risk prediction method, including: collecting vehicle driving data, determining whether a high-risk event exists based on the vehicle driving data, and generating an event signal. High-risk events are sudden events; identifying sudden events can improve the timeliness of subsequent risk prediction. Based on the vehicle driving data and the event signal, a pre-built vehicle posture prediction model predicts multiple frames of future vehicle posture data. These multiple frames of future vehicle posture data can predict the future driving state of the vehicle, which is beneficial for early identification of vehicle risks. If a high-risk event exists, the event signal input to the vehicle posture prediction model can drive the model to predict multiple frames of future vehicle posture data with higher risk, improving prediction accuracy. Based on the vehicle driving data, the event signal, and the multiple frames of future vehicle posture data, a risk level is obtained through a pre-built risk level classification model. The risk level can provide a fast and clear basis for subsequent risk decisions. Based on the event signal, the multiple frames of future vehicle posture data, and the risk level, target risk decision data is determined. By integrating event signals, multi-frame vehicle future attitude data, and risk levels, a comprehensive and reliable basis can be provided for risk decision-making. This is beneficial for improving the stability and consistency of target risk decision-making data in complex scenarios, thereby enhancing the accuracy and timeliness of vehicle risk prediction.

[0082] Since multi-sensor data is multi-source and asynchronous, it is necessary to perform data cleaning and unified timeline reconstruction on the collected multi-sensor data. For multi-source asynchronous data, missing value imputation, time series synchronization, and outlier removal are performed to reconstruct vehicle time series data with a unified timeline, generating a high-quality structured dataset for subsequent model prediction and model training.

[0083] Although multi-sensor data originates from high-precision sensor modules, it often contains noise information such as missing frames, time misalignments, abnormal jumps, drift, and jitter due to issues such as different sampling frequencies, communication delays, transmission bandwidth fluctuations, and electromagnetic interference under vehicle operating conditions. Therefore, a robust data cleaning mechanism is needed, with temporal consistency as the core, while also considering data integrity, physical rationality, and processing efficiency.

[0084] All sensor signals are initially sorted according to their local timestamps and cached in a unified processing queue. Then, using the IMU channel's timeline as the primary reference axis, time interpolation or truncation is performed on the remaining channels to ensure each time frame has complete feature dimensions. If a channel experiences a short-term signal interruption, interpolation is performed within the maximum permissible time window using a linear prediction method based on the first derivative trend. For data gaps exceeding a threshold in the time window, these gaps are marked as unusable regions to prevent distortion during subsequent model training. After timestamp synchronization is complete, a smoothing filter strategy is applied to each channel signal to eliminate transient glitches, high-frequency noise, and oscillations. The filtering process must consider the true frequency distribution of the vehicle's dynamic response to avoid weakening key features.

[0085] For key attitude and terrain variables such as vehicle speed, acceleration, roll angle, and slope angle, physical boundary detection logic also needs to be constructed. For example, vehicle speed should not be negative, and slope angle should not experience a reverse jump greater than 30 degrees in a short period of time. For data points that do not conform to physical laws, they are not directly discarded. Instead, a multivariate collaborative detection strategy is used to identify their context, classifying such data into three different levels: highly anomaly, low confidence, and boundary fluctuation. These are then stored in different labeled datasets for selective use in the subsequent fusion stage.

[0086] Unified timeline reconstruction involves not only sampling point alignment but also collaborative reconstruction of cross-sensor events. During sudden slope changes or attitude shifts in a vehicle, the response delays of different channels naturally differ; for example, GPS slope estimation has a computational delay of 100 to 200 milliseconds, while IMU angular velocity can respond in real time. To eliminate this information asymmetry, a dynamic window alignment strategy is used. After detecting a sudden change in a channel, the resonant behavior of other channels within similar time windows is searched and slid-paired to form event anchors. These event anchors can serve as causal relationships in subsequent vehicle attitude prediction model construction, improving the effectiveness and reliability of temporal modeling. After data cleaning and unified timeline reconstruction, the data is packaged and stored in standard data segments, each with a fixed length of 10 seconds and containing 1000 frames. Each data segment has a unique sequence number and event label index, serving as the core training input data for subsequent vehicle attitude prediction and risk level classification models.

[0087] The data cleaning and unified timeline process is highly dependent on the changing patterns of real road test data within the automaker. Therefore, the statistical distribution of the automaker's past operating conditions data was introduced as a basic threshold setting standard during the design process to ensure that the cleaning strategy is close to the actual operating conditions rather than a generalization rule, thereby further improving the generalization ability and stability of the subsequent model in extreme slope scenarios.

[0088] In some embodiments, determining whether a high-risk event exists based on the vehicle driving data and generating an event signal includes:

[0089] Based on the vehicle driving data, the values ​​corresponding to each preset indicator are calculated.

[0090] Based on the values ​​corresponding to each preset indicator and the preset event trigger threshold, it is determined whether a high-risk event exists and an event signal is generated.

[0091] Specifically, vehicle driving data includes multiple frames of continuous data. Analysis of these frames can identify high-risk events. High-risk events are those with a low incidence but decisive impact. They are characterized by their suddenness and are often accompanied by rapid changes in slope, vehicle posture, and nonlinear load distribution disturbances. To identify high-risk events, it is necessary to filter data from the multiple frames of continuous data that exhibit significant characteristic changes, high rates of change, and high deviations from historical averages. The process of generating event signals employs the Temporal Point Process (TPP) analysis method. TPP is a type of stochastic process used to simulate discrete events occurring over continuous time, and it is widely applied in event sequence modeling, causal inference, and cross-domain analysis.

[0092] In practice, all vehicle driving data is differentially processed to extract first-order and second-order derivative features to construct various preset indicators. These preset indicators can characterize the dynamic trends of vehicles, and then the presence of high-risk events can be determined based on these trends.

[0093] Vehicle driving data primarily includes structured time series data that has undergone data cleaning and temporal reconstruction. This structured time series contains multiple data points such as roll angle, pitch angle, yaw angle, vehicle speed, acceleration, wheel speed difference, slope angle, GPS elevation change rate, terrain curvature index, and sensor confidence level. Preset indicators include slope angle change rate, roll acceleration, and lateral acceleration change rate. The values ​​of each preset indicator can be calculated from the vehicle driving data. For each preset indicator, a corresponding event trigger threshold is pre-set. If the value of a preset indicator exceeds the corresponding event trigger threshold, it indicates the presence of a high-risk event. The event trigger threshold can be determined based on the automaker's historical test data to ensure it is adapted to the actual operating characteristics of the vehicle. For example, if the slope angle change rate and roll acceleration exceed the corresponding event trigger thresholds in three consecutive frames of data, a high-risk event is determined to exist.

[0094] If a high-risk event exists, the generated event signal includes the high-risk event signal and its corresponding label. The label describes the high-risk event type, start and end timestamps, peak index, and sparsity score. The sparsity score characterizes the rarity of the high-risk event. For example, the peak index could be the maximum rate of change of gradient, the maximum roll angle, etc. Event types could include steep uphill climbs, roll edges, etc. If no high-risk event exists, the flag corresponding to the high-risk event signal in the generated event signal is set to zero, indicating that no high-risk event exists. The purpose of high-risk event identification and labeling is to identify and extract high-risk events that account for a low percentage of the overall sample but have a decisive impact on the stability of the vehicle from the collected continuous vehicle driving data. By calculating the values ​​of various preset indicators, the existence of high-risk events can be accurately determined, and then event signals can be generated. Event signals can serve as an important basis for subsequent model risk identification, improving the sensitivity and response speed of subsequent models in predicting vehicle rollover risks.

[0095] In some embodiments, the step of predicting multiple frames of future vehicle attitude data based on the vehicle driving data and the event signal using a pre-built vehicle attitude prediction model includes:

[0096] In response to the fact that the event signal does not include a high-risk trigger signal, based on the vehicle driving data, the vehicle body posture prediction model uses a first preset inference mechanism to predict multiple frames of future vehicle posture data.

[0097] In response to the event signal, including the high-risk trigger signal, based on the vehicle driving data and the high-risk trigger signal, the vehicle attitude prediction model uses a second preset inference mechanism to predict multiple frames of future vehicle attitude data.

[0098] Specifically, the prediction process of the vehicle posture prediction model includes two inference mechanisms: a first preset inference mechanism and a second preset inference mechanism. The first preset inference mechanism corresponds to a first conditional probability table, and the second preset inference mechanism corresponds to a second conditional probability table. The conditional probability table records the probability that the child nodes (results) will take each possible state given a specific combination of states processed by a parent node (cause).

[0099] The first preset inference mechanism is a conventional inference mechanism, based on normal vehicle dynamics for future projection. When the event signal does not include high-risk trigger signals, the vehicle attitude prediction model performs inference and prediction based on the first preset inference mechanism, which corresponds to the first conditional probability table. The first conditional probability table is learned from normal vehicle driving data and low-to-medium risk driving data. Based on the vehicle driving data, the vehicle attitude prediction model uses the first preset inference mechanism to predict multiple frames of future vehicle attitude data. After each parameter in the vehicle driving data is mapped to each node in the vehicle attitude prediction model, the vehicle attitude prediction model determines the possible probability distribution by querying the first conditional probability table. For example, based on the slope angle and vehicle speed in the vehicle driving data, the corresponding nodes in the vehicle posture prediction model are mapped. The state value is determined to be steep based on the slope angle S, and the state value is determined to be medium speed based on the vehicle speed V. Based on the query conditions S=steep and V=medium speed, the probability distribution of the roll angle is obtained by querying the first conditional probability table: {P(roll angle ∈ "normal") = 10%, P(roll angle ∈ "slight tilt") = 25%, P(roll angle ∈ "severe tilt") = 40%, P(roll angle ∈ "critical roll") = 25%}.

[0100] The second preset inference mechanism is an event-driven inference mechanism that extrapolates the future based on high-risk events. When the event signal includes a high-risk trigger signal, the vehicle attitude prediction model performs inference and prediction based on the second preset inference mechanism, which corresponds to the second conditional probability table. The second conditional probability table is learned from the vehicle's high-risk event driving data to learn the nonlinear dynamic behavior of the vehicle under sudden, extreme stress. For example, how the vehicle attitude will deteriorate uncontrollably when the wheels instantly lift off the ground or encounter severe potholes. Based on the vehicle driving data and the high-risk trigger signal, the vehicle attitude prediction model uses the second preset inference mechanism to predict multiple frames of future vehicle attitude data. After the parameters in the vehicle driving data and the high-risk event signal are mapped to the nodes in the vehicle attitude prediction model, the input of the high-risk event signal activates the high-risk event node in the DBN, and the vehicle attitude prediction model determines the possible probability distribution by querying the second conditional probability table. Compared to the first conditional probability table, the probability of an increase in roll angle due to the same small slope change is significantly increased under the action of the second conditional probability table. For example, based on the slope angle and vehicle speed in the vehicle driving data, the corresponding nodes in the vehicle posture prediction model are mapped. The state value is determined to be steep based on the slope angle S, and the state value is determined to be medium speed based on the vehicle speed V. Based on the query conditions S=steep and V=medium speed, the probability distribution of the roll angle is obtained by querying the second conditional probability table: {P(roll angle ∈ "normal") = 5%, P(roll angle ∈ "slight tilt") = 10%, P(roll angle ∈ "severe tilt") = 50%, P(roll angle ∈ "critical roll") = 35%}.

[0101] This embodiment presents a joint inference mechanism triggered by event signals and DBN. When the event signals do not include high-risk event signals, multiple frames of future vehicle attitude data are inferred according to a conventional first preset inference mechanism. When the event signals include high-risk event signals, multiple frames of future vehicle attitude data are inferred according to a second preset inference mechanism. This ensures that high-risk event signals are given sufficient consideration when predicting the vehicle's future attitude, and timely outputs vehicle attitude data that matches the high-risk event, thereby improving the vehicle's ability to cope with sudden risks and ensuring vehicle driving safety.

[0102] In some embodiments, the method for constructing the vehicle posture prediction model includes:

[0103] Collect historical vehicle time-series data;

[0104] Based on the historical vehicle time series data, pre-constructed high-risk event samples, and predefined physical prior knowledge, an initial dynamic Bayesian network is constructed.

[0105] Based on the historical vehicle time series data, the initial dynamic Bayesian network is subjected to parameter learning and structural optimization to obtain the vehicle posture prediction model.

[0106] Specifically, historical vehicle time-series data includes driving data for each vehicle within consecutive time periods, such as roll angle, pitch angle, yaw rate, lateral acceleration, longitudinal acceleration, vehicle speed, slope angle, rate of change of slope angle, center of gravity offset, lateral stability factor, etc. Historical vehicle time-series data provides the state changes of variables over time. These variables exhibit clear temporal dependencies and physical coupling characteristics in actual vehicle dynamic behavior. High-risk event samples are pre-constructed data containing specific vehicle driving data within the time period of high-risk events and corresponding labels. Physical prior knowledge refers to the prior knowledge associated with physical variables defined by the automaker.

[0107] Based on the historical vehicle time-series data, pre-constructed high-risk event samples, and predefined physical prior knowledge, an initial dynamic Bayesian network is constructed. Each variable in the historical vehicle time-series data can serve as a node in the initial dynamic Bayesian network, the high-risk event samples can serve as key nodes, and the physical prior knowledge is used to guide the initial structure design and edge weight biasing of the initial dynamic Bayesian network.

[0108] The nodes in the initial dynamic Bayesian network consist of various variables from historical vehicle time-series data, including but not limited to roll angle, pitch angle, yaw angle, lateral acceleration, longitudinal acceleration, slope angle, vehicle speed, wheel speed difference, center of mass offset, and lateral stability factor. The initial dynamic Bayesian network is composed of multiple identical time slices (e.g., time slice at time t-1, time slice at time t, and time slice at time t+1). The time slice at time t represents the "present and past," and the time slice at time t+1 represents the "future."

[0109] Each time slice contains nodes representing all the variables mentioned above. Edges between nodes represent direct causal dependencies between variables at that time (e.g., slope angle t → roll angle t, indicating that the slope angle at time t directly causes vehicle roll at time t). Edges between adjacent time slices represent the transmission and evolution paths of variables over time (e.g., lateral acceleration t → roll angle t+1, indicating that the acceleration at time t has a cumulative effect on the roll angle at time t+1). This structure constitutes a typical dynamic Bayesian network, where each node can receive influences from similar variables from the previous time step, as well as cross-influences from other variables at the current time step.

[0110] High-risk event samples are added as key nodes to the time slice. These key nodes, acting as temporary causes, are connected to nodes directly affected by them. This allows for the dynamic activation of specific predicted paths when a high-risk event is triggered, simulating the vehicle's attitude evolution response under sudden events. For example, a high-risk slope event t → slope angle t+1 means that the high-risk slope event at time t directly affects the slope angle at time t+1.

[0111] In the initial construction of the dynamic Bayesian network, it is also necessary to build the structural framework based on prior physical knowledge. This prior physical knowledge includes two types of information: one is the variable relationships mined from real vehicle operation data using criteria such as information gain and maximum expected entropy change; the other is the prior dependency of variables extracted based on vehicle dynamics theory and structural models. For example, the high coupling between lateral acceleration and roll angle during rapid curvature changes, or the hysteresis response between the rate of change of slope and longitudinal acceleration, these known relationships will serve as bounding constraints in the initial construction of the dynamic Bayesian network. Based on the prior physical knowledge, it can be determined which relationships between variables should be preserved and which should be weakened or pruned.

[0112] In the initial dynamic Bayesian network, each edge corresponds to a conditional dependency. The directionality of the edge represents the information propagation path, and the existence of the edge indicates whether there is a significant statistical relationship between the variables. During the construction of the initial dynamic Bayesian network, it is also necessary to determine whether there are closed loops, redundant paths, or self-excited feedback loops in the causal chain, and to perform penal pruning on paths that cannot be validated by data.

[0113] The final constructed initial dynamic Bayesian network outputs in a topologically stable form. The node states and edge connectivity information for each time slice are encoded into a standard graph description structure and stored. The initial dynamic Bayesian network possesses complete node coverage, defined edge directions, and causal chain connectivity across multiple time slices, ensuring that it can perform rolling predictions of the vehicle's state for several future frames based on partial observation data when entering the prediction phase. The initial dynamic Bayesian network is a reasoning-enabled, interpretable graph structure network with temporal logic, capable of supporting early prediction and path deduction of future rollover risks.

[0114] The initial dynamic Bayesian network also supports the insertion of future vehicle state variables such as real-time tire pressure and driving force distribution for scalability enhancement. The initial dynamic Bayesian network construction process is technically implemented using an automotive-grade graphing engine, supporting automatic graph generation, structure evaluation, and version rollback management, ensuring a complete closed-loop lifecycle management for the initial dynamic Bayesian network.

[0115] After the initial dynamic Bayesian network is constructed, it needs to undergo parameter learning and structural optimization to obtain the vehicle posture prediction model. This process uses the automaker's own historical data for DBN parameter estimation (generating a Conditional Probability Table (CPT)) and structural learning. The graph structure of the initial dynamic Bayesian network is iteratively optimized using a scoring function and a heuristic search strategy to improve prediction accuracy and causal expression capabilities.

[0116] The goal of parameter learning is to transform the previously constructed initial dynamic Bayesian network from a static topology into a dynamic probabilistic model with temporal reasoning capabilities. In the initial dynamic Bayesian network, each node corresponds to a variable, and its state value originates from the discretized results of historical vehicle time-series data. Since vehicle operating states have continuous characteristics, intervals are divided according to the statistical distribution and engineering semantics of the variables, converting continuous values ​​into state values. For example, roll angle is divided into normal, slight tilt, severe tilt, and critical rollover states. The discretization operation must be combined with the vehicle's own dynamic response characteristics and safety thresholds, and tested to ensure no information loss in the results. The state value of each node, together with the state of its parent node, constitutes a conditional probability space. By statistically analyzing the co-occurrence frequency of each state combination in the historical vehicle time-series data, the value of each conditional probability is estimated, generating a complete conditional probability table. That is, for each node, the historical vehicle time-series data is traversed, and the frequency of the node itself taking each state under each possible state combination of its parent node is statistically analyzed. This process, provided there is sufficient sample coverage, can be directly accomplished using frequency estimation, directly normalizing the frequencies to probabilities to form a conditional probability table. In sparse sample regions, smoothing mechanisms are introduced, such as Laplace correction or weight averaging based on Bayesian priors, to ensure the model is robust in boundary states.

[0117] After parameter learning is complete, a structure optimization process is performed to re-evaluate the rationality and predictive ability of the current initial dynamic Bayesian network structure. Structure optimization uses the Bayesian information criterion or minimum description length criterion as the core evaluation indicators, optimizing the generalization ability and computational efficiency of the initial dynamic Bayesian network while ensuring the interpretability of the network structure. A combination of greedy search and heuristic search is employed to perform local perturbations and global evaluations in the network structure space. Local perturbation operations include adding, deleting, and adjusting the direction of edges. After each perturbation, the initial dynamic Bayesian network is evaluated based on its log-likelihood score on the validation set or its accuracy in identifying rollover events. Local perturbations that significantly improve the evaluation indicators are selected for modification. During structure optimization, multiple initial dynamic Bayesian networks formed after local perturbations are compared in terms of performance, and the initial dynamic Bayesian network corresponding to the optimal balance between structural complexity and inference ability is ultimately selected as the output.

[0118] For redundant edges that are highly correlated but lack causality between dependent variables, pruning is performed based on the predicted marginal benefit. For paths that are currently insensitive to certain anomalous events, the sensitivity of the initial dynamic Bayesian network is enhanced by introducing new edges or guiding edge direction adjustments.

[0119] The output initial dynamic Bayesian network (vehicle attitude prediction model) includes a set of model parameters, including the state definition of each node, the conditional probability distribution of each parent node combination, the topology of all edges, edge directions, and activation conditions. These model parameters are serialized and saved, and can be used for model loading or online deployment in the subsequent prediction stage. To ensure the engineering usability of the model parameters, the parameter learning and structure optimization processes are trained based on the automaker's own test scenarios and data resources, covering typical scenarios such as urban driving conditions, highway scenarios, and extreme mountain slopes. Scenario sampling and balancing are performed to avoid overfitting or failure of the model in specific scenarios. The model parameters also include a set of model uncertainty indicators, such as variable state boundary ambiguity and event trigger probability confidence intervals, used for target risk level assessment and control command adjustment in the subsequent decision output stage. The parameter learning and structure optimization process runs on an automotive-grade edge computing platform and supports an incremental model training mechanism, allowing for continuous structural fine-tuning and parameter updates based on newly collected data, ensuring the model's long-term online deployment capability.

[0120] The construction method in this embodiment yields a vehicle attitude prediction model that can understand the causal relationship of vehicle dynamics and infer future changes in vehicle attitude, greatly improving the accuracy and generalization ability of the prediction.

[0121] In some embodiments, the method for constructing the high-risk event sample includes:

[0122] Based on the historical vehicle time series data, calculate the historical values ​​corresponding to each preset indicator;

[0123] Based on the historical data and preset event trigger thresholds, determine whether a high-risk event exists;

[0124] In response to the existence of a high-risk event, data related to a preset time period is extracted from the historical change data based on the trigger time of the high-risk event to form event window data corresponding to the high-risk time.

[0125] The event window data is labeled to form event tags;

[0126] Calculate the sparsity score based on the event window data;

[0127] The event window data, the event labels, and the sparsity scores are used as samples of high-risk events.

[0128] Specifically, historical vehicle time-series data includes multiple consecutive frames of data. Analysis of these frames can identify high-risk events. High-risk events are those with a low incidence but significant impact; they are characterized by suddenness and are often accompanied by rapid changes in slope, vehicle posture, and nonlinear load distribution disturbances. To identify high-risk events, it is necessary to filter data from the multiple consecutive frames that exhibit significant feature changes, high rates of change, and high deviations from historical averages. In practice, all vehicle driving data is differentially processed to extract first- and second-order derivative features, constructing various preset indicators. These indicators characterize the dynamic trends of vehicles, allowing for the determination of whether high-risk events exist.

[0129] Historical vehicle time-series data primarily comprises structured sequences that have undergone data cleaning and time-series reconstruction. These sequences contain multiple data points, including roll angle, pitch angle, yaw angle, vehicle speed, acceleration, wheel speed difference, slope angle, GPS elevation change rate, terrain curvature index, and sensor confidence level. Preset indicators include slope angle change rate, roll acceleration, and lateral acceleration change rate. Historical values ​​for each preset indicator can be calculated from vehicle driving data. For each preset indicator, a corresponding event trigger threshold is pre-set. If the value of a preset indicator exceeds the corresponding event trigger threshold, a high-risk event is identified. The event trigger threshold can be determined based on the automaker's historical test data to ensure it is adapted to the actual operating characteristics of the vehicle. For example, if the slope angle change rate and roll acceleration exceed the corresponding event trigger thresholds in three consecutive frames of data, a high-risk event is identified.

[0130] If a high-risk event exists, the moment of the high-risk event is taken as the center moment. Contextual information from several frames before and after the center moment is extracted to form event window data. For example, the length of the event window data is 1 to 2 seconds. All data within the event window data is grouped into the same high-risk event sample, and data such as the occurrence time, direction of change, duration, maximum offset, and peak attitude index are labeled as event tags. The direction of change can include whether it is a rollover precursor, the slope change trend, and the direction of vehicle attitude deviation.

[0131] Furthermore, a distance metric is calculated between each event window's data and the global data distribution. This distance metric includes the mean difference, standard deviation shift, and outlier score. The distance metric allows for the ranking of the sparsity of the time window data, generating a sparsity score. Event window data that statistically deviates significantly from the normal distribution are defined as high-risk sparse events, and their weights are correspondingly increased to avoid training bias caused by the low frequency of original samples for high-risk sparse events. The event window data, event labels, and sparsity scores are used as samples for these high-risk events.

[0132] The construction of high-risk event samples will serve as important basic information in the subsequent construction of the vehicle body attitude prediction model. It will be used to build causal dependencies between vehicle body variables and activate high-risk prediction paths in the prediction stage, so that the vehicle body attitude prediction model can make predictions according to the second preset inference mechanism, thereby improving the sensitivity and response speed of the vehicle body attitude prediction model to sudden rollover risks.

[0133] In some embodiments, obtaining the risk level based on the vehicle driving data, the event signal, and the multi-frame vehicle future attitude data through a pre-built risk level classification model includes:

[0134] The current operating condition of the vehicle is determined based on the vehicle driving data.

[0135] The target risk level classification model is selected from multiple pre-built risk level classification models based on the current operating condition of the vehicle.

[0136] Based on the vehicle driving data, the event signal, and the multi-frame vehicle future attitude data, the risk level is obtained through the target risk level classification model.

[0137] Specifically, vehicle driving data can determine the vehicle's current operating condition, which may include high-speed driving, low-speed urban turning, and off-road driving. Different risk level classification models can be trained based on historical data under different operating conditions. Selecting different risk level classification models can improve the accuracy of risk level classification under different operating conditions. Based on the current operating condition, the corresponding risk level classification model is selected from multiple risk level classification models as the target risk level classification model.

[0138] The input data for the risk level classification model includes vehicle driving data, event signals, and multiple frames of future vehicle attitude data. Vehicle driving data includes roll angle, pitch angle, lateral acceleration, longitudinal acceleration, vehicle speed, wheel speed, roll angle change rate, lateral acceleration change rate, and slope angle change rate. Event signals indicate to the risk level classification model whether the vehicle is currently in a high-risk event. Multiple frames of future vehicle attitude data provide "foresight" for the risk level classification model.

[0139] The risk level classification model can classify the risk level of the current frame data, providing a fast and clear basis for judgment for the vehicle control system. The risk level classification model, the vehicle attitude prediction model, and the high-risk event identification form a synergistic relationship: the vehicle attitude prediction model is responsible for modeling the attitude evolution trend over a long period of time, the high-risk event identification is responsible for making advance predictions of risks, and the risk level classification model completes the risk assessment of the current frame of vehicle driving data at the millisecond level when the system needs an immediate response.

[0140] The risk level classification model outputs a five-dimensional vector, corresponding to the confidence levels of four risk levels and the recommended vehicle control priority weights. The four risk levels can be set as safe, slightly unstable, moderately unbalanced, and severely rollover-bound. The vehicle control priority weights provide the weight coefficients for the final vehicle control commands, which can be represented by specific values ​​ranging from 0 to 1. For example, a vehicle control priority weight of 0 indicates a recommendation for no intervention; 0.3 indicates a recommendation for low-level intervention; 0.7 indicates a recommendation for medium-level intervention; and 1 indicates a recommendation for high-level intervention. The data output by the risk level classification model provides rich diagnostic information for subsequent decision-making, ensuring the transparency and interpretability of the decisions, while also providing concise control commands to ensure the vehicle control system can act quickly and smoothly.

[0141] In the training process of the risk level classification model, the training data comes from the historical road test data of car manufacturers. Risk level labels are generated by labeling the training data, and training samples are formed based on the training data and risk level labels. The risk level classification model is trained using these training samples. Rollover boundary scenarios labeled in the historical road test data of car manufacturers are used as positive samples, and large-scale normal driving data are used as negative samples. Downsampling and weight adjustment are used to address the sample ratio imbalance problem, and temporal consistency constraints are introduced to ensure the interpretability of the prediction results in terms of temporal continuity. The internal structure of the risk level classification model consists of multiple decision trees with a depth of 6 to 10. Each tree learns the nonlinear mapping relationship between input features and risk levels through training samples. Each leaf node of the tree outputs a risk probability value, and the outputs of all trees are weighted and averaged to form the final classification result. The risk level classification model uses grouped cross-validation to evaluate training accuracy.

[0142] The training and testing of the risk classification model are entirely based on the automaker's own road test data, covering different terrains, vehicle loads, tire configurations, and driving styles to ensure sufficient generalization ability and adaptability. Simultaneously, online learning and model fine-tuning mechanisms are introduced, supporting optimization and iteration through vehicle-side data feedback and remote model updates after system deployment, achieving continuous performance improvement throughout the system's lifecycle. As the only single-frame real-time judgment path in the system, the risk classification model exhibits extremely high stability, efficiency, and control adaptability in actual deployment, providing a clear, stable, and engineered risk identification foundation for target risk decision-making data.

[0143] The risk level classification model is deployed in the vehicle gateway or central processing unit, with an execution cycle of less than 10 milliseconds, enabling high-frequency operation without affecting the resource allocation of the vehicle control system. By incorporating multiple frames of future vehicle attitude data, the model enhances the prior perception capability of the current frame classification, achieving cross-information enhancement between multiple models. For sudden risk frames where the predicted rollover risk from future vehicle attitude data is weak but the current signal changes drastically, the risk classification model can supplement the identification by using event signals from the input features, avoiding system misjudgment. Conversely, for future rollover risk frames where the current frame signal changes little but the confidence level of the future vehicle attitude data is extremely high, the risk classification model can assist the vehicle control system in entering a protection state in advance.

[0144] In some embodiments, determining the target risk decision data based on the event signal, the multi-frame vehicle future attitude data, and the risk level includes:

[0145] Determine whether the event signal, the multi-frame vehicle future attitude data, and the risk level meet preset high-risk conditions; in response to meeting the preset high-risk conditions, use the preset risk decision data corresponding to the preset high-risk conditions as the target risk decision data.

[0146] Specifically, target risk decision data includes the final target risk level and control instructions with a certain priority. Two methods are used to determine target risk decision data, depending on whether preset high-risk conditions are met. One method is based on event signals, multi-frame vehicle future attitude data, and the risk level to determine whether preset high-risk conditions are met.

[0147] The preset high-risk conditions specifically include: if the event signal includes a high-risk event signal with extremely high confidence (e.g., confidence > 0.95), and the danger probability value of the vehicle's future attitude data predicted over multiple frames exceeds a preset threshold (e.g., 60%), then the high-risk condition is determined to be met. In this case, regardless of whether the risk level output by the risk level classification model is high-risk, the target risk decision data is directly determined as the preset risk decision data corresponding to this high-risk condition. The target risk level in the corresponding preset risk decision data is the highest risk level, and the control command is the highest priority control command, to avoid delays caused by fusion averaging under extreme danger.

[0148] If the risk level output by the risk level classification model jumps sharply from a safe risk level to a severe boundary risk level within consecutive data frames (e.g., 3 frames), then the high-risk condition is determined to be met. In this case, the target risk decision data is determined as the preset risk decision data corresponding to this high-risk condition. The target risk level in the preset risk decision data is higher than the risk level determined by fusing event signals, multi-frame vehicle future attitude data, and the risk level itself. The control command is a medium-priority control command to address entirely new, sudden hazards that neither the vehicle attitude prediction model nor the risk level classification model has learned.

[0149] If any of the event signals, multi-frame vehicle future attitude data, or risk levels have a confidence level lower than a preset threshold, then a high-risk condition is determined to be met. In this case, the target risk decision data is determined to be the preset risk decision data corresponding to the high-risk condition at this time. The target risk level in the corresponding preset risk decision data is the highest risk level, and the control command is to prohibit active vehicle control, only issue an audible and visual alarm, and return vehicle control to the driver.

[0150] The target risk level can be set to four levels, including low risk, medium risk, high risk, and extremely high risk. Control instructions also include control instructions with four different priorities, such as low priority control instructions (no active intervention), medium priority control instructions (early warning), high priority control instructions (mild braking), and extremely high priority control instructions (full braking and stabilization control).

[0151] The target risk decision data may also include a control trigger time window, trigger source description, and confidence level. The control trigger time window indicates the recommended timeframe within which the vehicle controller must initiate control actions. The trigger source description indicates which of the aforementioned models or high-risk events the target risk decision data primarily originates from. The confidence level, ranging from 0 to 1, indicates the reliability of the target risk decision data.

[0152] The method described in this embodiment provides a method for determining target risk decision data that meets preset high-risk conditions, which can ensure vehicle driving safety when dealing with extreme and boundary situations.

[0153] In some embodiments, in response to the failure to meet a preset high-risk condition, the weight coefficients corresponding to the event signal, the multi-frame vehicle future attitude data, and the risk level are calculated respectively.

[0154] The target risk decision data is determined based on the weighting coefficients, the event signals, the multi-frame vehicle future attitude data, and the risk level.

[0155] Specifically, when the preset high-risk conditions are not met, this embodiment fuses event signals, multi-frame vehicle future attitude data, and risk levels to form target risk decision data that includes target risk assessment and control commands. Through the fusion process, information complementarity and dynamic weight adjustment between models are achieved, thereby improving prediction stability and output consistency in complex scenarios. First, the weight coefficients corresponding to the event signals, multi-frame vehicle future attitude data, and risk levels are calculated separately.

[0156] For multi-frame vehicle future attitude data, when the prediction results of multiple frames are highly consistent and have a high confidence level (e.g., the probability of danger for a certain vehicle attitude variable is more than 80% in 5 consecutive frames), the weight coefficient corresponding to the multi-frame vehicle future attitude data is increased. If the prediction results of multiple frames differ significantly and have a low confidence level, the weight coefficient corresponding to the multi-frame vehicle future attitude data is decreased.

[0157] For event signals, if the confidence level of high-risk event signals included in the event signal is high and multiple high-risk events are continuously identified, the weight coefficient of the corresponding event signal is increased. If there are no high-risk event signals in the event signal, the weight coefficient of the corresponding event signal is decreased.

[0158] Regarding risk levels, when the confidence level of a risk level exceeds a certain threshold (e.g., 90%) and the quality of the input sensor signal is high, the weighting coefficient corresponding to the risk level is increased. When the confidence level of a risk level is low, or the sensor signal quality is poor, the weighting coefficient corresponding to the risk level is decreased.

[0159] After determining the weighting coefficients corresponding to the event signal, multi-frame vehicle future attitude data, and risk level, a final risk score is calculated by weighting and summing the weighting coefficients and corresponding confidence levels. The target risk level is then determined based on the risk score. The priority of control commands is then determined according to the target risk level; the higher the target risk level, the higher the priority of the corresponding control commands.

[0160] In the fusion process, in addition to event signals, multi-frame vehicle future attitude data and risk levels, several context-aware indicators are introduced as auxiliary inputs, such as risk level fluctuation curves in the past few frames, the number of control commands triggered, driving style recognition results, and vehicle current load and center of gravity position information, to enhance the adaptability of the fusion process.

[0161] To ensure the stability and verifiability of the aforementioned fusion mechanism during actual deployment, the system supports online fusion weight backtracking, strategy evolution trajectory recording, and model consistency checking mechanisms. The component paths, model contribution ratios, and trigger rule sources for each fusion are packaged and recorded, forming an auditable chain in the system log for later vehicle testing, recall analysis, and model iteration optimization. The fusion mechanism supports flexible deployment strategies. For specific vehicle models and operating conditions, developers can set the preferred risk assessment source models and dynamically adjust the model fusion strategy based on data collected during operation, achieving rapid adaptation and stable deployment across different platforms and control systems.

[0162] The method described in this embodiment achieves information complementarity and cross-validation between the output data of different models and the results of high-risk event identification, avoiding misjudgments of risk due to errors in a single model. The weights of different models can be dynamically adjusted based on real-time data, improving the accuracy of target risk decision data. Furthermore, in extremely dangerous situations such as meeting preset high-risk conditions, the target risk level and control command priority can be rapidly increased, or the vehicle can be handed over to the driver, ensuring absolute safety during the driving process.

[0163] In some embodiments, the method further includes: sending the target risk decision data to relevant controllers of the vehicle, so that each relevant controller executes the control instructions carried in the target risk decision data.

[0164] Specifically, the target risk decision data includes control commands with a certain priority. These commands are pushed in real time to actuators such as the electronic stability control system, four-wheel drive control system, active suspension adjustment system, and electronic steering system, serving as reference signals for proactively adjusting vehicle dynamics. The target risk decision data format conforms to automotive-grade Controller Area Network (CLAN) messages or Ethernet protocols, and can be stably transmitted to the relevant controllers through the vehicle's communication architecture. The method in this embodiment sends the target risk decision data to the relevant vehicle controllers, providing clear and executable commands so that the controllers can proactively adjust the vehicle's state in a timely manner, achieving reliable proactive safety protection.

[0165] The fusion mechanism can be implemented through a fusion model, which is deployed in the vehicle's high-performance gateway domain controller. The output signal interface connects to the active safety controller, enabling the generation of real-time attitude-stabilized control commands. The deployment optimization and control interface integration of the fusion model not only require the engineering deployment of the model but also the establishment of underlying signal communication with the actual control system. This ensures that target risk decision data can drive active safety actuators such as braking, steering, and suspension to complete stable interventions within millisecond timescales. The primary goal of deployment is to ensure that the model inference process has sufficient real-time performance and resource controllability, while the control output results can be transmitted to the vehicle controller in a standardized interface format without affecting the operation of other safety modules.

[0166] The target platform for deploying the fusion model is either a central gateway domain controller or an independent dynamic control auxiliary domain, depending on the vehicle's electronic architecture configuration. Within the limits of platform resources, the fusion model's runtime environment and fusion strategy logic are merged into the same processing unit to reduce cross-domain communication latency. The vehicle attitude prediction model framework requires quantization compression, graph optimization, and batch processing structure adjustments to ensure stable operation in fixed-cycle tasks. The risk level classification model undergoes structural pruning before compilation and deployment, limiting the decision tree depth and the number of nodes. The structure of the vehicle attitude prediction model is transformed into a runtime state machine. The event signal triggering module performs state transition judgments using a finite state machine pattern, with the overall inference time controlled within 30 milliseconds.

[0167] All model input channels must be adapted to the sensing system at the signal level. Signals including IMU, vehicle speed, wheel speed, gradient, GPS elevation difference, and attitude angle must be transmitted to the model execution framework via CAN bus or Ethernet according to a fixed message format. To prevent incomplete input due to communication jitter or packet loss, the system introduces a buffer queue and redundant update mechanism. When critical channel signals are missing, predictive interpolation data is used for compensation without interrupting model operation. The system publishes target risk decision data every 10 milliseconds, using CAN ID for broadcast or point-to-point transmission.

[0168] The key to the control interface integration is ensuring that the model output format matches the target controller protocol. Each type of controller has different frame structures, trigger thresholds, and behavior patterns for received signals. For example, the electronic stability control system responds to instability identification signals based on yaw rate and wheel speed, while the steering system responds to steering angle change rate and lateral stability control logic. Before deployment, interface mapping logic configuration is required for each type of controller, encoding the target risk decision data into intervention requests that the controller can recognize. For instance, setting a high level for a specific bit could represent initiating torque distribution adjustment, or passing a rollover warning level code through a specific data field. To avoid false triggers, a dynamic confidence threshold adjustment mechanism is introduced. When the accuracy of target risk level identification is insufficient, the system output remains in a read-only monitoring state, only recording data without triggering actual control commands.

[0169] After the fusion model is deployed, the system enters the integration testing phase, where closed-loop testing is conducted with the vehicle control strategy. The tests include signal on / off response delay testing, model response behavior under different sensor data qualities, controller's ability to accept control commands and its behavioral response, and whether there are any contradictions or conflicts between the vehicle safety logic and the model output. Extreme tests are performed under different road conditions, driving styles, and combinations of loads and attitudes to ensure that the model does not cause safety risks such as misjudgments, missed judgments, or false triggers in real-world operation. All outputs during the testing process are written to the log system, with frame-level annotations of the inputs, outputs, and controller behavior for each model inference.

[0170] The deployment of the integrated model is the critical point for the model to move from R&D verification to mass production delivery, impacting the stability of the entire vehicle system and the user's driving experience. Therefore, the design retains an online rollback mechanism and a manual overwrite control mode. When the model outputs consecutive low confidence levels or a sensor fault is detected, the system automatically shuts down all control signals, retaining only data logging functionality, and reports the current model status to the vehicle's main control system. This mechanism achieves controllable coupling between model control and the vehicle's safety system, providing a standardized and modular implementation foundation for subsequent replication and deployment in more vehicle models and more complex scenarios.

[0171] It should be noted that the method in this embodiment can be executed by a single device, such as a computer or server. The method can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method in this embodiment, and the multiple devices will interact with each other to complete the method described.

[0172] It should be noted that some embodiments of this application have been described above. In some cases, the actions or steps described in the above embodiments can be performed in a different order than that shown in the above embodiments and the desired result can still be achieved. In addition, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0173] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides a vehicle risk prediction device.

[0174] refer to Figure 2 The vehicle risk prediction device includes:

[0175] The data acquisition module 202 is configured to collect vehicle driving data;

[0176] The determination module 204 is configured to determine whether a high-risk event exists based on the vehicle driving data and generate an event signal;

[0177] The prediction module 206 is configured to predict multiple frames of future vehicle posture data based on the vehicle driving data and the event signal using a pre-built vehicle posture prediction model.

[0178] The classification module 208 is configured to obtain the risk level based on the vehicle driving data, the event signal, and the multi-frame vehicle future attitude data through a pre-built risk level classification model.

[0179] The decision module 210 is configured to determine target risk decision data based on the event signal, the multi-frame vehicle future attitude data, and the risk level.

[0180] In some embodiments, the determining module 204 is further configured to calculate the values ​​corresponding to each preset indicator based on the vehicle driving data; determine whether a high-risk event exists based on the values ​​corresponding to each preset indicator and a preset event triggering threshold, and generate an event signal.

[0181] In some embodiments, the prediction module 206 is further configured to, in response to the event signal not including a high-risk trigger signal, predict multiple frames of future vehicle posture data based on the vehicle driving data using the vehicle posture prediction model and employing a first preset inference mechanism.

[0182] In response to the event signal, including the high-risk trigger signal, based on the vehicle driving data and the high-risk trigger signal, the vehicle attitude prediction model uses a second preset inference mechanism to predict multiple frames of future vehicle attitude data.

[0183] In some embodiments, a construction module is also included, configured to collect historical vehicle time-series data;

[0184] Based on the historical vehicle time series data, pre-constructed high-risk event samples, and predefined physical prior knowledge, an initial dynamic Bayesian network is constructed.

[0185] Based on the historical vehicle time series data, the initial dynamic Bayesian network is subjected to parameter learning and structural optimization to obtain the vehicle posture prediction model.

[0186] In some embodiments, the construction module is further configured to: calculate historical values ​​corresponding to each preset indicator based on the historical vehicle time series data; determine whether a high-risk event exists based on the historical values ​​and a preset event trigger threshold; in response to the existence of a high-risk event, extract associated preset time period data from the historical change data according to the trigger time point of the high-risk event to form event window data corresponding to the high-risk time; label the event window data to form event tags; calculate a sparsity score based on the event window data; and use the event window data, the event tags, and the sparsity score as samples of the high-risk event.

[0187] In some embodiments, the classification module 208 is further configured to determine the current operating condition of the vehicle based on the vehicle driving data; select a target risk level classification model from a plurality of pre-built risk level classification models based on the current operating condition of the vehicle; and obtain the risk level through the target risk level classification model based on the vehicle driving data, the event signal, and the multi-frame vehicle future attitude data.

[0188] In some embodiments, the decision module 210 is further configured to determine whether the event signal, the multi-frame vehicle future attitude data, and the risk level meet a preset high-risk condition; in response to meeting the preset high-risk condition, the preset risk decision data corresponding to the preset high-risk condition is used as the target risk decision data.

[0189] In some embodiments, the decision module 210 is further configured to, in response to the failure to meet a preset high-risk condition, calculate the weight coefficients corresponding to the event signal, the multi-frame vehicle future attitude data, and the risk level respectively; and determine the target risk decision data based on the weight coefficients, the event signal, the multi-frame vehicle future attitude data, and the risk level.

[0190] In some embodiments, an execution module is further included, configured to send the target risk decision data to relevant controllers of the vehicle, so that each relevant controller executes the control instructions carried in the target risk decision data.

[0191] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware.

[0192] The apparatus of the above embodiments is used to implement the corresponding vehicle risk prediction method in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0193] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the vehicle risk prediction method described in any of the above embodiments.

[0194] Figure 3 This embodiment illustrates a more specific hardware structure of an electronic device. The device may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0195] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0196] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0197] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0198] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0199] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0200] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0201] The electronic devices described above are used to implement the corresponding vehicle risk prediction methods in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0202] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the vehicle risk prediction method as described in any of the above embodiments.

[0203] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0204] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the vehicle risk prediction method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0205] Based on the same concept, corresponding to any of the above embodiments, this application also provides a computer program product, including computer program instructions, which, when run on a computer, cause the computer to perform the method described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0206] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0207] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0208] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0209] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. A vehicle risk prediction method, characterized in that, include: Collect vehicle driving data; Based on the vehicle driving data, determine whether a high-risk event exists and generate an event signal; Based on the vehicle driving data and the event signal, multiple frames of future vehicle attitude data are predicted using a pre-built vehicle attitude prediction model, including: In response to the fact that the event signal does not include a high-risk trigger signal, based on the vehicle driving data, the vehicle body posture prediction model uses a first preset inference mechanism to predict multiple frames of future vehicle posture data. In response to the event signal including the high-risk trigger signal, based on the vehicle driving data and the high-risk trigger signal, the vehicle attitude prediction model uses a second preset inference mechanism to predict multiple frames of future vehicle attitude data. Based on the vehicle driving data, the event signal, and the multi-frame vehicle future attitude data, the risk level is obtained through a pre-built risk level classification model. Based on the event signal, the multi-frame vehicle future attitude data, and the risk level, target risk decision data is determined.

2. The method according to claim 1, characterized in that, The step of determining whether a high-risk event exists based on the vehicle driving data and generating an event signal includes: Based on the vehicle driving data, the values ​​corresponding to each preset indicator are calculated. Based on the values ​​corresponding to each preset indicator and the preset event trigger threshold, it is determined whether a high-risk event exists and an event signal is generated.

3. The method according to claim 1, characterized in that, The method for constructing the vehicle body attitude prediction model includes: Collect historical vehicle time-series data; Based on the historical vehicle time series data, pre-constructed high-risk event samples, and predefined physical prior knowledge, an initial dynamic Bayesian network is constructed. Based on the historical vehicle time series data, the initial dynamic Bayesian network is subjected to parameter learning and structural optimization to obtain the vehicle posture prediction model.

4. The method according to claim 3, characterized in that, The method for constructing the high-risk event samples includes: Based on the historical vehicle time series data, calculate the historical values ​​corresponding to each preset indicator; Based on the historical data and preset event trigger thresholds, determine whether a high-risk event exists; In response to the existence of a high-risk event, data related to a preset time period is extracted from the historical change data based on the trigger time of the high-risk event to form event window data corresponding to the high-risk event; The event window data is labeled to form event tags; Calculate the sparsity score based on the event window data; The event window data, the event labels, and the sparsity scores are used as samples of high-risk events.

5. The method according to claim 1, characterized in that, The risk level is obtained based on the vehicle driving data, the event signal, and the multi-frame vehicle future attitude data through a pre-built risk level classification model, including: The current operating condition of the vehicle is determined based on the vehicle driving data. The target risk level classification model is selected from multiple pre-built risk level classification models based on the current operating condition of the vehicle. Based on the vehicle driving data, the event signal, and the multi-frame vehicle future attitude data, the risk level is obtained through the target risk level classification model.

6. The method according to claim 1, characterized in that, The determination of target risk decision data based on the event signal, the multi-frame vehicle future attitude data, and the risk level includes: Determine whether the event signal, the multi-frame vehicle future attitude data, and the risk level meet the preset high-risk conditions; In response to the fulfillment of preset high-risk conditions, preset risk decision data corresponding to the preset high-risk conditions is used as the target risk decision data.

7. The method according to claim 6, characterized in that, Also includes: In response to the failure to meet the preset high-risk conditions, the weight coefficients corresponding to the event signal, the multi-frame vehicle future attitude data, and the risk level are calculated respectively. The target risk decision data is determined based on the weighting coefficients, the event signals, the multi-frame vehicle future attitude data, and the risk level.

8. The method according to claim 1, characterized in that, Also includes: The target risk decision data is sent to the relevant controllers of the vehicle so that the relevant controllers can execute the control instructions carried in the target risk decision data.

9. A vehicle, characterized in that, The vehicles include: Memory, used to store executable program code; A processor for calling and running the executable program code from the memory, causing the vehicle to perform the method as described in any one of claims 1 to 8.

Images