Vehicle risk prediction method and vehicle

Abstract

This application provides a vehicle risk prediction method and a vehicle, relating to the field of vehicle safe driving technology. The method includes: acquiring vehicle operating status data and driving operation data over a period of time; based on the vehicle operating status data, performing structural response prediction using a trained structural response prediction model, and outputting a vehicle status prediction result for a future period of time; determining whether the vehicle status prediction result meets a preset first roll risk detection condition; in response to meeting the first roll risk detection condition, indicating that the current vehicle has a certain roll risk, then determining the vehicle's roll risk data based on the vehicle operating status data, driving operation data, and vehicle status prediction result. The method of this application can not only improve the accuracy of chronic roll risk prediction and reduce the roll misjudgment rate, but also determine roll risk data to identify the roll risk level, helping the vehicle to perform corresponding safety intervention operations and improving driving safety.

Description

Technical Field

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

[0002] Chronic body roll during vehicle operation is common in mountainous areas with continuous curves, long and gentle slopes, and crosswind disturbances. Existing vehicle body roll detection methods cannot effectively identify chronic body roll phenomena during vehicle operation, resulting in a high rate of false positives and untimely warnings. 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 high misjudgment rate of chronic vehicle roll.

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

[0005] Acquire vehicle operating status data and driving operation data over a specified period of time; Based on the vehicle operating status data, structural response prediction is performed using the trained structural response prediction model, and the vehicle status prediction result for a future period of time is output. Determine whether the vehicle state prediction result meets the preset first roll risk detection condition. In response to meeting the first roll risk detection condition, determine the vehicle roll risk data based on the vehicle operating state data, the driving operation data, and the vehicle state prediction result.

[0006] Optionally, determining the vehicle's roll risk data based on the vehicle operating status data, the driving operation data, and the vehicle status prediction results includes: Based on the vehicle operating status data, the driving operation data, and the vehicle status prediction results, attitude prediction is performed using a trained roll evolution prediction model, and the vehicle attitude prediction results for a future period of time are output. It is then determined whether the vehicle attitude prediction results meet the preset second roll risk monitoring conditions. In response to meeting the second roll risk detection conditions, the roll risk data is determined based on the vehicle attitude prediction results and the vehicle status prediction results.

[0007] Optionally, determining the roll risk data based on the vehicle attitude prediction result and the vehicle state prediction result includes: Based on the vehicle attitude prediction results and the vehicle state prediction results, topological data for characterizing vehicle roll risk is determined using a topological data analysis method; it is determined whether the topological data meets the preset third roll risk monitoring conditions, and in response to meeting the third roll risk detection conditions, the roll risk data is determined.

[0008] Optionally, the step of predicting the structural response based on the vehicle operating state data using a trained structural response prediction model and outputting a vehicle state prediction result for a future period includes: The feature extraction network in the structural response prediction model extracts features from the vehicle operating state data to obtain a hidden state vector; the fully connected layer network in the structural response prediction model maps the hidden state vector to the vehicle state prediction result.

[0009] Optionally, the training method for the structural response prediction model includes: Vehicle operating status data for different vehicle models is collected. Multiple training tasks are constructed based on this data, with each training task corresponding to a specific vehicle model and including multiple training samples. An initial structural response prediction model is built, comprising an initial feature extraction network and an initial fully connected layer network. For each training task, the following operations are performed: The initial fully connected layer network is updated with gradients based on a portion of the training samples in the training task to obtain an updated initial structural response prediction model; the updated initial structural response prediction model is trained based on multiple training tasks to update the network parameters of the initial feature extraction network in reverse; after training is completed, the structural response prediction model is obtained.

[0010] Optionally, the vehicle attitude prediction results include predicted roll angle and predicted roll rate. The step of predicting vehicle attitude based on the vehicle operating status data, driving operation data, and vehicle state prediction results, using a trained roll evolution prediction model, and outputting a vehicle attitude prediction result for a future period includes: encoding the vehicle operating status data, driving operation data, and vehicle state prediction results through the encoding layer of the roll evolution prediction model to obtain a joint state vector; fitting the attitude evolution law through the state fitting network of the roll evolution prediction model based on the joint state vector to output an initial roll angle prediction value for a future period; calculating the roll angular velocity prediction value for a future period through automatic differentiation based on the initial roll angle prediction value; and smoothing and boundary constraint processing the initial roll angle prediction value through the post-processing layer of the roll evolution prediction model to obtain the final roll angle prediction value.

[0011] Optionally, the training method for the roll evolution prediction model includes: constructing a training dataset, which includes training samples and ground truth labels; constructing an initial roll evolution prediction model; inputting the training samples into the initial roll evolution prediction model, and outputting a continuous roll angle state function through an initial state fitting network in the initial roll evolution prediction model; calculating the roll angular velocity based on the continuous roll angle state function using automatic differentiation; constructing a data fitting loss function based on the continuous roll angle state function and the ground truth labels; constructing a physical constraint loss function based on the roll angular velocity and the vehicle lateral dynamics equations; constructing a temporal smoothing constraint loss function based on the continuous roll angle state function; weighting and fusing the data fitting loss function, the physical constraint loss function, and the temporal smoothing constraint loss function to obtain a joint loss function; and iteratively training the initial roll evolution prediction model by minimizing the joint loss function to obtain the roll evolution prediction model.

[0012] Optionally, the method further includes: During the training process of the tilt evolution prediction model, the weight of the data fitting loss function in the joint loss function gradually decreases, while the weight of the physical constraint loss function in the joint loss function gradually increases.

[0013] Optionally, before performing structural response prediction based on the vehicle operating state data using a trained structural response prediction model and outputting the vehicle state prediction result for a future period, the method includes: Obtain vehicle model information; in response to determining that the model information is unknown, obtain training samples corresponding to the model information; perform gradient updates on the fully connected layer network in the structural response prediction model based on the training samples to obtain the updated structural response prediction model.

[0014] Based on the same inventive concept, a second aspect of this application also provides a vehicle, said vehicle comprising: A memory for storing 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 the first aspect.

[0015] As described above, the vehicle risk prediction method and vehicle provided in this application include: acquiring vehicle operating status data and driving operation data over a period of time; based on the vehicle operating status data, performing structural response prediction using a trained structural response prediction model, and outputting a vehicle status prediction result for a future period of time. The structural response prediction model can capture the cumulative changes in vehicle status due to chronic roll over a period of time, and then obtain a vehicle status prediction result for a future time period through reasonable prediction, providing a reliable data foundation for predicting vehicle roll risk. It is determined whether the vehicle status prediction result meets a preset first roll risk detection condition. If the first roll risk detection condition is met, indicating that the current vehicle has a certain roll risk, then based on the vehicle operating status data, the driving operation data, and the vehicle status prediction result, the vehicle's roll risk data is determined. The method of this application can not only improve the accuracy of chronic roll risk prediction and reduce the roll misjudgment rate, but also determine roll risk data to identify the roll risk level, helping the vehicle to perform corresponding safety intervention operations and improve driving safety. Attached Figure Description

[0016] 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.

[0017] Figure 1 This is a flowchart illustrating the vehicle risk prediction method according to an embodiment of this application; Figure 2 This is a schematic diagram of the vehicle risk prediction device according to an embodiment of this application; Figure 3 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0018] 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.

[0019] 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.

[0020] As described in the background section, most existing methods for determining vehicle roll stability and providing risk warnings rely on the amplitude thresholds of dynamic signals such as instantaneous vehicle attitude angles and lateral acceleration for simple discrimination. These methods can only achieve limited identification of acute and severe roll danger scenarios such as sudden changes in roll angle and sharp turns. They are difficult to effectively capture and characterize the structural stress accumulation, energy buildup, and gradual instability risks caused by long-term, low-amplitude, and continuous chronic roll.

[0021] This type of chronic roll often occurs in typical operating conditions such as continuous curves in mountainous areas, long gentle slopes, continuous crosswind disturbances, and asymmetric load driving. Its characteristics include small individual roll amplitudes, gradual changes, and weak short-term impacts, making it difficult to classify as a dangerous state using traditional threshold-based methods. However, because a real vehicle body is not an ideal rigid body, it exhibits nonlinear physical characteristics under continuous weak excitation, causing the roll risk to accumulate and slowly evolve over time. As driving time increases, local structural stress continuously rises, the suspension dynamic response gradually becomes unbalanced, and the body stiffness characteristics undergo implicit attenuation. Ultimately, without any obvious violent operation, it can suddenly trigger critical rollover, attitude instability, structural fatigue damage, and other severe consequences.

[0022] Existing technical solutions generally have the following limitations: First, they only focus on instantaneous dynamic states, ignoring the long-term evolution of structural micro-responses (such as body flexible deformation, suspension elastic hysteresis, bushing creep, structural stress accumulation, and gradual accumulation of lateral inertia and potential energy). They do not incorporate structural signals such as body stress, suspension micro-deformation, and local stiffness changes into risk modeling, and therefore cannot reflect the cumulative nature of chronic roll. Second, most are purely data-driven models, lacking physical constraints on vehicle dynamics. The prediction results are prone to false trends that do not conform to the laws of mechanics, making it difficult to achieve reliable predictions over long periods and through extrapolation. Third, they do not characterize energy accumulation and state topology changes at the system level, and cannot identify early critical transitions where "the amplitude is normal but the system has become unstable," resulting in significant warning lag. Fourth, the models rely on specific vehicle calibration, have weak cross-platform and cross-structural generalization capabilities, and are difficult to quickly transfer and reuse between vehicles with different body stiffness, suspension types, and curb weights.

[0023] In summary, this application proposes a vehicle risk prediction method that combines vehicle structural response prediction, physical constraint evolution, and topological mutation identification to predict roll risk. This method not only improves the accuracy of predicting chronic roll risk but also provides timely warnings of roll risk, thereby enhancing driving safety.

[0024] The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0025] This application provides a vehicle risk prediction method, referencing... Figure 1 This includes the following steps: Step 102: Obtain vehicle operating status data and driving operation data over a certain period of time.

[0026] Specifically, vehicle operating status data includes dynamic behavior variables collected by the onboard inertial measurement unit, such as roll angle, roll rate, lateral acceleration, vehicle speed, steering wheel angle, and longitudinal acceleration. These dynamic behavior variables capture the vehicle's overall attitude and motion, reflecting the dynamic process of roll. Vehicle operating status data also includes microstructural response data acquired by body structure monitoring sensors such as strain gauges and accelerometers, including local frame deformation, body stress distribution, suspension compression displacement, and vehicle pitch angle. Microstructural response data captures minute changes at the vehicle structural level, reflecting the cumulative effect of roll. Driving operation data includes steering wheel angle rate and acceleration / deceleration data.

[0027] To construct a model capable of accurately identifying the risk of chronic roll accumulation, the aforementioned multi-source data first requires high-frequency acquisition and rigorous time-series alignment preprocessing. All sensor data must be acquired at millisecond-level sampling frequencies to ensure sufficient time resolution and capture minute dynamic changes during the chronic accumulation process. In the data preprocessing stage, the multi-source heterogeneous signals are sequentially synchronized through channels, their sampling frequencies are unified, and their timestamps are strictly aligned to eliminate data drift and transmission delay errors. All signals are then uniformly mapped to the vehicle's main timeline, forming a structured multi-channel time-series dataset.

[0028] When constructing the training dataset for subsequent models, based on the aforementioned high-frequency acquisition and time-aligned preprocessing, a sliding window strategy is employed to prune and augment the time-series data, improving the generalization ability of subsequent models in complex driving scenarios. Further data normalization and outlier detection are performed, setting dynamic thresholds for structural signals such as stress and deformation to eliminate pseudo-data generated by sensor malfunctions or external interference. The acquisition and preprocessing process relies on the automaker's self-built data closed-loop system, achieving synchronous aggregation and unified management of multi-module data through a local area network bus and Ethernet interface. After the above processing, the final output is a high-efficiency, high-quality time-series sample data pool covering vehicle attitude dynamics and microstructure response, providing a unified input format for subsequent multimodal data fusion, model training, and joint inference.

[0029] Step 104: Based on the vehicle operating status data, perform structural response prediction using the trained structural response prediction model, and output the vehicle status prediction result for a future period of time.

[0030] Specifically, the structural response prediction model incorporates a neural network model based on the Model-Agnostic Meta-Learning (MAML) algorithm. This model enables the modeling and prediction of structural response change patterns induced by chronic roll under different structural parameters and platform architectures. The model structure includes a feature extraction network and a fully connected layer network. The feature extraction network extracts a hidden state vector that integrates short-term disturbances and long-term accumulated features. The fully connected layer network maps this hidden state vector to the predicted vehicle state. The predicted vehicle state includes predicted stress changes, predicted suspension displacement, and predicted local stiffness.

[0031] Step 106: Determine whether the vehicle state prediction result meets the preset first roll risk detection condition. In response to meeting the first roll risk detection condition, determine the vehicle roll risk data based on the vehicle operating state data, the driving operation data, and the vehicle state prediction result.

[0032] Specifically, by determining whether the vehicle state prediction results meet the preset first roll risk detection conditions, it is possible to judge whether there is abnormal accumulation in the body, suspension, and stiffness, and to promptly detect minor abnormalities during vehicle operation. The first roll risk detection conditions define the risk boundary conditions of the vehicle state prediction results. These conditions include: the difference in compression displacement between the left and right suspensions continuously exceeding a preset threshold, and the roll angle change rate exceeding a preset range for multiple consecutive time windows; or, the cumulative increase in stress at key body parts exceeding a safety threshold dynamically calculated based on vehicle speed and load. Key body parts may include the frame, suspension supports, body longitudinal / crossbeams, wheel wells, and subframe. The safety threshold is negatively correlated with vehicle speed; the faster the speed, the lower the safety threshold. The safety threshold is also negatively correlated with load; the heavier the load, the lower the safety threshold.

[0033] If the vehicle status prediction result meets the first roll risk detection condition, it indicates that the vehicle currently has an abnormal structural response. Further analysis based on vehicle operating status data, driving operation data, and the vehicle status prediction result is needed to accurately determine the vehicle's roll risk data. Roll risk data includes the roll risk level, roll risk trigger time window, and system stability status label. Roll risk data can provide reliable, proactive, and interpretable decision-making basis for vehicle roll safety control.

[0034] If the vehicle status prediction result does not meet the first roll risk detection condition, it indicates that the vehicle currently does not have any structural response anomalies. Real-time collection of vehicle operating status data can continue to monitor whether the vehicle status prediction result meets the first roll risk detection condition. At this time, the roll risk level in the roll risk data is safe, the roll risk trigger time window is none, and the system stability status label is stable.

[0035] Based on steps 102 to 106 above, the vehicle risk prediction method provided in this embodiment includes: acquiring vehicle operating status data and driving operation data over a period of time. Based on the vehicle operating status data, structural response prediction is performed using a trained structural response prediction model, and the vehicle status prediction result for a future period of time is output. The structural response prediction model can capture the cumulative changes in vehicle status due to chronic roll over a period of time, and then obtain the vehicle status prediction result for the future time period through reasonable prediction, providing a reliable data foundation for predicting vehicle roll risk. It is determined whether the vehicle status prediction result meets a preset first roll risk detection condition. In response to meeting the first roll risk detection condition, indicating that the current vehicle has a certain roll risk, the vehicle roll risk data is determined based on the vehicle operating status data, the driving operation data, and the vehicle status prediction result. The method of this application can not only improve the accuracy of chronic roll risk prediction and reduce the roll misjudgment rate, but also determine roll risk data to identify the roll risk level, help the vehicle perform corresponding safety intervention operations, and improve driving safety.

[0036] In some embodiments, the step of predicting the structural response based on the vehicle operating state data using a trained structural response prediction model and outputting a vehicle state prediction result for a future period includes: The feature extraction network in the structural response prediction model extracts features from the vehicle operating state data to obtain a hidden state vector; the fully connected layer network in the structural response prediction model maps the hidden state vector to the vehicle state prediction result.

[0037] Specifically, the vehicle operating state data input to the structural response prediction model is a preprocessed standard time series. The structural response prediction model includes a feature extraction network and a fully connected layer network. For example, the feature extraction network structure, in order of data propagation direction, includes a one-dimensional convolutional layer, a batch normalization and nonlinear activation layer, and a Bi-GRU (Bidirectional Gated Recurrent Unit) / Bi-LSTM (Bidirectional Long Short-Term Memory) encoding layer. The one-dimensional convolutional layer captures "short-term micro-perturbation features" in the input time series, such as small changes in roll rate, lateral acceleration fluctuations, and minor adjustments in steering wheel angle; these are weak signals in the early stages of chronic roll. Multiple one-dimensional convolutional layers can be set, with each layer outputting a "local feature map" for the corresponding channel, preserving the temporal dimension information of the time series data while extracting the local correlations between different features. Batch normalization normalizes the local feature maps output by the convolutional layers, eliminating data scale differences between different vehicle models and driving conditions, making the feature distribution more consistent, ensuring the universality of shared features, and accelerating the computational efficiency of subsequent temporal coding while avoiding gradient vanishing. Nonlinear activation uses nonlinear activation functions such as ReLU (Rectified Linear Unit) and Tanh (Hyperbolic Tangent) to perform nonlinear mapping on the normalized features, breaking the linear relationship and enhancing the model's ability to fit the "weakly changing, strongly nonlinear" features of chronic roll, extracting more discriminative universal features. Bi-GRU / Bi-LSTM can capture "long-term cumulative features across time windows" during chronic roll (such as slow stress accumulation and continuous suspension displacement, features that cannot be captured by a single short-term window). The final output data of the feature extraction network is a high-dimensional hidden state vector that integrates short-term disturbances and long-term cumulative features.

[0038] The hidden state vector is input into a fully connected layer network, which maps it to the vehicle state prediction result. The vehicle state prediction result includes predicted stress changes, predicted suspension displacement, and predicted local stiffness. For example, the fully connected layer network consists of 2 to 3 fully connected layers. The structural response prediction model in this embodiment can predict the vehicle structural response over a future period, obtaining the vehicle state prediction result. This provides important intermediate variables for subsequent models and a stable and physically consistent dynamic response basis for roll risk prediction.

[0039] In some embodiments, the training method for the structural response prediction model includes: Collect vehicle operation status data for different vehicle models, and construct multiple training tasks based on the vehicle operation status data for different vehicle models. Each training task corresponds to a vehicle model and includes multiple training samples. An initial structural response prediction model is constructed, which includes an initial feature extraction network and an initial fully connected layer network. For each training task, the following operations are performed: the initial fully connected layer network is updated with gradients based on a portion of the training samples in the training task to obtain an updated initial prediction model; the updated initial prediction model is trained based on multiple training tasks, and the network parameters of the initial feature extraction network are updated in reverse; after training is completed, the structural response prediction model is obtained.

[0040] Specifically, multiple training tasks are constructed using data from vehicles of different models or with different structural configurations, with each training task corresponding one-to-one with a vehicle model. That is, each training task corresponds to a specific combination of vehicle structural parameters, including body stiffness, suspension type, and curb weight. Each training task includes multiple training samples. Each training sample includes time-series data after preprocessing historical vehicle operating state data, as well as vehicle state prediction labels, including stress change labels, suspension displacement labels, and local stiffness labels. Constructing multiple training tasks allows the model to learn the general patterns of different vehicle models, improving the model's generalization ability.

[0041] An initial structural response prediction model is constructed, which includes an initial feature extraction network and an initial fully connected layer network. The initial feature extraction network has the same network structure as the feature extraction network in the previous embodiment, and the initial fully connected layer network has the same network structure as the fully connected layer network in the previous embodiment, which will not be described again here.

[0042] This step employs a model-independent meta-learning algorithm for model training. First, the inner loop of the model-independent meta-learning algorithm updates the gradient of the initial fully connected layer network using a small number of training samples from each training task, while keeping the initial feature extraction network parameters unchanged. This yields network parameters suitable for the current vehicle model. In this way, the model can quickly learn the roll response of the current vehicle model using a small amount of continuation data, achieving personalized adaptation. Then, the outer loop of the model-independent meta-learning algorithm trains the updated initial prediction model on multiple training tasks. Training samples are input into the updated initial prediction model for prediction, the overall prediction error is calculated, and the network parameters of the initial feature extraction network are updated in reverse, allowing the model to learn common features across different vehicle models. The loss function used during training employs a mean squared error loss function and a trend consistency constraint. The mean squared error loss function minimizes the difference between the vehicle state prediction label and the output prediction result of the initial structural response prediction model. The trend consistency constraint ensures that the model's prediction results conform to physical laws; for example, if roll increases slowly, stress should also increase slowly; if the roll angle increases, the corresponding suspension compression should also increase accordingly. By constructing this loss function, the fitting effect on the chronic tilt accumulation process is enhanced.

[0043] Finally, after multiple rounds of training and iterative optimization, a structural response prediction model with cross-vehicle generalization ability is obtained. The model training method in this embodiment enables the prediction of structural response patterns induced by chronic roll under different vehicle models and structural parameters. During training, only the initial fully connected layer network is updated in an inner loop, while the initial feature extraction network is updated slowly, thus achieving a decoupled structure of "general features + individual adaptation".

[0044] In some embodiments, before performing structural response prediction based on the vehicle operating state data using a trained structural response prediction model and outputting the vehicle state prediction result for a future period, the method includes: Obtain vehicle model information; in response to determining that the model information is unknown, obtain training samples corresponding to the model information; perform gradient updates on the fully connected layer network in the structural response prediction model based on the training samples to obtain the updated structural response prediction model.

[0045] Specifically, the vehicle model information is obtained. If the model information is unknown, it indicates that the current vehicle is a new model or includes new structural parameters. In this case, a small number of training samples for that model can be used to update the gradient of the fully connected layer network in the structural response prediction small model. This allows for fine-tuning of the fully connected layer network for the new model, achieving rapid fitting of the structural response characteristics of the new model. The fine-tuning process does not require retraining the entire structural response prediction model, achieving rapid adaptation to small samples of the new model and enabling rapid model personalization. The updated structural response prediction model can then be executed online for prediction, improving the adaptation efficiency of the structural response prediction model. The inference process of the structural response prediction model supports sliding window updates and asynchronous computation, and can be deployed on the vehicle or in the cloud to achieve multi-model adaptation and continuous vehicle state prediction.

[0046] In some embodiments, determining the vehicle's roll risk data based on the vehicle operating status data, the driving operation data, and the vehicle status prediction results includes: Based on the vehicle operating status data, the driving operation data, and the vehicle status prediction results, attitude prediction is performed using a trained roll evolution prediction model, and the vehicle attitude prediction results for a future period of time are output. It is then determined whether the vehicle attitude prediction results meet the preset second roll risk monitoring conditions. In response to meeting the second roll risk detection conditions, the roll risk data is determined based on the vehicle attitude prediction results and the vehicle status prediction results.

[0047] Specifically, in this embodiment, the roll evolution prediction model can be a Physics-Informed Neural Network (PINN). A PINN is a deep learning method that directly embeds physical laws (usually in the form of partial differential equations or ordinary differential equations) into the neural network training process. It is not purely data-driven, but rather integrates prior physical knowledge with observational data to ensure that the model's predictions strictly adhere to known physical laws. Based on vehicle operating status data, driving operation data, and vehicle state prediction results, the roll evolution prediction model can output vehicle attitude predictions for a future time period. This model accurately models the long-term, weakly changing, and gradually accumulating process of chronic roll. The vehicle attitude prediction results include predicted roll angle and roll rate values ​​for a future time period. The predicted roll angle values ​​represent the continuous evolution trajectory of the roll angle, forming a prediction curve showing the continuous change of vehicle attitude over time based on the vehicle attitude prediction results.

[0048] Next, it is determined whether the vehicle attitude prediction results meet the preset second roll risk monitoring conditions. These conditions define the risk boundary conditions for the vehicle attitude prediction results, further determining whether the roll is slowly increasing and becoming increasingly dangerous. If the second roll risk monitoring conditions are met, it indicates that the vehicle roll evolution has transitioned from a slow accumulation phase to a rapid deterioration phase. The second roll risk detection conditions include: the vehicle roll angle monotonically increases over multiple consecutive sliding time windows, and the increase exceeds the preset safety range upper limit. At this point, it is necessary to further determine the roll risk data based on the vehicle attitude prediction results and the vehicle state prediction results. If the vehicle attitude prediction results do not meet the second roll risk monitoring conditions, the corresponding roll risk data will have a roll risk level of safety 1 (warning level), the roll risk trigger time window will be continuously observed, and the system stability status label will be slightly abnormal.

[0049] It should be noted that during normal vehicle operation, the structural response prediction model continuously outputs vehicle state prediction results. To better monitor vehicle roll response, the roll evolution prediction model can be run simultaneously at a low frequency, meaning the roll evolution prediction model outputs vehicle attitude prediction results at a lower frequency. If it is determined that the vehicle state prediction results meet the preset first roll risk detection condition, the running frequency of the roll evolution prediction model is increased, meaning the roll evolution prediction model outputs vehicle attitude prediction results at a higher frequency to promptly detect changes in vehicle roll.

[0050] The roll evolution prediction model in this embodiment can predict how the vehicle roll angle will continuously change over a future period, determining whether the vehicle roll is slowly deteriorating and becoming increasingly dangerous, thus achieving a more reliable assessment of roll prediction. If it is determined that the vehicle roll trend is intensifying, the roll risk data is determined by combining the vehicle attitude prediction results and the vehicle state prediction results, in order to achieve an accurate prediction of vehicle roll risk.

[0051] In some embodiments, the vehicle attitude prediction results include predicted roll angle and predicted roll rate; The method involves using a trained roll evolution prediction model to predict vehicle attitude based on the vehicle operating status data, driving operation data, and vehicle status prediction results, and outputting vehicle attitude prediction results for a future period, including: The vehicle operating state data, driving operation data, and vehicle state prediction results are encoded through the encoding layer in the roll evolution prediction model to obtain a joint state vector. Based on the joint state vector, the attitude evolution law is fitted through the state fitting network in the roll evolution prediction model to output the initial roll angle prediction value for a future period. According to the initial roll angle prediction value, the roll angular velocity prediction value for a future period is obtained through automatic differentiation calculation. The initial roll angle prediction value is smoothed and boundary constraint processed through the post-processing layer in the roll evolution prediction model to obtain the roll angle prediction value.

[0052] Specifically, the roll evolution prediction model includes an encoding layer, a state fitting network, and a post-processing layer. The encoding layer performs a unified dimensional mapping on multi-channel heterogeneous time-series data, linearly embedding vehicle operating state data, driving operation data, and vehicle state prediction results, followed by feature alignment and concatenation to form a high-dimensional joint state vector. Based on this joint state vector, the state fitting network learns the nonlinear temporal evolution law of the roll state. For example, the state fitting network structure includes multiple fully connected layers, gated recurrent units, and a regression output head. The multiple fully connected layers complete the nonlinear mapping of the vehicle's dynamic state space, the gated recurrent units capture the continuous dependence and cumulative effect of chronic roll over long time scales, and the regression output head outputs the predicted roll angle value within future time windows. Since the roll evolution prediction model has embedded dynamic constraints during the training phase, its output naturally satisfies the laws of vehicle dynamics. The roll rate is obtained through automatic differentiation and used for subsequent risk assessment and state determination. Subsequently, the initial roll angle prediction value is smoothed and subjected to boundary constraints through the post-processing layer in the roll evolution prediction model: the post-processing layer filters the predicted trajectory to eliminate high-frequency jitter, and the boundary constraints limit the roll angle to a reasonable physical range to avoid abnormal fluctuations, ultimately yielding the predicted roll angle value. Simultaneously, the confidence level of the predicted roll angle value is evaluated based on historical prediction residuals, outputting the prediction confidence level for the current roll angle prediction value.

[0053] The vehicle attitude prediction results output by the roll evolution prediction model in this embodiment can be used to determine whether chronic roll has accumulated and intensified, changed its trend, or is approaching an instability threshold, thus achieving early warning of vehicle roll and improving vehicle driving safety.

[0054] In some embodiments, the training method of the tilt evolution prediction model includes: Construct a training dataset, which includes training samples and ground truth labels; Construct an initial tilt evolution prediction model; The training samples are input into the initial roll evolution prediction model, and the roll angle continuous state function is output through the initial state fitting network in the initial roll evolution prediction model. Based on the continuous state function of the roll angle, the roll angular velocity is obtained by automatic differentiation calculation; Based on the roll angle continuous state function and the truth label, a data fitting loss function is constructed; Based on the roll rate and the vehicle lateral dynamics equations, a physical constraint loss function is constructed; Based on the roll angle continuous state function, a time-series smoothing constraint loss function is constructed; The data fitting loss function, the physical constraint loss function, and the temporal smoothing constraint loss function are weighted and fused to obtain a joint loss function; The initial tilt evolution prediction model is iteratively trained by minimizing the joint loss function to obtain the tilt evolution prediction model.

[0055] Specifically, a training dataset is constructed by jointly using real-vehicle data and multi-condition simulation data, covering typical scenarios of chronic roll such as long curves, gentle slopes, variable loads, and crosswinds. The training data includes training samples, which consist of time-series feature data constructed from vehicle operating status data and driving operation data, as well as vehicle state prediction results output by the structural response prediction model. The ground truth label is the measured value of the vehicle roll angle within the future prediction time.

[0056] An initial roll evolution prediction model is constructed. The model structure of the initial roll evolution prediction model is the same as that of the roll evolution prediction model in the previous embodiment, and will not be repeated here. Training samples are input into the initial roll evolution prediction model. The initial state fitting network in the initial roll evolution prediction model outputs a continuous roll angle state function. This continuous roll angle state function determines the continuous roll angle values ​​at various future time points. Based on the continuous roll angle state function, the first derivative of the roll angle, i.e., the roll angular velocity, is calculated using automatic differentiation. (Continuous roll angle state function) The specific format is as follows: (1), Wherein, NN represents the initial tilt evolution prediction model. This represents vehicle operating status data and driving operation data. This indicates the predicted vehicle status.

[0057] During training, a joint loss function is used to iteratively train the initial roll evolution prediction model. The joint loss function is constructed from a data fitting loss function, a physical constraint loss function, and a temporal smoothing constraint loss function. Among them, the data fitting loss function is constructed based on the roll angle continuous state function and the ground truth label, and its goal is to minimize the difference between the roll angle obtained from the roll angle continuous state function and the ground truth label, so that the roll angle trajectory output by the model is consistent with the actual label in numerical terms.

[0058] The physical constraint loss function is constructed based on the roll rate and the vehicle's lateral dynamics equations. Based on these equations, it establishes physical residual terms such as lateral force balance, roll moment balance, and energy changes to force the model output to satisfy the vehicle's kinematic and mechanical conservation laws. The temporal smoothing constraint loss function is constructed based on the continuous roll angle state function. It applies a smoothing regularization to the temporal variation of the roll angle, suppressing high-frequency jitter and ensuring continuous and stable roll evolution. Data fitting loss function. The specific format is as follows: (2), in, This represents the data fitting loss function. Represents the physical constraint loss function. This represents the time-series smoothing constraint loss function. This represents the weights corresponding to each loss function.

[0059] Finally, by minimizing the joint loss function, the initial roll evolution prediction model is iteratively trained. Combined with an early stop strategy and cross-validation, a roll evolution prediction model is obtained. During training, the vehicle body structural characteristics and the vehicle's lateral dynamics are deeply integrated, outputting a continuous, smooth, and physically reliable roll angle trajectory, providing high-quality, high-dimensional continuous input for subsequent topological change identification. The training method directly embeds the vehicle's lateral dynamics equations into the loss function and training process, forcing the model to satisfy fundamental physical laws such as force balance, torque conservation, and energy change while fitting the data. This fundamentally avoids invalid predictions that are "numerically accurate but physically impossible," solving the problems of weak signals being easily covered by noise and the difficulty in fitting long-term cumulative trends in chronic roll scenarios.

[0060] In some embodiments, the method further includes: During the training process of the tilt evolution prediction model, the weight of the data fitting loss function in the joint loss function gradually decreases, while the weight of the physical constraint loss function in the joint loss function gradually increases.

[0061] Specifically, during the training process of the tilt evolution prediction model, the formula (2) in the aforementioned embodiment... and It is dynamically changing. In the early stages of training, the weight of the data fitting loss function in the joint loss function is increased, that is, the weight of the loss function is increased. The value of this value reduces the weight of the physical constraint loss function in the joint loss function, which means reducing... The numerical value is used to enable the model to quickly learn the basic temporal features of tilt dynamics, achieving rapid convergence. In the later stages of training, the weight of the data fitting loss function in the joint loss function is reduced, i.e., the weight of the loss function is decreased. The value of this value increases the weight of the physical constraint loss function in the joint loss function, that is, it increases the weight of the physical constraint loss function. The numerical values ​​are gradually increased by simultaneously strengthening the physical constraint loss weights and smoothing constraints, allowing the model to conform to the basic laws of vehicle dynamics on top of the fundamental features, locking in the long-term cumulative trend of chronic roll, and avoiding overfitting and prediction divergence.

[0062] By dynamically adjusting the weights of the loss term in this embodiment, the inherent contradiction between traditional physical information neural network training and purely data-driven models is resolved: high physical constraints throughout the process lead to model convergence difficulties, gradient vanishing, and failure to learn weak signals of chronic lateral tilt; high data weights throughout the process lead to black-box overfitting, poor physical consistency, long-term prediction divergence, and weak anti-interference ability. The training method in this embodiment can significantly improve the model convergence speed and training stability, solving the problem of balancing training convergence and physical consistency.

[0063] In some embodiments, determining the roll risk data based on the vehicle attitude prediction result and the vehicle state prediction result includes: Based on the vehicle attitude prediction results and the vehicle state prediction results, topological data for characterizing vehicle roll risk is determined using a topological data analysis method; it is determined whether the topological data meets the preset third roll risk monitoring conditions, and in response to meeting the third roll risk detection conditions, the roll risk data is determined.

[0064] Specifically, when it is determined that the vehicle roll risk is increasing, further roll risk prediction is needed. In this embodiment, based on the vehicle attitude prediction results and vehicle state prediction results, a high-dimensional topological model of the energy evolution path throughout the chronic roll process is performed using Topological Data Analysis (TDA). Early abrupt change points are identified at the system structure level, and topological data used to characterize vehicle roll risk is determined, enabling advanced capture and early warning of potential risk turning points. The topological data may include topological anomaly data, which can specifically include quantifiable topological anomalies such as changes in homology group structure, jumps in Betti number, and drastic reconstruction of persistence maps. Among them, changes in homology group structure refer to the gradual splitting of the state point cloud, composed of multidimensional variables such as roll angle, structural stress, suspension displacement, and system energy, from an originally connected and compact geometric shape in the topological space into multiple discontinuous subclusters or forming obvious void structures, reflecting the transformation of vehicle roll behavior from a stable and coordinated state to a locally unstable state. Betti number mutations are used to quantify the aforementioned topological change process. Betti0 represents the number of system state clusters; when the vehicle enters the abnormal roll phase, this value abruptly changes from a single cluster to multiple clusters, indicating that the responses of various vehicle subsystems begin to separate. Betti1 represents the number of ring structures in the system; its rapid increase usually corresponds to the formation of repetitive oscillating paths of energy in the suspension-body system, a significant precursor to chronic roll accumulation. Drastic reconstruction of the persistent topological pattern manifests as the rapid disappearance of previously persistent topological features, while a large number of new features with short lifecycles suddenly appear, indicating that the vehicle dynamics system has entered a highly unstable region.

[0065] In practical implementation, based on the vehicle attitude prediction results, the values ​​of variables such as the vehicle state prediction results at the same moment are combined and defined as a high-dimensional state vector. Combining all state vectors over a continuous time period forms a high-dimensional state point cloud. This point cloud fully depicts the state distribution and structural characteristics of the vehicle roll dynamic system throughout its evolution. To extract geometric structural features from the point cloud, Rips complex is used for topological modeling of the state point cloud: expanding stepwise with distance thresholds at different scales, connecting closely spaced state points into edges, triangles, and higher-dimensional units, forming a topological spatial structure that is nested layer by layer from fine to coarse. Through this multi-scale nested structure, the geometric shape of the system at both the level of subtle fluctuations and the overall trend can be captured simultaneously, avoiding feature loss caused by a single scale. Based on the Rips complex, the persistence cohomology of the point cloud is calculated, that is, the appearance and disappearance process of topological structures at different scales is analyzed, and the results are represented as a persistence diagram. The persistence diagram can objectively reflect the life cycle of various topological structures in the system: structures with long life cycles represent the stable, essential, and unchanging core characteristics of the system; structures with short life cycles are mostly noise or temporary disturbances. Quantifiable topological invariants are extracted from the persistent graph, primarily including: Betti numbers (Betti0, Betti1), used to describe the connectivity, number of holes, and number of closed loops in the point cloud; and the lifetime of topological features, used to measure the stability of the structure. These topological invariants do not change with coordinate transformations, translations, or scaling, and can stably reflect the intrinsic structural characteristics of the system, rather than the magnitude of surface signals. During vehicle operation, the system continuously monitors abnormal behaviors such as abrupt changes, jumps, and reconstructions of the above topological invariants: when the Betti number suddenly jumps, the persistent graph undergoes drastic reconstruction, a large number of long-period features disappear, and short-period features surge, it indicates that the internal structure of the system has undergone fundamental changes. Even if intuitive values ​​such as roll angle and acceleration are still within safe ranges, it can be determined that the vehicle has entered the critical instability zone.

[0066] The method in this embodiment can assist in identifying "pre-rollover" states, i.e., the stage where the system has not experienced severe disturbances but the topology has undergone drastic changes, thereby establishing a highly sensitive risk identification mechanism. The identified topology data serves as a strong trigger signal for the roll risk model, which can be used to drive control strategy adjustments, activate safety strategy warnings, and provide feedback to adjust the model parameter update path. The method in this embodiment does not require any prior assumptions about the data distribution and is unaffected by characteristics such as nonlinearity, weak signals, slow time-varying, and long accumulation. It has a natural advantage for dynamic systems like chronic roll, which exhibit slow quantitative changes, strong concealment, and no severe early signs. It can accurately capture early risk moments when "quantitative changes have not yet manifested, but qualitative changes have already occurred in the structure," providing key technical support for early warning of chronic roll.

[0067] The system determines whether the topological data meets the preset third roll risk monitoring conditions. If so, it identifies the vehicle as entering the risk state transition zone, i.e., the critical state transitioning from stability to rollover / structural instability. The third roll risk monitoring conditions define the risk boundary conditions of the topological data, further determining whether the vehicle is about to enter the risk state transition zone, characterizing the precursors to rollover. Upon entering the risk state transition zone, the system increases the prediction frequency and initiates risk warning or control strategies, providing an early intervention window for chronic rollover and ensuring driving safety. The third roll risk monitoring conditions include topological anomalies in the topological data, i.e., quantifiable topological anomalies such as changes in homology group structure, jumps in Betti numbers, and drastic reconstruction of persistence maps. At this point, the roll risk level in the roll risk data is level 3 (highly dangerous), the roll risk trigger time window is a short-term deterioration window, and the system stability state label is critical instability. If the topology data does not meet the preset third tilt risk monitoring conditions, the tilt risk level in the tilt risk data is level 2, the tilt risk trigger time window is predicted to be a few seconds to tens of seconds in the future, and the system stability status label is unstable.

[0068] In summary, based on the step-by-step assessment of vehicle roll risk in this application, when the vehicle state prediction result does not meet the preset first roll risk detection condition, the vehicle roll risk is level 0, and the system stability status is labeled as stable; when the vehicle state prediction result meets the preset first roll risk detection condition, and the vehicle attitude prediction result does not meet the preset second roll risk monitoring condition, the vehicle roll risk is level 1, and the system stability status is labeled as slightly abnormal; when the vehicle attitude prediction result meets the preset second roll risk monitoring condition, and the topology data does not meet the preset third roll risk monitoring condition, the vehicle roll risk is level 2, and the system stability status is labeled as unstable; when the topology data meets the preset third roll risk monitoring condition, the vehicle roll risk is level 3, and the system stability status is labeled as critically unstable.

[0069] By combining multiple models to progressively assess vehicle roll risk, a joint inference process is constructed that features hierarchical progression, step-by-step activation, low computational consumption, and high early warning lead time. During vehicle operation, the system gradually upgrades from routine monitoring to high-sensitivity risk scanning, ultimately outputting the chronic roll risk level, risk trigger time window, and system stability status label, achieving full-cycle identification and early warning of chronic roll from its initial accumulation to critical instability.

[0070] It should be noted that, based on inherent vehicle attributes such as vehicle structure, curb weight, suspension type, and load conditions, the output weights of the structural response prediction model, roll evolution prediction model, and topology data analysis algorithm in risk assessment can be adjusted in real time, enabling automakers to achieve personalized and adaptive chronic roll risk identification. Specifically, three types of dynamically adjustable weights are set, with the structural response prediction model corresponding to a specific weight. Weights corresponding to the tilt evolution prediction model Weights corresponding to topology data analysis algorithms and satisfy The system can adaptively calculate weights online based on the following inherent vehicle operating parameters: vehicle platform and body structure (monocoque / body-on-frame), curb weight and actual weight, suspension type (e.g., independent suspension / non-independent suspension, air suspension / coil spring suspension), vehicle center of gravity height, vehicle-to-vehicle distance, and driving conditions.

[0071] Scenario-specific dynamic weighting strategies: ① Weighting mode that emphasizes structural response: suitable for heavy vehicles, high center of gravity vehicles, non-load-bearing bodies, and vehicles under long-term loads. Increase, and The system is appropriately reduced to focus more on chronic, weak-signal, long-cycle structural anomalies such as suspension asymmetry, stress accumulation, and stiffness decay, triggering warnings earlier and increasing risk anticipation. ② A weighted mode emphasizing dynamic evolution is suitable for light passenger vehicles, low center of gravity, independent suspension, and high-speed driving conditions. Increase, and The values ​​are moderately reduced to make the system rely more on real-time dynamic signals such as roll angle, lateral acceleration, and steering input, thereby improving the response speed to rapid changes in roll. ③ Weighted mode with emphasis on topological stability: suitable for high-risk scenarios, continuous curves, long downhill slopes, strong crosswinds, and complex road conditions. Increase, and The values ​​are appropriately reduced so that the system can focus more on topological changes, structural reconstructions, and critical transitions in the high-dimensional state space, and can identify signs of instability before the attitude amplitude exceeds the limit.

[0072] The system performs a tiered, progressive risk assessment based on dynamic weights: Pre-monitoring of the driving structure response prediction model, that is... Adjust the threshold in the first roll risk detection condition to determine whether to proceed to the step of judging whether the second roll risk detection condition is met. Trend judgment of the driving side tilt evolution prediction model, that is, Adjust the threshold of the second roll risk detection condition to determine whether to proceed to the step of judging whether the third roll risk detection condition is met. Mutation determination in driving topology data analysis methods, i.e. Adjust the threshold of the third roll risk detection condition to determine whether the risk transition zone has been entered. After determining that the risk transition zone has been entered, increase the prediction frequency of the roll evolution prediction model and the calculation frequency of the topology data analysis method, and activate vehicle stability control, early warning prompts, or active safety strategies to provide a proactive, interpretable, and quantifiable intervention window for chronic roll.

[0073] It should be noted that, to ensure the long-term stability and accuracy of roll risk identification of the above model in actual road environments, this application also provides a joint model tuning and closed-loop correction mechanism based on long-term real-vehicle operation data. The input data for this mechanism comes from measured driving data under different vehicle models, road conditions, and climate conditions, covering common triggering scenarios for chronic roll, such as continuous gentle slopes, long curves, and strong crosswinds. The output data includes the model's parameter update path, key weight adjustment values, and correction results for trigger thresholds. Internally, the mechanism compares the prediction results of each model with the actual vehicle response through a data feedback module deployed on the vehicle or cloud, statistically analyzes the model residuals, calculates the error distribution trend, and constructs a feedback loss index. Based on this, the model parameters are fine-tuned according to the residual gradient trend, updating the adaptation module weights of the structural response prediction model, the physical weight coefficients of the roll evolution prediction model, and the topological mutation judgment threshold of the topological data analysis method. To ensure that the model evolution does not overfit, a periodic calibration mechanism and stability constraints are introduced to ensure that the update process converges to the steady-state solution domain. Furthermore, this mechanism supports centralized management and feature extraction of data from different vehicle models by automakers' back-end systems, enabling rapid re-adaptation of different vehicle models by generating vehicle model feature embedding vectors. The entire process operates in a closed-loop manner, combined with the adaptive computing resource management capabilities of the in-vehicle system, to achieve continuous evolution and long-term effective operation of the risk model, providing key guarantees for automakers to build a highly robust, real-time responsive chronic roll risk identification system. The system periodically compares model predictions with actual responses using long-term measured data, calculates residuals and error distributions, and generates an update plan (parameter update path) based on "which parameters to adjust, how much to adjust, and in what order." It also provides the increase or decrease in the weights of each module when the three models are fused (key weight adjustment values), and automatically adjusts the risk trigger boundaries / thresholds (such as structural anomaly threshold, trend aggravation threshold, and topological mutation threshold) according to false positives and false negatives, forming a new deployable configuration.

[0074] 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.

[0075] 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.

[0076] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides a vehicle risk prediction device.

[0077] refer to Figure 2 The vehicle risk prediction device includes: The acquisition module 202 is configured to acquire vehicle operating status data and driving operation data over a period of time. The prediction module 204 is configured to predict the structural response based on the vehicle operating status data and through a trained structural response prediction model, and output the vehicle status prediction result for a future period of time. The determination module 206 is configured to determine whether the vehicle state prediction result meets the preset first roll risk detection condition, and in response to meeting the first roll risk detection condition, determine the vehicle roll risk data based on the vehicle operating state data, the driving operation data and the vehicle state prediction result.

[0078] In some embodiments, the determining module 206 is configured to perform attitude prediction based on the vehicle operating status data, the driving operation data and the vehicle status prediction result, using a trained roll evolution prediction model, and output the vehicle attitude prediction result for a future period of time. Determine whether the vehicle attitude prediction result meets the preset second roll risk monitoring condition. In response to meeting the second roll risk detection condition, determine the roll risk data based on the vehicle attitude prediction result and the vehicle state prediction result.

[0079] In some embodiments, the determining module 206 is configured to determine topological data for characterizing vehicle roll risk based on the vehicle attitude prediction result and the vehicle state prediction result using a topological data analysis method. Determine whether the topology data meets the preset third roll risk monitoring conditions, and in response to meeting the third roll risk detection conditions, determine the roll risk data.

[0080] In some embodiments, the prediction module 204 is configured to extract features from the vehicle operating state data through the feature extraction network in the structural response prediction model to obtain a hidden state vector. The hidden state vector is mapped to the vehicle state prediction result through the fully connected layer network in the structural response prediction model.

[0081] In some embodiments, a training module is further included, configured to collect vehicle operating status data for different vehicle models, construct multiple training tasks based on the vehicle operating status data for different vehicle models, wherein each training task corresponds one-to-one with a vehicle model and includes multiple training samples; construct an initial structural response prediction model, which includes an initial feature extraction network and an initial fully connected layer network; for each training task, perform the following operations: update the gradient of the initial fully connected layer network based on a portion of the training samples in the training task to obtain an updated initial structural response prediction model; train the updated initial structural response prediction model based on multiple training tasks, and update the network parameters of the initial feature extraction network in reverse; after training is completed, the structural response prediction model is obtained.

[0082] In some embodiments, the vehicle attitude prediction result includes a predicted roll angle and a predicted roll rate. The determination module 206 is configured to encode the vehicle operating state data, the driving operation data, and the vehicle state prediction result through the encoding layer in the roll evolution prediction model to obtain a joint state vector; based on the joint state vector, fit the attitude evolution law through the state fitting network in the roll evolution prediction model to output an initial predicted roll angle for a future period; based on the initial predicted roll angle, calculate the predicted roll rate for a future period using automatic differentiation; and perform smoothing and boundary constraint processing on the initial predicted roll angle through the post-processing layer in the roll evolution prediction model to obtain the predicted roll angle value.

[0083] In some embodiments, the training module is configured to: construct a training dataset including training samples and ground truth labels; construct an initial roll evolution prediction model; input the training samples into the initial roll evolution prediction model, and output a roll angle continuous state function through an initial state fitting network in the initial roll evolution prediction model; calculate the roll angular velocity based on the roll angle continuous state function using automatic differentiation; construct a data fitting loss function based on the roll angle continuous state function and the ground truth labels; construct a physical constraint loss function based on the roll angular velocity and the vehicle lateral dynamics equations; construct a temporal smoothing constraint loss function based on the roll angle continuous state function; perform weighted fusion of the data fitting loss function, the physical constraint loss function, and the temporal smoothing constraint loss function to obtain a joint loss function; and iteratively train the initial roll evolution prediction model by minimizing the joint loss function to obtain the roll evolution prediction model.

[0084] In some embodiments, the training module is configured such that, during the training of the tilt evolution prediction model, the weight of the data fitting loss function in the joint loss function gradually decreases, while the weight of the physical constraint loss function in the joint loss function gradually increases.

[0085] In some embodiments, before performing structural response prediction based on the vehicle operating status data using a trained structural response prediction model and outputting the vehicle status prediction result for a future period, an adjustment module is further included, configured to acquire vehicle model information; in response to determining that the model information is unknown model information, acquire training samples corresponding to the model information; and perform gradient updates on the fully connected layer network in the structural response prediction model according to the training samples to obtain an updated structural response prediction model.

[0086] 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.

[0087] 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.

[0088] 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.

[0089] Figure 3This 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.

[0090] 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.

[0091] 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.

[0092] 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.

[0093] 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.).

[0094] 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.

[0095] 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.

[0096] 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.

[0097] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides a vehicle, including: a memory for storing executable program code; and a processor for calling and running the executable program code from the memory, so that the vehicle performs the vehicle risk prediction method as described in the foregoing embodiments.

[0098] 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.

[0099] 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.

[0100] 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.

[0101] 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.

[0102] It is understood that before using the technical solutions of the various embodiments in this disclosure, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.

[0103] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to choose, based on the prompt message, whether to provide personal information to the software or hardware such as the electronic device, application, server, or storage medium performing the operations of this disclosed technical solution.

[0104] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" regarding the provision of personal information by the electronic device.

[0105] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0106] 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.

[0107] 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.

[0108] 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.

[0109] 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: Acquire vehicle operating status data and driving operation data over a specified period of time; Based on the vehicle operating status data, structural response prediction is performed using the trained structural response prediction model, and the vehicle status prediction result for a future period of time is output. Determine whether the vehicle state prediction result meets the preset first roll risk detection condition. In response to meeting the first roll risk detection condition, determine the vehicle roll risk data based on the vehicle operating state data, the driving operation data, and the vehicle state prediction result.

2. The method according to claim 1, characterized in that, The determination of vehicle roll risk data based on the vehicle operating status data, the driving operation data, and the vehicle status prediction results includes: Based on the vehicle operating status data, the driving operation data, and the vehicle status prediction results, the attitude prediction is performed using the trained roll evolution prediction model, and the vehicle attitude prediction results for a future period of time are output. Determine whether the vehicle attitude prediction result meets the preset second roll risk monitoring condition. In response to meeting the second roll risk detection condition, determine the roll risk data based on the vehicle attitude prediction result and the vehicle state prediction result.

3. The method according to claim 2, characterized in that, The determination of the roll risk data based on the vehicle attitude prediction result and the vehicle state prediction result includes: Based on the vehicle attitude prediction results and the vehicle state prediction results, topological data for characterizing vehicle roll risk are determined using topological data analysis methods. Determine whether the topology data meets the preset third roll risk monitoring conditions, and in response to meeting the third roll risk detection conditions, determine the roll risk data.

4. The method according to claim 1, characterized in that, The step of predicting the structural response based on the vehicle operating status data using a trained structural response prediction model, and outputting the vehicle status prediction result for a future period of time, includes: The hidden state vector is obtained by extracting features from the vehicle operating state data through the feature extraction network in the structural response prediction model. The hidden state vector is mapped to the vehicle state prediction result through the fully connected layer network in the structural response prediction model.

5. The method according to claim 1, characterized in that, The training method for the structural response prediction model includes: Collect vehicle operation status data for different vehicle models, and construct multiple training tasks based on the vehicle operation status data for different vehicle models. Each training task corresponds to a vehicle model and includes multiple training samples. Construct an initial structural response prediction model, which includes an initial feature extraction network and an initial fully connected layer network; For each training task, perform the following operations: The initial fully connected layer network is updated with gradients based on a portion of the training samples in the training task to obtain an updated initial structural response prediction model; the updated initial structural response prediction model is trained based on multiple training tasks to inversely update the network parameters of the initial feature extraction network. After training, the structural response prediction model is obtained.

6. The method according to claim 2, characterized in that, The vehicle attitude prediction results include the predicted roll angle and the predicted roll rate. The method involves using a trained roll evolution prediction model to predict vehicle attitude based on the vehicle operating status data, driving operation data, and vehicle status prediction results, and outputting vehicle attitude prediction results for a future period, including: The vehicle operating state data, driving operation data, and vehicle state prediction results are encoded through the coding layer in the roll evolution prediction model to obtain a joint state vector. Based on the joint state vector, the attitude evolution law is fitted by the state fitting network in the roll evolution prediction model, and the predicted value of the initial roll angle for a future period of time is output. Based on the initial roll angle prediction value, the roll angular velocity prediction value for a future time period is obtained through automatic differential calculation. The initial roll angle prediction value is smoothed and subjected to boundary constraints through the post-processing layer in the roll evolution prediction model to obtain the roll angle prediction value.

7. The method according to claim 2, characterized in that, The training method for the tilt evolution prediction model includes: Construct a training dataset, which includes training samples and ground truth labels; Construct an initial tilt evolution prediction model; The training samples are input into the initial roll evolution prediction model, and the roll angle continuous state function is output through the initial state fitting network in the initial roll evolution prediction model. Based on the continuous state function of the roll angle, the roll angular velocity is obtained by automatic differentiation calculation; Based on the roll angle continuous state function and the truth label, a data fitting loss function is constructed; Based on the roll rate and the vehicle lateral dynamics equations, a physical constraint loss function is constructed; Based on the roll angle continuous state function, a time-series smoothing constraint loss function is constructed; The data fitting loss function, the physical constraint loss function, and the temporal smoothing constraint loss function are weighted and fused to obtain a joint loss function; The initial tilt evolution prediction model is iteratively trained by minimizing the joint loss function to obtain the tilt evolution prediction model.

8. The method according to claim 7, characterized in that, The method further includes: During the training process of the tilt evolution prediction model, the weight of the data fitting loss function in the joint loss function gradually decreases, while the weight of the physical constraint loss function in the joint loss function gradually increases.

9. The method according to claim 1, characterized in that, Before predicting the structural response based on the vehicle operating status data using a trained structural response prediction model and outputting the vehicle status prediction result for a future period, the method includes: Obtain vehicle model information; In response to determining that the vehicle model information is unknown vehicle model information, the training sample corresponding to the vehicle model information is obtained; The fully connected layer network in the structural response prediction model is updated with gradients based on the training samples to obtain the updated structural response prediction model.

10. 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 9.

Images