A vehicle control method and a vehicle

Abstract

This application provides a vehicle control method and a vehicle. The method, applied in the field of vehicle control technology, includes: acquiring vehicle driving parameter information; determining a target feature vector based on the driving parameter information, the target feature vector reflecting the disturbance state of the vehicle before transitioning from an airborne state to a landing state; determining a target impact level of the vehicle during the transition from an airborne state to a landing state based on the driving parameter information and the target feature vector, the target impact level including multiple different landing impact levels and their corresponding confidence levels; determining a buffer control strategy based on the driving parameter information and the target impact level; and using the buffer control strategy to control the vehicle to perform a landing buffer operation before landing. This method can control the vehicle to perform a landing buffer operation before landing, thereby effectively reducing the impact intensity experienced by the vehicle at the moment of landing after airborne operation and improving vehicle safety.

Description

Technical Field

[0001] This application relates to the field of vehicle control technology, and more specifically, to a vehicle control method and a vehicle within the field of vehicle control technology. Background Technology

[0002] When vehicles are off-roading, they often encounter rugged mountain roads or slopes, which can easily cause the vehicle to become airborne. When the vehicle lands after being airborne, it will be subjected to a huge impact. In severe cases, this can easily cause chassis impact overload, loss of vehicle posture control, or even structural damage, thereby affecting the safety of vehicle use.

[0003] Currently, the main approach is a passive response control strategy that is triggered the instant the vehicle lands after it has taken off the air, in order to make remedial adjustments to the vehicle after it has landed.

[0004] However, even with this method, the vehicle will still experience a huge impact upon landing after being airborne. Therefore, how to reduce the impact intensity on the vehicle upon landing is an urgent problem to be solved. Summary of the Invention

[0005] This application provides a vehicle control method and a vehicle, which can effectively reduce the impact intensity of the vehicle upon landing after it has taken off the ground, thereby improving the safety of vehicle use.

[0006] Firstly, a vehicle control method is provided, comprising: acquiring vehicle driving parameter information; determining a target feature vector based on the driving parameter information, the target feature vector reflecting the disturbance state of the vehicle before switching from an airborne state to a landing state; determining a target impact level of the vehicle when switching from an airborne state to a landing state based on the driving parameter information and the target feature vector, the target impact level including multiple different landing impact levels and their corresponding confidence levels; determining a buffer control strategy based on the driving parameter information and the target impact level; and using the buffer control strategy to control the vehicle to perform a landing buffer operation before landing.

[0007] The above technical solution can predict the target impact level of a vehicle at the moment of landing after takeoff based on the perceived disturbance state of the vehicle before landing. Based on the target impact level, a buffer control strategy can be determined, thereby controlling the vehicle to perform landing buffer operations before landing. This can solve the problem of landing buffer control lag, thereby effectively reducing the impact intensity of the vehicle at the moment of landing after takeoff and improving the safety of vehicle use.

[0008] In conjunction with the first aspect, in some possible implementations, the driving parameter information includes inertial parameter information; determining the target feature vector based on the driving parameter information includes: performing continuous wavelet transform processing on the inertial parameter information to obtain the target time-frequency tensor; and processing the target time-frequency tensor using a disturbance sensing model to obtain the target feature vector.

[0009] The above technical solution can extract key disturbance features from the inertial parameter information before the vehicle lands based on the disturbance perception model, enhance the sensitivity to high-frequency impact signs, and provide higher-dimensional representation support for the subsequent identification of the landing impact level and the decision-making of buffer control strategies.

[0010] Combining the first aspect and the above implementation methods, in some possible implementation methods, the perturbation sensing model is used to process the target time-frequency tensor to obtain the target feature vector, including: using multiple feature extraction modules in the perturbation sensing model to process the target time-frequency tensor sequentially to obtain target extracted features; and using the flattening layer in the perturbation sensing model to perform a flattening operation on the target extracted features to obtain the target feature vector.

[0011] Combining the first aspect and the above implementation methods, in some possible implementation methods, the target impact level of the vehicle when switching from the airborne state to the landing state is determined based on the driving parameter information and the target feature vector, including: extracting the time-domain features corresponding to the driving parameter information to obtain a first feature parameter set; and using an impact level identification model to process the first feature parameter set and the target feature vector to obtain the target impact level.

[0012] By integrating time-domain statistics and frequency-domain depth features, the above technical solution can accurately identify the impact level of a landing event.

[0013] Combining the first aspect and the above implementation methods, in some possible implementation methods, a buffer control strategy is determined based on driving parameter information and target impact level, including: obtaining the target state parameters of the vehicle, which include longitudinal vehicle speed and / or the driving torque corresponding to each wheel in the vehicle; extracting the time-domain features corresponding to the driving parameter information to obtain a second set of feature parameters; and using a buffer control strategy model to process the driving parameter information, the second set of feature parameters, the target state parameters, and the target impact level to obtain the buffer control strategy.

[0014] Combining the first aspect and the above implementation methods, in some possible implementation methods, the buffer control strategy includes a suspension stiffness adjustment value and a torque limiting coefficient corresponding to each wheel in the vehicle; the buffer control strategy is used to control the vehicle to perform a landing buffer operation before landing, including: controlling the vehicle to reduce the suspension stiffness before landing based on the suspension stiffness adjustment value; and / or, controlling the vehicle to reduce the driving torque corresponding to each wheel before landing based on the torque limiting coefficient.

[0015] Through the above technical solution, within a short window before the vehicle lands, suspension stiffness adjustment commands and drive torque limiting strategies are generated in real time based on the current disturbance state and impact level prediction results, thereby achieving active mitigation control of the landing impact.

[0016] Combining the first aspect and the above implementation methods, in some possible implementation methods, the driving parameter information includes inertial parameter information; before controlling the vehicle to reduce the suspension stiffness before landing based on the suspension stiffness adjustment value, it also includes: determining the vehicle body attitude in the air state based on the inertial parameter information; and adjusting the suspension stiffness adjustment value corresponding to the suspension on the target side when the vehicle body attitude is that the target side of the vehicle is sinking.

[0017] Through the above technical solution, the suspension stiffness adjustment value can also be proportionally corrected according to the vehicle's current inertial parameter information to achieve differentiated control of the stiffness of the left and right / front and rear suspensions, reduce the suspension stiffness adjustment value corresponding to the target side suspension, so that the target side suspension will receive a lower stiffness command to resist uneven impacts.

[0018] In combination with the first aspect and the above implementation methods, in some possible implementation methods, before controlling the vehicle to reduce the driving torque corresponding to each wheel before landing based on the torque limit coefficient, the method further includes: determining the road surface type of the road where the vehicle is located, including low-adhesion road surface and high-adhesion road surface; and adjusting the torque limit coefficient based on the road surface type.

[0019] The above technical solution can correct the torque limitation coefficient of each wheel based on the road surface type of the road where the vehicle is located, so as to maintain a higher limitation range in low-friction road surface scenarios, thereby further reducing the driving torque of each wheel, and allowing appropriate retention of some driving force in high-friction road surface scenarios to achieve rapid recovery of traction.

[0020] In combination with the first aspect and the above implementation methods, in some possible implementation methods, the method further includes: determining whether the vehicle is in a target warning state, the target warning state being used to indicate that the vehicle is about to or is in a state of takeoff; if the vehicle is in a target warning state, executing the step of using a buffer control strategy to control the vehicle to perform a landing buffer operation before landing; if the vehicle is not in a target warning state, freezing the buffer control strategy.

[0021] The above technical solution allows for the use of a buffer control strategy to control the vehicle to perform a landing buffer operation only when the vehicle is in a target warning state; when the vehicle is not in a target warning state, the buffer control strategy can be frozen, meaning that the landing buffer operation will not be performed.

[0022] Secondly, a vehicle control device is provided, comprising: a parameter acquisition module for acquiring vehicle driving parameter information; a first determination module for determining a target feature vector based on the driving parameter information, the target feature vector reflecting the disturbance state of the vehicle before switching from an airborne state to a landing state; a second determination module for determining a target impact level of the vehicle when switching from an airborne state to a landing state based on the driving parameter information and the target feature vector, the target impact level including multiple different landing impact levels and their corresponding confidence levels; a third determination module for determining a buffer control strategy based on the driving parameter information and the target impact level; and a control module for controlling the vehicle to perform a landing buffer operation before landing using the buffer control strategy.

[0023] In conjunction with the second aspect, in some possible implementations, the driving parameter information includes inertial parameter information. The first determining module is specifically used to perform continuous wavelet transform processing on the inertial parameter information to obtain the target time-frequency tensor; and to process the target time-frequency tensor using a disturbance sensing model to obtain the target feature vector.

[0024] Combining the second aspect and the above implementation methods, in some possible implementation methods, the first determining module is specifically used to process the target time-frequency tensor sequentially using multiple feature extraction modules in the perturbation sensing model to obtain target extracted features; and to perform a flattening operation on the target extracted features using the flattening layer in the perturbation sensing model to obtain the target feature vector.

[0025] Combining the second aspect and the above implementation methods, in some possible implementation methods, the second determining module is specifically used to extract the time-domain features corresponding to the driving parameter information to obtain the first feature parameter set; and to process the first feature parameter set and the target feature vector using the impact level identification model to obtain the target impact level.

[0026] Combining the second aspect and the above implementation methods, in some possible implementation methods, the third determining module is specifically used to obtain the target state parameters of the vehicle, including the longitudinal vehicle speed and / or the driving torque corresponding to each wheel in the vehicle; extract the time-domain features corresponding to the driving parameter information to obtain the second feature parameter set; and use a buffer control strategy model to process the driving parameter information, the second feature parameter set, the target state parameters and the target impact level to obtain the buffer control strategy.

[0027] Combining the second aspect and the above implementation methods, in some possible implementation methods, the buffer control strategy includes a suspension stiffness adjustment value and a torque limiting coefficient corresponding to each wheel in the vehicle. The control module is specifically used to control the vehicle to reduce the suspension stiffness before landing based on the suspension stiffness adjustment value; and / or, based on the torque limiting coefficient, control the vehicle to reduce the driving torque corresponding to each wheel before landing.

[0028] In conjunction with the second aspect and the above implementation methods, in some possible implementation methods, the driving parameter information includes inertial parameter information, and the vehicle control device may further include: a vehicle attitude determination module, used to determine the vehicle attitude in the airborne state based on the inertial parameter information; and a suspension stiffness adjustment module, used to adjust the suspension stiffness adjustment value corresponding to the target side suspension when the vehicle attitude is that the target side of the vehicle is sinking.

[0029] In conjunction with the second aspect and the above-described implementation methods, in some possible implementation methods, the vehicle control device may further include: a road surface type determination module, used to determine the road surface type of the road where the vehicle is located, the road surface type including low-adhesion road surface and high-adhesion road surface; and a torque limit coefficient adjustment module, used to adjust the torque limit coefficient based on the road surface type.

[0030] In conjunction with the second aspect and the above implementation methods, in some possible implementation methods, the vehicle control device may further include: a warning state determination module, used to determine whether the vehicle is in a target warning state, the target warning state being used to indicate that the vehicle is about to or is currently in a state of takeoff; when the vehicle is in a target warning state, the control module is used to use a buffer control strategy to control the vehicle to perform a landing buffer operation before landing; and a strategy freezing module, used to freeze the buffer control strategy when the vehicle is not in a target warning state.

[0031] Thirdly, a vehicle is provided, comprising: a memory for storing executable program code; and a processor for calling and running the executable program code from the memory, causing the vehicle to perform the methods described in the first aspect or any possible implementation thereof.

[0032] Fourthly, a computer program product is provided, comprising: computer program code, which, when run on a computer, causes the computer to perform the methods described in the first aspect or any possible implementation thereof.

[0033] Fifthly, a computer-readable storage medium is provided that stores computer program code, which, when executed on a computer, causes the computer to perform the methods described in the first aspect or any possible implementation thereof.

[0034] The possible implementations of aspects two through five have similar effects to those of aspect one and its possible implementations, and will not be elaborated upon here. Attached Figure Description

[0035] Figure 1 This is a schematic diagram illustrating an application scenario of the vehicle control method provided in the embodiments of this application; Figure 2 This is a schematic flowchart of a vehicle control method provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a vehicle control device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a vehicle provided in an embodiment of this application. Detailed Implementation

[0036] The technical solutions in this application will be clearly and thoroughly described below with reference to the accompanying drawings. In the description of the embodiments of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. "And / or" in the text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Furthermore, in the description of the embodiments of this application, "multiple" refers to two or more than two.

[0037] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0038] To facilitate understanding of the technical solutions in the embodiments of this application, some terms involved in the embodiments of this application will be briefly explained below.

[0039] Wavelet Convolutional Neural Network (CNN) model: This is a deep learning model that combines wavelet transform and convolutional neural network (CNN). In the wavelet CNN model, data obtained after continuous wavelet transform processing can be used as the model input.

[0040] In some embodiments, the wavelet convolutional neural network model may include multiple cascaded feature extraction modules and flattening layers, and each feature extraction module includes cascaded convolutional layers, batch normalization (BN) layers, activation functions, squeeze and excitation (SE) attention mechanism layers, and max pooling layers.

[0041] The convolutional layer extracts raw features from the input data; the batch normalization layer normalizes the features extracted by the convolutional layer to ensure stable data distribution; the activation function introduces non-linear changes, enabling the model to learn and fit complex patterns; the SE attention mechanism layer weights the features output by the activation function according to the channel dimension, allowing the model to learn the importance of features in different channels and assign higher weights to key channel features, thereby improving the model's feature extraction capability; and the max pooling layer downsamples the features output by the SE attention mechanism layer, reducing data dimensionality and retaining the most significant feature information.

[0042] In its implementation, the SE attention mechanism layer comprises three key steps: squeezing, excitation, and reweighting. The squeezing step primarily performs global average pooling on the features of each channel, thereby aggregating global information from each channel's features and compressing spatial information into a single channel descriptor. The excitation step mainly uses two fully connected layers and an activation function to learn the dependencies between channels and generate weight coefficients for each channel. Specifically, the first fully connected layer reduces the number of channels from C to C / r, where r is the dimensionality reduction coefficient; the second fully connected layer restores the number of channels from C / r back to C; and the activation function maps the weight values ​​to the [0, 1] interval, thus obtaining the final channel weights. The reweighting step mainly multiplies the channel weights with the features input to the SE attention mechanism layer channel by channel, thereby completing the feature calibration of the channel dimension.

[0043] TabNet model: A deep learning model designed specifically for tabular data. Its working principle is based on a sequential multi-step decision framework. It uses an attention mechanism to dynamically select the most relevant subset of features for processing at each step, thereby simulating the hierarchical partitioning process of a decision tree.

[0044] Specifically, the input features are first normalized by a batch normalization layer to ensure the consistency of the model's learning speed across different feature dimensions. Then, they are input into the feature transformer module to generate a set of high-order features containing complex interaction information. These high-order features are then input into the split module. After splitting, they are used as the preprocessed data and input into the attention transformation module of the first decision step.

[0045] In each decision step, contextual information (i.e., in the first decision step, this refers to the preprocessed data) is input to the attentional transformer module to calculate the importance score of each feature, thus obtaining the feature selection mask. The feature selection mask is multiplied by the original input features to filter out the features relevant to the current decision step. These filtered features are then input to the feature transformation module for nonlinear transformation. The output of the feature transformation module is split into two parts by the segmentation module. One part is processed by the rectified linear unit (ReLU) activation function to generate the output result, while the other part is input to the attentional transformer module of the next decision step as contextual information for calculating the next feature selection mask, so that it can be reused in subsequent decision processes. Finally, the outputs generated from all decision steps are summed and then passed through a fully connected (FC) layer to generate the final output of the TabNet model.

[0046] Proximal policy optimization (PPO) is a policy-based reinforcement learning algorithm that continuously interacts with the environment, selects actions based on the current policy, and updates the policy based on the rewards from the environment to maximize long-term cumulative rewards.

[0047] The PPO algorithm employs an Actor (policy neural network model)-Critic (value neural network model) architecture. The policy neural network model selects actions based on the environment state, outputting the mean and variance of the actions; the value neural network model evaluates the value of the environment state, outputting the expected cumulative reward for that state. Specifically, the value neural network model measures the advantage estimate of each action relative to the average level. Based on the advantage signal, the parameters of the policy neural network model are adjusted through gradient ascent, increasing the probability of actions with positive advantage. Furthermore, by minimizing the mean squared error between the value prediction and the actual reward, the value neural network model can more accurately evaluate the state. Thus, by alternately or simultaneously updating the policy neural network model and the value neural network model, they co-evolve during training, ultimately enabling the policy neural network model to make increasingly better decisions.

[0048] In some embodiments, both the policy neural network model and the value neural network model include two fully connected layers, a Dropout layer (i.e., a regularization layer), and an activation function, all cascaded in sequence. The Dropout layer is a regularization technique used in neural networks. By randomly dropping a subset of neurons during training, it prevents overfitting of the model. Its core idea is to force the network to not rely on specific neurons, thereby improving generalization ability.

[0049] Figure 1 This is a schematic diagram illustrating an application scenario of the vehicle control method provided in the embodiments of this application. For example... Figure 1 As shown, the scene may include vehicle 110 and road 120 where vehicle 110 is located. The road surface of road 120 is uneven. The uneven road surface can not only indicate the smoothness and slope of the road surface, but also indicate obstacles on the road surface.

[0050] Vehicle 110 is engaged in off-road driving or driving on roads with poor surface smoothness (such as...). Figure 1 When the road surface 120 is uneven, a lifting-off condition may occur. That is, when vehicle 110 is traveling to position A, if the road surface of 120 is uneven, vehicle 110 may experience the following situation: Figure 1 The vehicle 110 is shown as being airborne at position B. Furthermore, when the vehicle 110 lands at position C after being airborne at position B, it may experience a relatively strong impact. This airborne landing scenario can be referred to as the "jump landing scenario."

[0051] Currently, the main passive response control strategy is based on the moment the vehicle lands after being airborne, in order to make remedial adjustments to the vehicle after landing. However, it cannot identify the risk of landing impact in advance, resulting in lag in suspension adjustment, excessive inertial output of the drive system, etc., which means that the vehicle will still be subjected to a huge impact at the moment of landing after being airborne.

[0052] In related technologies, passive response control strategies based on threshold triggering by a single sensor cannot make coordinated judgments on multi-source disturbances, nor can they analyze the signs of impact in advance. As a result, the system can only make remedial adjustments after landing, and cannot achieve proactive mitigation of impact from the source.

[0053] To address the aforementioned problems, this application provides a vehicle control method. This method involves acquiring vehicle driving parameter information; determining a target feature vector based on the driving parameter information, whereby the target feature vector reflects the vehicle's disturbance state before transitioning from an airborne state to a landing state; determining the target impact level of the vehicle during the transition from airborne to landing state based on the driving parameter information and the target feature vector, whereby the target impact level includes multiple different landing impact levels and their corresponding confidence levels; determining a buffer control strategy based on the driving parameter information and the target impact level; and using the buffer control strategy to control the vehicle to perform a landing buffer operation before landing. Therefore, this application can predict the target impact level experienced by the vehicle at the moment of landing after airborne operation based on the perceived disturbance state of the vehicle before landing, and determine a buffer control strategy based on the target impact level, thereby controlling the vehicle to perform a landing buffer operation before landing. This solves the problem of landing buffer control lag, effectively reducing the impact intensity experienced by the vehicle at the moment of landing after airborne operation and improving vehicle safety.

[0054] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments.

[0055] Figure 2 This is a flowchart illustrating a vehicle control method provided in an embodiment of this application. This vehicle control method can be applied to vehicles, such as... Figure 2 As shown, the vehicle control method may specifically include the following steps: S201, obtain vehicle driving parameter information.

[0056] Vehicles are equipped with inertial measurement units (IMUs), wheel speed sensors, and suspension displacement sensors (such as active suspension displacement sensors).

[0057] The inertial measurement unit (IMU) is used to collect inertial parameter information, including triaxial acceleration and triaxial angular velocity. The triaxial acceleration can include lateral acceleration, longitudinal acceleration, and vertical acceleration. The triaxial angular velocity can include roll angular velocity (i.e., the angular velocity of the vehicle body around the x-axis), pitch angular velocity (i.e., the angular velocity of the vehicle body around the y-axis), and yaw angular velocity (i.e., the angular velocity of the vehicle body around the z-axis). Wheel speed sensors are used to collect the wheel speeds of each wheel in the vehicle. Suspension displacement sensors are used to collect the suspension displacements corresponding to each wheel in the vehicle, which characterize the relative displacement of each wheel relative to the vehicle body in the vertical direction. Therefore, in this embodiment, the driving parameter information may include inertial parameter information, wheel speeds, and suspension displacements, etc.

[0058] In one implementation, the sampling frequency of the inertial measurement unit (IMU) can be set to 1000Hz to fully reproduce the detailed dynamics during the jump, especially the high-frequency disturbance information before landing. Wheel speed sensors and suspension displacement sensors provide intuitive feedback on the vehicle's longitudinal impact, vertical motion trend, and changes in suspension compression characteristics. The sampling frequencies of the wheel speed sensors and suspension displacement sensors can be the same as the sampling frequency of the IMU.

[0059] It should be noted that the data collected by the inertial measurement unit, wheel speed sensors, and suspension displacement sensors are aligned using synchronized timestamps and uniformly sampled and cached in the embedded system by the central data acquisition gateway, ensuring that the multi-source data has a unified time reference and completeness without missing data. Furthermore, for abnormal data situations, the vehicle also incorporates a rule-based noise removal module, such as sensor drift correction, abrupt value filtering, and spectrum anomaly smoothing, to ensure data quality stability.

[0060] S202, determine the target feature vector based on driving parameter information. The target feature vector is used to reflect the disturbance state of the vehicle before switching from the air state to the landing state.

[0061] The "airborne state" refers to the state in which the vehicle's wheels are off the road surface, while the "landed state" refers to the state in which the vehicle is in contact with the road surface. The disturbance state of the vehicle before it switches from the airborne state to the landed state represents the unexpected and non-ideal dynamic interference signals caused by sudden changes in the external environment or motion state before the vehicle lands.

[0062] In one possible implementation, the driving parameter information includes inertial parameter information. The above-mentioned S202 "determine the target feature vector based on the driving parameter information" may specifically include the following steps: perform continuous wavelet transform processing on the inertial parameter information to obtain the target time-frequency tensor; process the target time-frequency tensor using a disturbance sensing model to obtain the target feature vector.

[0063] To improve the perception capability of the disturbance perception model in the short-term response before vehicle landing, a time-frequency joint preprocessing strategy needs to be constructed. First, a sliding window extraction mechanism is built for the inertial parameter information collected by the inertial measurement unit (IMU). For example, the sliding window size is set to 500ms and the step size to 50ms, thereby extracting the inertial parameter information based on the sliding window to obtain the inertial parameter information within each time window. Next, continuous wavelet transform processing is performed on the inertial parameter information within each time window to obtain the target time-frequency tensor. The wavelet basis function used in the continuous wavelet transform processing can be Morlet wavelet, etc. The target time-frequency tensor is a three-channel two-dimensional time-frequency tensor related to acceleration and angular velocity. The three channels of the target time-frequency tensor correspond to the energy distribution of longitudinal acceleration, pitch angular velocity, and roll angular velocity at different frequencies, respectively. The time axis dimension of the target time-frequency tensor represents the sampling points within the time window, and the frequency axis dimension represents the spectral resolution. The target time-frequency tensor is mainly used to capture the high-frequency disturbance distribution and energy density change patterns of the vehicle in the hundreds of milliseconds before landing.

[0064] To ensure the stability of the preprocessing output, the target time-frequency tensor corresponding to each time window can be standardized to a unified range. Then, the standardized target time-frequency tensor is used as the input of the perturbation sensing model, so that the perturbation sensing model can output the target feature vector.

[0065] Furthermore, the aforementioned "processing the target time-frequency tensor using a perturbation-aware model to obtain the target feature vector" can specifically include the following steps: processing the target time-frequency tensor sequentially using multiple feature extraction modules in the perturbation-aware model to obtain target extracted features; and performing a flattening operation on the target extracted features using a flattening layer in the perturbation-aware model to obtain the target feature vector.

[0066] In one example, the disturbance perception model can be a wavelet convolutional neural network model, which is used to extract key disturbance features from the inertial parameter information of the vehicle before landing in the scenario of the vehicle switching from the air state to the landing state (i.e., the jump scenario), enhance the sensitivity to high-frequency impact signs, and provide higher-dimensional representation support for the subsequent identification of the landing impact level and the decision of buffer control strategy.

[0067] Specifically, the overall structure of the perturbation-aware model adopts a three-layer stacked convolutional design. That is, the perturbation-aware model can include three sequentially cascaded feature extraction modules. The target time-frequency tensor is used as the input of the first feature extraction module, the output of the first feature extraction module is used as the input of the second feature extraction module, the output of the second feature extraction module is used as the input of the third feature extraction module, and the output of the third feature extraction module is the target extracted feature.

[0068] Each feature extraction module includes a series of cascaded convolutional layers, batch normalization layers, activation functions, SE attention mechanisms, and max pooling layers. Furthermore, from the first to the last feature extraction module, the receptive field of the convolutional layers gradually increases. This reflects a layer-by-layer expansion strategy in the perturbation-aware model, where the kernel size progresses from a smaller receptive field to a larger one, facilitating the extraction of perturbation evolution trends from fine-grained to coarse-grained.

[0069] The receptive field refers to the size of the input data that each neuron in a perturbation-aware model can perceive. The larger the receptive field, the more contextual information the neuron can perceive. The size of the receptive field depends on the structure of the perturbation-aware model, including the number of convolutional layers, the size of the convolutional kernels, the stride, and padding.

[0070] It should be noted that the convolutional layer in the first feature extraction module focuses on extracting short-term local perturbation patterns, such as the subtle attitude changes of the vehicle just before landing; the convolutional layers in the second and third feature extraction modules focus on changes in the global frequency domain energy structure, which can form a deep recognition capability for specific perturbation patterns before the vehicle lands.

[0071] Furthermore, in each feature extraction module, a batch normalization layer and an activation function are added after the convolutional layer. The activation function is uniformly ReLU to enhance nonlinear expressiveness, and a max pooling layer is connected to compress the feature map dimension. To enhance the perturbation-aware model's ability to perceive important frequency channels, an SE attention mechanism layer is introduced to explicitly model the inter-channel response weights, thereby improving the discriminative power under unstructured perturbations.

[0072] In this way, the target time-frequency tensor is processed sequentially by the three feature extraction modules in the disturbance perception model, so that the last feature extraction module can output the target extracted features. Furthermore, the disturbance perception model also includes a flattening layer. The target extracted features output by the last feature extraction module can be used as input to the flattening layer, allowing the flattening layer to perform a flattening operation on the target extracted features to obtain a target feature vector. This target feature vector serves as a compressed representation of the disturbance state before the vehicle impact within that time window.

[0073] The flattening operation refers to the process of converting multi-dimensional input data into a one-dimensional vector. The target feature vector can have 128 dimensions to ensure sufficient expressiveness and computational efficiency in subsequent model fusion.

[0074] S203, based on driving parameter information and target feature vector, determines the target impact level when the vehicle switches from the airborne state to the landing state. The target impact level includes multiple different landing impact levels and their corresponding confidence levels.

[0075] In some embodiments, after determining the target feature vector based on driving parameter information, the target impact level experienced by the vehicle at the moment of landing can be further determined based on the driving parameter information and the target feature vector.

[0076] In one possible implementation, the above-mentioned S203 "determines the target impact level of the vehicle when it switches from the airborne state to the landing state based on driving parameter information and target feature vector" may specifically include the following steps: extracting the time-domain features corresponding to the driving parameter information to obtain a first feature parameter set; and using an impact level identification model to process the first feature parameter set and the target feature vector to obtain the target impact level.

[0077] Therefore, this application embodiment achieves accurate identification of the impact level by fusing time-domain statistics and frequency-domain depth features, while ensuring the interpretability and training efficiency of the impact level identification model.

[0078] In one example, the impact level recognition model can be the TabNet model, whose core advantage lies in its end-to-end feature selection capability and multi-step inference mechanism. It can automatically perform hierarchical filtering and dynamic weight allocation of input features, and is suitable for situations with complex feature distribution and diverse data modalities in vehicle jumping scenarios.

[0079] In this embodiment, a sliding window extraction mechanism can be constructed based on the inertial parameter information collected by the inertial measurement unit, the wheel speed of each wheel in the vehicle collected by the wheel speed sensor, and the suspension displacement collected by the suspension displacement sensor. This allows for the extraction of the time-domain features corresponding to the inertial parameter information, the time-domain features corresponding to the wheel speed, and the time-domain features corresponding to the suspension displacement within each time window, thus obtaining a first set of feature parameters.

[0080] For example, the first set of characteristic parameters may include multiple time-domain features within the time window, such as the maximum value of longitudinal acceleration, the mean value of longitudinal acceleration, the variance of longitudinal acceleration, the standard deviation of longitudinal acceleration, the slope of the rising edge of longitudinal acceleration, the slope of the falling edge of longitudinal acceleration, the peak duration of longitudinal acceleration, the rate of change of pitch angular velocity, the rate of change of roll angular velocity, the rate of change of speed corresponding to the wheel rotation speed of each wheel, and the suspension compression rate corresponding to each wheel. These time-domain features reflect the macroscopic motion trend and changes in impact risk of the vehicle throughout the entire process from the start of takeoff to landing. It should be understood that the suspension compression rate refers to the rate of change of suspension displacement over time.

[0081] The input to the impact level identification model consists of two parts: a target feature vector output by the perturbation sensing model and a first set of feature parameters. All input features in the first set of feature parameters are normalized before being input into the impact level identification model to ensure consistency in the learning speed of the model across different feature dimensions. This allows the impact level identification model to output the target impact level.

[0082] The target impact level can include three different landing impact levels and their corresponding confidence levels, such as a light landing impact level and its corresponding confidence level, a moderate landing impact level and its corresponding confidence level, and a severe landing impact level and its corresponding confidence level. In other words, the target impact level output by the impact level identification model is a three-dimensional probability vector, representing the confidence level (i.e., predicted probability) for each of the light, moderate, and severe landing impact levels.

[0083] In other embodiments, the landing impact level with the highest confidence can be selected from the target impact levels as the vehicle's landing impact risk level.

[0084] For example, if the confidence level for a mild landing impact is 0.7, the confidence level for a moderate landing impact is 0.2, and the confidence level for a severe landing impact is 0.1, then the vehicle's landing impact risk level is a mild landing impact.

[0085] S204 determines the buffer control strategy based on driving parameter information and target impact level.

[0086] In some embodiments, after determining the target impact level of the vehicle when switching from an airborne state to a landing state based on driving parameter information and target feature vector, a buffer control strategy can be further determined based on driving parameter information and target impact level. This buffer control strategy is used to control the vehicle to perform a landing buffer operation before landing.

[0087] In one possible implementation, the above-mentioned S204 "determine the buffer control strategy based on driving parameter information and target impact level" may specifically include the following steps: obtaining the target state parameters of the vehicle, including longitudinal vehicle speed and / or the driving torque corresponding to each wheel in the vehicle; extracting the time-domain features corresponding to the driving parameter information to obtain a second set of feature parameters; and using a buffer control strategy model to process the driving parameter information, the second set of feature parameters, the target state parameters and the target impact level to obtain the buffer control strategy.

[0088] In one example, the buffer control strategy model can be a policy neural network model trained based on the PPO algorithm. PPO, as a proximal policy optimization algorithm, is suitable for policy learning in high-dimensional, continuous action spaces, possessing advantages such as high sample efficiency and strong training stability. It is particularly suitable for vehicle control problems with rapid dynamic response and complex state transitions. The buffer control strategy model is used to generate suspension stiffness adjustment commands and drive torque limiting strategies in real time within a short window before the vehicle lands, based on the current disturbance state and impact level prediction results, thereby achieving active mitigation control of the landing impact.

[0089] In the embodiments of this application, target state parameters of the vehicle can be obtained, including longitudinal vehicle speed and / or the driving torque (i.e., driving torque distribution state) corresponding to each wheel in the vehicle.

[0090] Furthermore, a sliding window extraction mechanism can be constructed based on the inertial parameter information collected by the inertial measurement unit and the suspension displacement collected by the suspension displacement sensor, thereby extracting the time-domain features corresponding to the inertial parameter information and the time-domain features corresponding to the suspension displacement in each time window, and obtaining the second feature parameter set.

[0091] For example, the second set of feature parameters may include the suspension compression rate and the average longitudinal acceleration of each wheel within the time window. The second set of feature parameters may be completely identical to the first set of feature parameters, or the second set of feature parameters may share only some temporal features with the first set of feature parameters.

[0092] In addition, it can also obtain driving parameter information such as pitch rate, roll rate and vehicle longitudinal acceleration.

[0093] The input to the buffer control strategy model consists of two parts. One part is the target impact level output by the impact level identification model, namely, the light impact level and its corresponding confidence level, the moderate impact level and its corresponding confidence level, and the heavy impact level and its corresponding confidence level, which are encoded into a set of three-dimensional probability vectors to guide the buffer control strategy model to identify the current risk level and make targeted responses. The other part is the vehicle's current dynamic state parameters, including the aforementioned driving parameter information, the second feature parameter set, and the target state parameters, namely, pitch rate, roll rate, vehicle longitudinal acceleration, suspension compression rate corresponding to each wheel, mean longitudinal acceleration, longitudinal vehicle speed, and drive torque corresponding to each wheel in the vehicle.

[0094] The buffer control strategy model is used to process driving parameter information, second characteristic parameter set, target state parameters and target impact level, so that the buffer control strategy model can output buffer control strategy.

[0095] The buffer control strategy model consists of two continuous action control signals, which correspond to the suspension stiffness adjustment value and the torque limiting coefficient of each wheel in the vehicle, respectively.

[0096] It should be understood that the suspension stiffness adjustment value output by the buffer control strategy model and the torque limit coefficient corresponding to each wheel in the vehicle are actually the average values ​​of the actions output by the strategy neural network model.

[0097] S205 employs a buffer control strategy to control the vehicle to perform a landing buffer operation before landing.

[0098] In this embodiment, the buffer control strategy output by the buffer control strategy model can be effectively mapped to the actual actuator in the vehicle chassis control system. The actual actuator in the vehicle chassis control system controls the vehicle to perform a landing buffer operation before landing, thereby completing a precise control closed loop from strategy decision result to physical action.

[0099] In one possible implementation, the buffer control strategy includes a suspension stiffness adjustment value and a torque limiting coefficient corresponding to each wheel in the vehicle; the above-mentioned S205 "using a buffer control strategy to control the vehicle to perform a landing buffer operation before landing" may specifically include the following steps: based on the suspension stiffness adjustment value, controlling the vehicle to reduce the suspension stiffness before landing; and / or, based on the torque limiting coefficient, controlling the vehicle to reduce the driving torque corresponding to each wheel before landing.

[0100] The suspension stiffness adjustment value in the buffer control strategy corresponds to a specific control channel on the active suspension system in the vehicle, which is used to send a real-time adjustment request to the active suspension controller. The torque limiting coefficient corresponding to each wheel in the buffer control strategy corresponds to a specific control channel on the power system controller in the vehicle, which is used to pre-limit the power system controller to prevent the driving impact from expanding at the moment of landing, and to perform parameter decoupling and response linkage adjustment according to the current vehicle state.

[0101] For example, an active suspension system can be a continuous damping control (CDC) system or an e-active body control (E-ABC) system; a powertrain controller can be a traction control system (TCS) or an electric vehicle control unit (VCU) system.

[0102] In the buffer control strategy, the suspension stiffness adjustment value can be a value between 0 and 1, which represents the normalization coefficient of the feedforward desired suspension response stiffness. The vehicle maps this suspension stiffness adjustment value to the four independent suspension solenoid valves or motor controllers, so that they complete the desired stiffness switching within a specific time window before landing, thereby controlling the vehicle to reduce the suspension stiffness of each suspension in the vehicle before landing.

[0103] In one example, a preset mapping relationship exists between the suspension stiffness adjustment value and the final adjusted suspension stiffness. This mapping relationship allows the final adjusted suspension stiffness to be determined based on the suspension stiffness adjustment value. Specifically, the suspension stiffness adjustment value and the final adjusted suspension stiffness are positively correlated: a larger suspension stiffness adjustment value results in a larger final adjusted suspension stiffness, and a smaller suspension stiffness adjustment value results in a smaller final adjusted suspension stiffness.

[0104] In another example, the suspension stiffness adjustment value can also represent the final adjusted suspension stiffness.

[0105] The torque limiting coefficient in the buffer control strategy is actually four values, which correspond to the torque limiting coefficients of the four wheels of the vehicle. They are used to control the vehicle to reduce the driving torque of each wheel before landing, thereby reducing the structural impact or body attitude disturbance caused by the sudden change in the instantaneous grip force of the drive wheels under high impact conditions.

[0106] Specifically, there is a preset mapping relationship between the torque limiting coefficient and the driving torque of each wheel. This mapping relationship allows the driving torque to be determined based on the torque limiting coefficient. The torque limiting coefficient and the driving torque of each wheel are negatively correlated. A larger torque limiting coefficient results in a smaller driving torque for that wheel; conversely, a smaller torque limiting coefficient results in a larger driving torque for that wheel.

[0107] In other embodiments, the driving parameter information includes inertial parameter information; before the above-mentioned "controlling the vehicle to reduce suspension stiffness before landing based on the suspension stiffness adjustment value", the following steps may also be included: determining the vehicle body posture in the air state based on the inertial parameter information; when the vehicle body posture is that the target side of the vehicle is sinking, adjusting the suspension stiffness adjustment value corresponding to the suspension on the target side.

[0108] Taking into account the effects of uneven vehicle posture and asymmetrical terrain, the suspension stiffness adjustment value can also be proportionally corrected based on the vehicle's current inertial parameter information to achieve differentiated control of the stiffness of the left and right / front and rear suspensions.

[0109] Specifically, the vehicle's attitude in the air can be determined based on the pitch and roll angular velocities in the inertial parameters. The roll angular velocity determines the height difference between the left and right sides of the vehicle, while the pitch angular velocity determines whether the front of the vehicle rises or falls.

[0110] Therefore, when the vehicle's body posture is determined to be a target side depression, the suspension stiffness adjustment value corresponding to the target side is reduced, while the suspension stiffness adjustment value corresponding to the other sides remains unchanged.

[0111] For example, when the vehicle's body posture is such that the right side of the vehicle is noticeably sloping down, the suspension stiffness adjustment value corresponding to the right side suspension can be reduced, so that the right side suspension will receive a lower stiffness command to resist uneven impacts.

[0112] It should be understood that before the suspension stiffness adjustment values ​​for each suspension are corrected based on the vehicle body posture, the suspension stiffness adjustment values ​​for each suspension are the same.

[0113] It should be noted that the specific reduction range of the suspension stiffness adjustment value can be set according to actual conditions, and this application embodiment does not limit this. For example, the reduction range of the suspension stiffness adjustment value can be the product of the suspension stiffness adjustment value and a first preset ratio, such as 5% or 10%.

[0114] In other embodiments, before the above-mentioned "controlling the vehicle to reduce the driving torque corresponding to each wheel before landing based on the torque limiting coefficient", the following steps may be included: determining the road surface type of the road where the vehicle is located, including low-adhesion road surface and high-adhesion road surface; adjusting the torque limiting coefficient based on the road surface type.

[0115] Before mapping the driving torque of each wheel based on the torque limit coefficient of each wheel, the torque limit coefficient of each wheel can be corrected based on the road surface type of the road where the vehicle is located.

[0116] Specifically, on low-traction surfaces such as snow, mud, and slippery surfaces, the torque limiting coefficient for each wheel can be increased to maintain a higher limiting range in low-traction scenarios, thereby further reducing the driving torque for each wheel. On high-traction surfaces such as asphalt surfaces, the torque limiting coefficient for each wheel can be kept unchanged or decreased, allowing for the retention of some driving force to achieve rapid traction recovery in high-traction scenarios.

[0117] It should be noted that the specific increase or decrease in the torque limiting coefficient can be set according to the actual situation, and this application embodiment does not limit this. For example, the increase in the torque limiting coefficient can be the product of the torque limiting coefficient and a second preset ratio, such as 5%, 6%, 10%, or 15%, etc.; the decrease in the torque limiting coefficient can be the product of the torque limiting coefficient and a third preset ratio, such as 5%, 6%, 10%, or 15%, etc.; or, the increase in the torque limiting coefficient can be a preset first coefficient threshold, such as 0.05 or 0.1, etc.; the decrease in the torque limiting coefficient can be a preset second coefficient threshold, such as 0.05 or 0.1, etc.

[0118] In addition, before controlling the vehicle to reduce suspension stiffness before landing based on the suspension stiffness adjustment value, and before controlling the vehicle to reduce the driving torque corresponding to each wheel before landing based on the torque limit coefficient, the vehicle's on-board safety inspection module can be used for constraint processing to set the upper and lower limits corresponding to suspension stiffness and driving torque, so as to prevent the actuator from exceeding the range command.

[0119] Thus, when the determined suspension stiffness and drive torque do not exceed the upper and lower limits, the vehicle is directly controlled to reduce its suspension stiffness before landing, and the drive torque corresponding to each wheel is also reduced before landing, based on the suspension stiffness adjustment value. When the determined suspension stiffness exceeds the upper and lower limits, the determined suspension stiffness is constrained based on these limits, thereby controlling the vehicle to adjust the suspension stiffness to the constrained stiffness before landing; and / or, when the determined drive torque exceeds the upper and lower limits, the drive torque is constrained based on these limits, thereby controlling the vehicle to adjust the drive torque corresponding to each wheel to the constrained drive torque before landing.

[0120] For example, if the determined suspension stiffness exceeds the upper limit, the constrained suspension stiffness will be the upper limit; if the determined suspension stiffness exceeds the lower limit, the constrained suspension stiffness will be the lower limit. Similarly, if the determined drive torque exceeds the upper limit, the constrained drive torque will be the upper limit; if the determined drive torque exceeds the lower limit, the constrained drive torque will be the lower limit.

[0121] In one possible implementation, the vehicle control method may further include the following steps: determining whether the vehicle is in a target warning state, the target warning state being used to indicate that the vehicle is about to or is in a state of takeoff; if the vehicle is in a target warning state, executing the step S205 above, "using a buffer control strategy to control the vehicle to perform a landing buffer operation before landing"; if the vehicle is not in a target warning state, freezing the buffer control strategy.

[0122] In this embodiment of the application, it can be determined whether the vehicle is in a target warning state by one or more of the following: the road surface smoothness of the road where the vehicle is located, the suspension displacement corresponding to each wheel in the vehicle, the wheel speed of each wheel in the vehicle, the vertical acceleration of the vehicle, the vehicle speed, and the accelerator pedal opening.

[0123] It should be noted that during the vehicle's operation, steps S201 to S204 described above will be continuously executed to determine the buffer control strategy. The buffer control strategy will only be adopted when the vehicle is in a target warning state to control the vehicle to perform a landing buffer operation before landing; if the vehicle is not in a target warning state, the buffer control strategy can be frozen, meaning that the landing buffer operation will not be executed.

[0124] Therefore, the embodiments of this application can predict the target impact level that the vehicle will receive at the moment of landing after takeoff based on the perceived disturbance state of the vehicle before landing, and determine the buffer control strategy based on the target impact level, thereby controlling the vehicle to perform landing buffer operation before landing. This can solve the problem of landing buffer control lag, thereby effectively reducing the impact intensity of the vehicle at the moment of landing after takeoff and improving the safety of vehicle use.

[0125] It should be noted that, to ensure the coordination of responses between actuators, the system is designed with a linkage response synchronization mechanism. This mechanism sends out multi-channel control frames via the main electronic control unit (ECU), packaging and sending out the suspension stiffness adjustment value and the torque limit coefficient (i.e., power control command) for each wheel. The response is then dynamically adjusted based on the feedback status within the control cycle. For example, if the delayed response of a certain actuator is detected to exceed a set threshold, the corresponding strategy output will be paused, and the system will revert to the default mechanical setting to prevent unilateral excitation from causing vehicle body posture imbalance.

[0126] Furthermore, to address the differences in hardware control interfaces between different vehicle models, the control mapping layer is designed as a configurable parameter module. This allows engineers to define the mapping function form and scaling factor between the suspension and power interfaces according to different chassis systems, supporting seamless integration of various control objects such as CDC, air suspension, and electric drive axle. The overall design ensures a stable, clearly decoupled, and rapidly responding path for the strategy control signals from model output to physical execution, making it the key control hub module for realizing the entire landing buffer operation.

[0127] It should be noted that the disturbance perception model, impact level identification model, and buffer control strategy model introduced in the embodiments of this application are all pre-trained models, which need to be trained before actual application.

[0128] During the training phase of the disturbance perception model, a sample set can be constructed using actual vehicle jump test data. The training samples in the sample set include the target time-frequency tensor corresponding to the training samples, the temporal features corresponding to the training samples, and a manually labeled landing impact level label. The landing impact level label can be obtained by recording the scene based on images captured by a camera and subsequently manually labeling and verifying it.

[0129] Specifically, during the training of the perturbation sensing model, the input is the target time-frequency tensor corresponding to the training samples, and the supervision target is the impact level label within 0.5 seconds after the corresponding time window. During training, a regression target can be constructed using the mean squared error loss function. By fitting the mapping relationship between the target feature vector corresponding to the training samples and the impact level label, the perturbation sensing model to be trained is guided to implicitly predict the severity of potential impacts in the short time before impact.

[0130] The optimizer uses the Adaptive Moment Estimation (Adam) algorithm with an initial learning rate of 0.001. During training, a gradient pruning strategy is employed to limit gradient explosion and prevent convergence instability. Simultaneously, to enhance the generalization ability of the perturbation-aware model, data augmentation strategies are used in the training samples to simulate perturbations in the original inertial parameters, including random Gaussian noise injection, short-term amplitude stretching, and simulation of sample point loss. The input tensor is dynamically updated in each batch during the training cycle.

[0131] In the model performance verification phase, the focus is on examining the correlation distribution between the output features of the perturbation perception model and the landing impact level label, as well as the feature robustness performance under different terrain types and jump scenarios. The final perturbation perception model can output a 128-dimensional target feature vector.

[0132] In training the impact level recognition model, a cross-entropy loss function can be used to construct the classification target. Considering that the proportion of light impact samples is significantly higher than heavy impact samples in actual off-road driving, a class weight penalty mechanism is introduced to increase the training focus weight for the heavy impact category, preventing the impact level recognition model from predicting the mainstream category. The AdamW optimizer is used, with an initial learning rate of 0.002. A learning rate decay strategy is enabled during training, with a decay rate of 0.9 per round, improving the convergence stability of the impact level recognition model in the later stages of training.

[0133] To address feature drift caused by scene changes, a cross-terrain, cross-speed, and multi-vehicle data hybrid strategy was adopted in the training set construction to ensure the impact level recognition model has good versatility and robustness. This structure not only improves training efficiency but also explicitly outputs the usage frequency of each feature dimension, thus providing vehicle engineers with a clear physical feature interpretation basis for subsequent system optimization. In the evaluation phase of the impact level recognition model, confusion matrix analysis revealed that it has higher recall and accuracy than traditional multilayer perceptrons in severe impact prediction, especially demonstrating significant recognition ability in scenarios with high suspension compression ratio and sudden pitch rate changes before landing.

[0134] During the training of the buffer control strategy model, the PPO algorithm uses a shearing objective function to update the policy parameters of the policy neural network model, ensuring that the policy updates remain stable in the policy space. A simulation environment is used to construct the feedback loop during training. A high-fidelity vehicle dynamics simulation platform is used to construct jump-landing scenarios. A closed-loop training system is formed by combining the CarSim vehicle model (a whole-vehicle dynamics simulation model) and the Python Gym interface (a standardized toolkit interface for developing and comparing reinforcement learning algorithms). In the simulation environment, diverse training samples are generated by introducing variables such as different terrain slopes, takeoff speeds, vehicle weight distributions, and wind resistance disturbances to improve the policy's generalization ability. The reward function is constructed with minimizing the peak longitudinal acceleration after landing as the primary objective, while suspension response time and torque output smoothness are introduced as auxiliary constraint factors, forming a multi-objective adjustment strategy. The peak longitudinal impact is set as the primary negative reward indicator. Penalties are generated for excessively long response times or drastic fluctuations in control actions, prompting the buffer control strategy model to achieve control smoothness while protecting the structure. After training stabilized, the strategy's effectiveness was verified through real-world data playback. The buffer control strategy model was able to perform buffer control actions 200 milliseconds before impact, significantly reducing the average impact intensity and keeping the response command latency within 20 milliseconds, demonstrating good real-time performance and stability. The final trained buffer control strategy model was deployed to the vehicle controller in embedded TorchScript, interacting with the main ECU via middleware to achieve feedforward adjustment functionality in the vehicle control system. This differentiates it from traditional ground-triggered control logic, creating a forward-looking intelligent control approach.

[0135] It should be noted that, since the embodiments of this application require the participation of disturbance perception models, impact level identification models, and buffer control strategy models when determining the buffer control strategy, in order to construct a well-structured and real-time-efficient model fusion framework, it is necessary to effectively connect the disturbance perception model, impact level identification model, and buffer control strategy model to form a closed-loop information processing chain in the impact buffering task. Each of the three models performs a different function: the disturbance perception model is responsible for sensing high-frequency disturbances, the impact level identification model is responsible for identifying the impact level, and the buffer control strategy model is responsible for generating the control strategy.

[0136] The fusion process begins with the 128-dimensional target feature vector output by the disturbance perception model. This target feature vector represents the energy distribution and disturbance intensity of inertial parameters in the frequency domain within the current time window, serving as one of the main inputs to the impact level identification model. Simultaneously, the first set of feature parameters generated during data preprocessing is also input into the impact level identification model, achieving a dual-channel fusion expression in the frequency and time domains. After completing feature selection and decision inference, the impact level identification model outputs a three-dimensional probability vector, representing the confidence level for each category: light, moderate, and severe impact. This three-dimensional probability vector serves not only as the direct output of the impact level classification result but also as a crucial input to the subsequent buffer control strategy model. Before entering the buffer control strategy model, this three-dimensional probability vector is concatenated into the vehicle's current dynamic state parameters, forming a complete set of state inputs to guide the buffer control strategy model in identifying the current operating condition level and potential risk level, thereby enabling targeted buffer control actions.

[0137] To ensure data alignment consistency across models at different frame rates and computation cycles, the information flow graph employs a sliding window caching mechanism to manage input-output relationships. The disturbance perception model and the impact level recognition model operate at high frequency, processing new data every 50ms, while the buffer control strategy model updates control actions every 100ms. The two time scales are decoupled and synchronized through a cache queue to ensure stable data transmission without excessively blocking the main control process.

[0138] In addition, the information flow structure between all models adopts a modular design, which is encapsulated and called through vehicle artificial intelligence (AI) middleware. Each model is encapsulated as an independent service node, and the output structure is transmitted through a standardized message format, which mainly includes data type identifier, timestamp, model output value and status flag bit.

[0139] During the fusion process, a confidence-weighted mechanism is introduced to assist the buffer control strategy model in action sampling. Specifically, the impact level probability vector output by the impact level identification model will trigger a control strategy adjustment process under low confidence conditions. The buffer control strategy model internally uses a penalty term to increase action conservatism, improving safety and fault tolerance under high uncertainty conditions. The entire fusion system is designed with a clear information thread, unified interface protocols, and reasonable computational load sharing as its goals. In actual deployment, the model can run in parallel, asynchronously, or sequentially according to the computing platform capabilities of different vehicle models, ensuring that the control response meets real-time requirements while also possessing the ability to generate the optimal buffer strategy. Finally, the information flow graph forms a closed-loop path in the functional architecture, from perception feature extraction and risk level identification to control strategy output. Each stage is optimized based on personalized vehicle data to meet the engineering requirements of complex impact management in off-road conditions.

[0140] It is worth noting that after each model is deployed to the vehicle and the entire vehicle is online, real vehicle data can be collected and uploaded remotely based on over-the-air (OTA) technology. The cloud can also make fine adjustments to each model based on the OTA real vehicle data.

[0141] The above combination Figure 2 The vehicle control method provided in the embodiments of this application has been described. The apparatus for performing the above method provided in the embodiments of this application is described below.

[0142] Figure 3 This is a schematic diagram of the structure of a vehicle control device provided in an embodiment of this application. Figure 3 As shown, the vehicle control device 300 may include: a parameter acquisition module 301, a first determination module 302, a second determination module 303, a third determination module 304, and a control module 305.

[0143] The system includes: a parameter acquisition module 301 for acquiring vehicle driving parameter information; a first determination module 302 for determining a target feature vector based on the driving parameter information, the target feature vector reflecting the disturbance state of the vehicle before switching from the airborne state to the landing state; a second determination module 303 for determining the target impact level of the vehicle when switching from the airborne state to the landing state based on the driving parameter information and the target feature vector, the target impact level including multiple different landing impact levels and their corresponding confidence levels; a third determination module 304 for determining a buffer control strategy based on the driving parameter information and the target impact level; and a control module 305 for controlling the vehicle to perform a landing buffer operation before landing using the buffer control strategy.

[0144] In one possible implementation, the driving parameter information includes inertial parameter information. The first determining module 302 is specifically used to perform continuous wavelet transform processing on the inertial parameter information to obtain the target time-frequency tensor; and to process the target time-frequency tensor using a disturbance sensing model to obtain the target feature vector.

[0145] In one possible implementation, the first determining module 302 is specifically used to process the target time-frequency tensor sequentially using multiple feature extraction modules in the perturbation sensing model to obtain target extracted features; and to perform a flattening operation on the target extracted features using a flattening layer in the perturbation sensing model to obtain a target feature vector.

[0146] In one possible implementation, the second determining module 303 is specifically used to extract the time-domain features corresponding to the driving parameter information to obtain a first feature parameter set; and to process the first feature parameter set and the target feature vector using an impact level identification model to obtain the target impact level.

[0147] In one possible implementation, the third determining module 304 is specifically used to obtain the target state parameters of the vehicle, including the longitudinal vehicle speed and / or the driving torque corresponding to each wheel in the vehicle; extract the time-domain features corresponding to the driving parameter information to obtain the second feature parameter set; and use a buffer control strategy model to process the driving parameter information, the second feature parameter set, the target state parameters and the target impact level to obtain the buffer control strategy.

[0148] In one possible implementation, the buffer control strategy includes a suspension stiffness adjustment value and a torque limiting coefficient corresponding to each wheel in the vehicle. The control module 305 is specifically used to control the vehicle to reduce the suspension stiffness before landing based on the suspension stiffness adjustment value; and / or, based on the torque limiting coefficient, control the vehicle to reduce the drive torque corresponding to each wheel before landing.

[0149] In one possible implementation, the driving parameter information includes inertial parameter information. The vehicle control device 300 may further include: a vehicle attitude determination module for determining the vehicle attitude in an airborne state based on the inertial parameter information; and a suspension stiffness adjustment module for adjusting the suspension stiffness adjustment value corresponding to the target side suspension when the vehicle attitude is a target side sag.

[0150] In one possible implementation, the vehicle control device 300 may further include: a road surface type determination module for determining the road surface type of the road where the vehicle is located, the road surface type including low-adhesion road surface and high-adhesion road surface; and a torque limit coefficient adjustment module for adjusting the torque limit coefficient based on the road surface type.

[0151] In one possible implementation, the vehicle control device 300 may further include: a warning state determination module, used to determine whether the vehicle is in a target warning state, the target warning state being used to indicate that the vehicle is about to or is in a state of takeoff; when the vehicle is in a target warning state, the control module 305 is used to control the vehicle to perform a landing buffer operation before landing using a buffer control strategy; and a strategy freezing module, used to freeze the buffer control strategy when the vehicle is not in a target warning state.

[0152] Figure 4 This is a schematic diagram of the structure of a vehicle provided in an embodiment of this application. For example, as shown... Figure 4 As shown, the vehicle 400 includes a memory 401 and a processor 402. The memory 401 stores executable program code 4011, and the processor 402 is used to call and execute the executable program code 4011 to perform a vehicle control method.

[0153] Furthermore, embodiments of this application also protect an apparatus that may include a memory and a processor, wherein the memory stores executable program code, and the processor is used to call and execute the executable program code to perform a vehicle control method provided in embodiments of this application.

[0154] This embodiment can divide the device into functional modules based on the above method example. For example, each module can correspond to a separate function, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware. It should be noted that the module division in this embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0155] When each functional module is divided according to its corresponding function, the device may further include a parameter acquisition module, a first determination module, a second determination module, a third determination module, and a control module. It should be noted that all relevant content regarding the steps involved in the above method embodiments can be referenced to the functional descriptions of the corresponding functional modules, and will not be repeated here.

[0156] It should be understood that the device provided in this embodiment is used to execute the above-described vehicle control method, and therefore can achieve the same effect as the above-described implementation method.

[0157] When using an integrated unit, the device may include a processing module and a storage module. When the device is applied to a vehicle, the processing module can be used to control and manage the vehicle's movements. The storage module can be used to support the vehicle in executing relevant program code.

[0158] The processing module may be a processor or a controller, which can implement or execute various exemplary logic blocks, modules, and circuits shown in conjunction with the disclosure of this application. The processor may also be a combination of functions that implement computing capabilities, such as a combination of one or more microprocessors, a combination of digital signal processing (DSP) and a microprocessor, etc., and the storage module may be a memory.

[0159] In addition, the device provided in the embodiments of this application may specifically be a chip, component or module. The chip may include a connected processor and a memory. The memory is used to store instructions. When the processor calls and executes the instructions, the chip can execute a vehicle control method provided in the above embodiments.

[0160] This embodiment also provides a computer-readable storage medium storing computer program code. When the computer program code is run on a computer, the computer executes the above-described related method steps to implement a vehicle control method provided in the above embodiment.

[0161] This embodiment also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned related steps to implement a vehicle control method provided in the above embodiment.

[0162] In this embodiment, the device, computer-readable storage medium, computer program product, or chip are all used to execute the corresponding methods provided above. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods provided above, and will not be repeated here.

[0163] Through the above description of the embodiments, those skilled in the art will understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0164] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0165] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A vehicle control method, characterized in that, The method includes: Obtain vehicle driving parameter information; A target feature vector is determined based on the driving parameter information. The target feature vector is used to reflect the disturbance state of the vehicle before it switches from the airborne state to the grounded state. Based on the driving parameter information and the target feature vector, the target impact level of the vehicle when switching from the airborne state to the landing state is determined. The target impact level includes multiple different landing impact levels and their corresponding confidence levels. Based on the driving parameter information and the target impact level, a buffer control strategy is determined; The aforementioned buffer control strategy is used to control the vehicle to perform a landing buffer operation before landing.

2. The method according to claim 1, characterized in that, The driving parameter information includes inertial parameter information; determining the target feature vector based on the driving parameter information includes: The inertial parameter information is subjected to continuous wavelet transform processing to obtain the target time-frequency tensor; The target time-frequency tensor is processed using a perturbation-sensing model to obtain the target feature vector.

3. The method according to claim 2, characterized in that, The step of processing the target time-frequency tensor using a perturbation-sensing model to obtain the target feature vector includes: The target time-frequency tensor is processed sequentially using multiple feature extraction modules in the disturbance perception model to obtain target extracted features; The target feature vector is obtained by performing a flattening operation on the extracted features of the target using the flattening layer in the perturbation sensing model.

4. The method according to claim 1, characterized in that, The step of determining the target impact level of the vehicle when switching from the airborne state to the landing state based on the driving parameter information and the target feature vector includes: Extract the time-domain features corresponding to the driving parameter information to obtain the first feature parameter set; The first set of feature parameters and the target feature vector are processed using an impact level identification model to obtain the target impact level.

5. The method according to claim 1, characterized in that, The step of determining a buffer control strategy based on the driving parameter information and the target impact level includes: Obtain the target state parameters of the vehicle, including the longitudinal speed and / or the driving torque corresponding to each wheel in the vehicle; Extract the time-domain features corresponding to the driving parameter information to obtain the second feature parameter set; The buffer control strategy is obtained by processing the driving parameter information, the second set of feature parameters, the target state parameters, and the target impact level using a buffer control strategy model.

6. The method according to any one of claims 1 to 5, characterized in that, The buffer control strategy includes a suspension stiffness adjustment value and a torque limiting coefficient corresponding to each wheel in the vehicle; controlling the vehicle to perform a landing buffer operation before landing using the buffer control strategy includes: Based on the suspension stiffness adjustment value, the vehicle is controlled to reduce the suspension stiffness before landing; And / or, based on the torque limiting coefficient, control the vehicle to reduce the driving torque corresponding to each wheel before landing.

7. The method according to claim 6, characterized in that, The driving parameter information includes inertial parameter information; before controlling the vehicle to reduce suspension stiffness before landing based on the suspension stiffness adjustment value, it also includes: The vehicle's body posture in the airborne state is determined based on the inertial parameter information; When the vehicle body posture is such that the target side of the vehicle is depressed, the suspension stiffness adjustment value corresponding to the suspension on the target side is adjusted.

8. The method according to claim 6, characterized in that, Before controlling the vehicle to reduce the drive torque corresponding to each wheel before landing based on the torque limiting coefficient, the method further includes: The road surface type of the road where the vehicle is located is determined, and the road surface type includes low-adhesion road surface and high-adhesion road surface; The torque limiting coefficient is adjusted based on the road surface type.

9. The method according to claim 1, characterized in that, The method further includes: Determine whether the vehicle is in a target warning state, the target warning state being used to indicate that the vehicle is about to or is currently in the airborne state; When the vehicle is in the target warning state, the step of using the buffer control strategy to control the vehicle to perform a landing buffer operation before landing is executed; If the vehicle is not in the target warning state, the buffer control strategy is frozen.

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