Control method of a vehicle wheel and vehicle

Abstract

The application provides a wheel control method and a vehicle. The method is applied to the field of intelligent driving and includes the following steps: obtaining a plurality of vehicle component dynamic data of the vehicle during vehicle driving, the vehicle component dynamic data being obtained by associating vehicle dynamic state data and vehicle component static data, the vehicle component dynamic data being used to represent the dynamic coupling strength between vehicle components, and the plurality of vehicle component dynamic data being continuously generated; determining a risk trend of suspension of at least one wheel of the vehicle in a future time step based on the plurality of vehicle component dynamic data; and generating a torque adjustment instruction for the corresponding wheel based on the risk trend of suspension and current power state data of the vehicle, so as to adjust the driving torque in advance before the wheel is suspended. The method can reduce the driving torque of the corresponding wheel before the wheel has a risk trend of suspension, prevent the wheel from idling, and reduce power loss.

Description

Technical Field

[0001] This application relates to the field of intelligent driving, and more specifically, to a wheel control method and vehicle in the field of intelligent driving. Background Technology

[0002] When a vehicle encounters puddles or potholes while driving, the vehicle's traction control system will wait until the wheels begin to slip before cutting off power or applying braking force to the wheels.

[0003] However, in practical applications, when a vehicle drives over continuous speed bumps or rough, cross-axle surfaces, the vehicle's wheels may become suspended in the air. That is, when the vehicle's left front wheel is about to bounce off the speed bump and leave the ground, the vehicle will send a large torque to the left front wheel, causing it to spin freely in the air. In this situation, when the left front wheel lands, the driving torque on it will be wasted, and the suspended left front wheel can easily lead to vehicle instability, reducing vehicle stability and even making the driver feel a loss of steering control, resulting in a poor driving experience. Summary of the Invention

[0004] This application provides a wheel control method and a vehicle. The method can reduce the driving torque of the corresponding wheel before the wheel of the vehicle tends to be suspended in the air, thereby preventing the wheel from spinning and reducing power loss.

[0005] Firstly, a method for controlling a wheel is provided, the method comprising: During vehicle operation, dynamic data of multiple vehicle components are acquired. This dynamic data is obtained by constructing a correlation based on the vehicle's dynamic state data and the static data of the vehicle components. The dynamic data of the vehicle components is used to represent the dynamic coupling strength between the vehicle components. These multiple dynamic data of the vehicle components are generated continuously. Based on the dynamic data of these multiple vehicle components, the risk trend of at least one wheel of the vehicle being suspended in the air is determined within a future time step. Based on the risk of the wheel being suspended in mid-air and the vehicle's current power status data, a torque adjustment command is generated for the corresponding wheel to adjust the drive torque in advance before the wheel is actually suspended in mid-air.

[0006] By employing the aforementioned method, dynamic data of vehicle components obtained through correlation can be constructed to predict the trend of wheel suspension risk within future time steps. Combined with current power state data, the driving torque of the wheels can be proactively adjusted before they actually become suspended, thus achieving forward-looking prediction and early proactive intervention of wheel suspension risk. Compared to traditional methods that only trigger torque adjustment measures when the wheel has already become suspended, this solution fundamentally overcomes the lag in intervention timing in related technologies. In other words, the method provided in this application can identify potential dangers and take early torque adjustment measures in the early stages of wheel suspension risk accumulation, i.e., before the wheel actually becomes suspended, thereby avoiding problems such as wheel spinning, wasted power, and vehicle posture imbalance, and effectively improving the vehicle's driving stability and safety under complex operating conditions.

[0007] Secondly, a wheel control device is provided, the device comprising: The acquisition module is used to acquire dynamic data of multiple vehicle components during vehicle operation. The dynamic data of vehicle components is obtained by constructing a correlation based on the vehicle's dynamic state data and the static data of the vehicle components. The dynamic data of vehicle components is used to represent the dynamic coupling strength between the vehicle components. The dynamic data of multiple vehicle components is generated continuously. The determination module is used to determine the air suspension risk trend of at least one wheel of the vehicle within a future time step based on the dynamic data of the multiple vehicle components. The generation module is used to generate torque adjustment commands for the corresponding wheels based on the risk of suspension and the current power status data of the vehicle, so as to adjust the drive torque in advance before the wheel actually suspends.

[0008] Thirdly, a vehicle is provided, including a memory and a processor. The memory stores executable program code, and the processor calls and runs the executable program code from the memory, causing the vehicle to perform the methods described above for controlling the wheels.

[0009] Fourthly, a computer program product is provided, comprising: computer program code, which, when run on a computer, causes the computer to execute the method described above for controlling the wheel.

[0010] 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 above for controlling the wheels. Attached Figure Description

[0011] Figure 1This is a schematic diagram of the implementation environment of a wheel control method provided in an embodiment of this application; Figure 2 This is a schematic flowchart of a wheel control method provided in an embodiment of this application; Figure 3 This is a schematic flowchart of another wheel control method provided in an embodiment of this application; Figure 4 This is a schematic flowchart illustrating a model building method provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a wheel control device provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of a vehicle provided in an embodiment of this application. Detailed Implementation

[0012] 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.

[0013] 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.

[0014] Before introducing the methods of the embodiments of this application, the technical terms that may be involved in the embodiments of this application will be explained first.

[0015] Evolutionary Graph Neural Networks (EvoGNNs) are used to process graph structures that evolve dynamically over time. Examples include changes in user relationships in social networks, real-time changes in traffic flow in transportation networks, and real-time changes in the components of a vehicle during its journey.

[0016] An Inertial Measurement Unit (IMU) is a hardware sensor module that typically includes an accelerometer and a gyroscope. The accelerometer measures linear acceleration, while the gyroscope measures angular velocity. IMUs have a high update rate and high short-term accuracy, but they suffer from integral drift over time.

[0017] Graph Attention Network (GAT) is a graph neural network architecture that aggregates information from neighboring nodes. GAT automatically learns the importance (attention coefficient) of each neighbor to the current node in the graph structure.

[0018] In related technologies, when vehicles encounter complex road conditions such as jumps, potholes, and unilateral slippage, their wheels may become suspended in the air. In such situations, vehicles generally rely on the slip ratio or longitudinal acceleration of the drive wheels for torque correction; that is, they control the output torque of the wheels after the vehicle is suspended in the air. This reactive response mechanism lacks the ability to anticipate and actively intervene in situations where the wheels are about to become suspended in the air, which can easily lead to problems such as wheel spin, wasted power, and vehicle imbalance.

[0019] To address at least one of the aforementioned technical problems, embodiments of this application provide a wheel control method. This method can predict the risk of wheel suspension by combining dynamic data of vehicle components, and generate a torque adjustment command based on the suspension risk, so as to adjust the drive torque in advance before the wheel actually becomes suspended.

[0020] In the following description of the embodiments of this application, it is used as... Figure 1 Taking an example, the implementation environment of the embodiments of this application will be introduced.

[0021] For example, such as Figure 1 As shown, the implementation environment includes a driver assistance controller 110 and an actuator 120.

[0022] The driver assistance controller 110 is a terminal installed on the vehicle. It can acquire and process relevant information. The driver assistance controller 110 is communicatively connected to the actuator 120, and the information processed by the driver assistance controller 110 can be executed by the actuator 120. The actuator 120 includes power components (engine / drive motor), braking system, steering system, and suspension system, etc.

[0023] The technical solution provided in this application can be applied to vehicles equipped with driver assistance controllers. By adopting the technical solution provided in this application, the driving torque of the corresponding wheel can be reduced before the vehicle wheel is at risk of being suspended in the air, which not only improves the safety of the vehicle but also reduces power loss.

[0024] After introducing the implementation environment and application scenarios of the embodiments of this application, the technical solutions provided by the embodiments of this application are described below. (See also...) Figure 2 Taking the driver assistance controller as the executing entity as an example, the method includes the following steps.

[0025] Step 201: During vehicle operation, acquire dynamic data of multiple vehicle components. The dynamic data of vehicle components is constructed based on the vehicle's dynamic state data and the static data of the vehicle components. The dynamic data of vehicle components is used to represent the dynamic coupling strength between the vehicle components. The dynamic data of multiple vehicle components is generated continuously.

[0026] In this application, "vehicle" refers to the vehicle that is of interest and controlled. The vehicle is equipped with various sensors, capable of collecting its own status information and environmental information in real time, and has the ability to execute control commands. Vehicle components are the various functional units that constitute the vehicle as a whole, as well as products that serve the vehicle. Vehicle components include mechanical or dynamic units such as the suspension, wheels, axles, body, and steering system. Vehicle component dynamic data is used to characterize the real-time interactions between vehicle components. Vehicle dynamic state data refers to real-time data continuously collected over a period of time, used to characterize vehicle components. Examples include real-time signals such as vehicle speed, wheel speed, vehicle attitude, acceleration, and suspension travel. Vehicle component static data characterizes the time-invariant design parameters of each vehicle component. Examples include parameters such as vehicle mass, center of gravity position, suspension stiffness, and wheel radius. The association representation fuses vehicle dynamic state data and vehicle component static data through a model. Dynamic coupling strength characterizes the degree to which a change in one vehicle component affects another component. Continuous generation generates a set of data to characterize each time step, forming a time series. That is, the dynamic data of multiple vehicle components are composed in the form of a time series.

[0027] It should be understood that during vehicle operation, dynamic data from vehicle components reflects the strong coupling relationships within the vehicle's dynamics system. Single-signal monitoring methods for vehicle components often suffer from lag, only able to make a judgment when a wheel actually becomes airborne or slips. In such cases, the vehicle's posture has already lost balance, easily leading to a rollover risk. To effectively avoid such risks, dynamic data from multiple vehicle components is acquired to determine the coupling state of the entire vehicle's dynamics system.

[0028] Step 202: Based on the dynamic data of the multiple vehicle components, determine the risk trend of at least one wheel of the vehicle being suspended in the air within a future time step.

[0029] The future time step is used to represent several points in the future. The wheel's levitation risk trend is used to represent the direction of change in the wheel's future levitation status.

[0030] It should be understood that when a vehicle encounters a bump, pothole, or sharp turn, the dynamic data of the vehicle components will change drastically. That is, the interaction between the dynamic data of the vehicle components will become abnormal, and the coupling strength between the vehicle components will fluctuate violently. For example, if a vehicle wheel hits a bump, the suspension will suddenly compress, and the coupling strength between the wheel and the suspension will increase instantaneously, resulting in abnormal wheel acceleration. In this case, the wheel is more likely to be at risk of being lifted off the ground. Based on the dynamic data of these multiple vehicle components, the trend of the risk of at least one wheel being lifted off the ground within a future time step can be determined.

[0031] Step 203: Based on the risk of suspension and the current power status data of the vehicle, generate torque adjustment commands for the corresponding wheels to adjust the drive torque in advance before the wheel actually becomes suspended.

[0032] The current dynamic state data characterizes the degree of matching between the vehicle's powertrain and road surface adhesion conditions. Torque adjustment commands are used to adjust the drive torque of the wheels. The corresponding wheel is the one predicted to have a risk of being lifted off the ground. "Before the wheel actually lifts off the ground" indicates the point before the wheel actually leaves the ground.

[0033] It should be understood that during vehicle operation, waiting until the wheels are already off the ground before reducing drive torque would result in wasted driving force and the high-speed rotation of the wheels upon landing could easily lead to loss of vehicle control. In such cases, it is crucial to anticipate the risk of wheel suspension in advance and proactively adjust the drive torque of the wheel before it actually becomes suspended, thus preventing it from spinning freely for a moment before it is suspended.

[0034] This application provides a wheel control method that, by constructing and associating dynamic data of vehicle components, predicts the risk trend of wheel suspension within a future time step and, combined with current power state data, proactively adjusts the wheel's driving torque before the wheel actually becomes suspended, thereby achieving forward-looking prediction and early proactive intervention of wheel suspension risk. Compared to traditional methods that only trigger torque adjustment measures when the wheel has already become suspended, this solution fundamentally overcomes the defect of delayed intervention in related technologies. That is, the method provided in this application can identify potential dangers and take early torque adjustment measures in the early stages of wheel suspension risk accumulation, i.e., before the wheel actually becomes suspended, thereby avoiding problems such as wheel spinning, wasted power, and vehicle posture imbalance, and effectively improving the vehicle's driving stability and safety under complex operating conditions.

[0035] It should be noted that step 201 above 203 is a brief description of a wheel control method provided in an embodiment of this application. The following will provide a more detailed description of the wheel control method provided in an embodiment of this application, using some examples. See also... Figure 3Taking the driver assistance controller as the executing entity as an example, the method includes the following steps.

[0036] Step 301: During vehicle operation, the driver assistance controller acquires dynamic data of multiple vehicle components.

[0037] Among them, the dynamic data of multiple vehicle components refers to the data obtained by collecting and constructing associations between the dynamic state data of the vehicle and the static data of the vehicle components at continuous or discrete time points.

[0038] In one possible implementation, vehicle dynamic state data and vehicle component static data are acquired; at different sampling times, multiple nodes are determined based on the vehicle component static data, each node representing the attributes of the vehicle components; based on the physical layout and functional association of the vehicle components, each vehicle dynamic state data is mapped to a spatial edge of the node, each spatial edge representing the physical coupling relationship between the nodes; based on the order of sampling times, a spatiotemporal graph structure is constructed at different sampling times based on the node and the spatial edge, the edges of the spatiotemporal graph structure representing the state transition relationship between the nodes across time steps; the spatiotemporal graph structure of multiple consecutive time steps is constructed sequentially to form multiple vehicle component dynamic data.

[0039] It should be understood that, since vehicle dynamic status data and vehicle component static data are independent of each other, in order to explore the static dynamic attribute characteristics of each vehicle component, the vehicle dynamic status data and the vehicle component static data are linked together to observe their change patterns.

[0040] Vehicle dynamic state data reflects multiple aspects of a vehicle, including kinematics, dynamics, and environmental perception. For example, dynamic data from multiple vehicle components includes kinematic and dynamic information such as suspension travel, real-time speed, acceleration, angular velocity, roll angle, pitch angle, and yaw rate. It also includes environmental perception information from onboard sensors (such as radar and cameras) and output information from internal vehicle controllers (such as engine speed, brake pressure, and steering angle). This vehicle component dynamic data represents the dynamic coupling strength between vehicle components, and these multiple vehicle component dynamic data are generated continuously. Vehicle component static data indicates the attribute information of the vehicle components. Multiple vehicle component dynamic data are arranged sequentially in a time-series diagram structure.

[0041] A spatiotemporal graph structure is a graphical model that can simultaneously represent the relationships between data in both spatial and temporal dimensions. Nodes in this spatiotemporal graph structure characterize the attributes of vehicle components. For example, information such as wheel information, vehicle center of gravity information, left and right suspension information, front motor information, and rear motor information can all be used as independent nodes. These nodes carry node feature vectors, such as wheel load magnitude and wheel speed vectors. Edges in this spatiotemporal graph structure characterize the physical coupling relationships or state transition relationships across time steps between nodes. Physical coupling relationships can refer to the inherent connections between different nodes at the same moment, such as the relationship between the left and right wheels and the vehicle suspension. State transition relationships across time steps refer to the evolution of the same sensor data at different moments, such as the dynamic correlation between the left wheel load at the current moment and the left wheel load at the next moment. In one implementation, the measurement value of each vehicle component at each sampling moment can be used as a node, and spatial edges are established between nodes corresponding to different sensors at the same moment, based on the physical location or functional relevance of the vehicle component on the vehicle. Simultaneously, temporal edges are established between nodes at consecutive sampling moments. In some embodiments, a multi-layered spatiotemporal graph structure can be adopted, where one layer represents spatial relationships and another layer represents temporal relationships. For example, a spatial graph is constructed at each time step, and then corresponding nodes in the spatial graphs of adjacent time steps are connected to form temporal edges.

[0042] In this implementation, by mapping the static data of vehicle components at each sampling time point to nodes, clear and discretized data units are provided for subsequent spatiotemporal graph structure processing. Based on this, spatial edges are constructed according to the physical layout and functional associations of vehicle components, reflecting the real-time physical coupling effects between different components of the vehicle at the same time, such as the longitudinal stability association between the front and rear wheels, and the steering coupling between the front wheels and the vehicle body. By fusing vehicle dynamic state data and vehicle component static data, a spatiotemporal graph structure characterizing physical coupling and cross-time step state transitions is constructed, forming a multi-dimensional dynamic data set of vehicle components from a continuous graph structure. Based on this, the driver assistance controller can capture the gradual trend of coupling strength between nodes, thereby enabling early identification of wheel suspension risks and prediction of their evolution. This method integrates a spatiotemporal graph structure with spatial and temporal information, providing high-quality input for subsequent spatiotemporal attention mechanisms.

[0043] To provide a clearer explanation of the above embodiments, the following description is divided into several parts.

[0044] The first part explains the acquisition of the vehicle's dynamic status data and the static data of the vehicle components.

[0045] The purpose of acquiring vehicle dynamic status data and vehicle component static data is to collect raw vehicle dynamic status data through sensors at a first preset frequency and a first preset time window length. The first preset frequency is automatically determined by the driver assistance controller, and this embodiment does not limit this frequency; for example, the first preset frequency is 100 Hz. The first preset time window length is automatically determined by the driver assistance controller, and this embodiment does not limit this length; for example, the first preset time window length is 5 to 10 seconds.

[0046] For example, suspension displacement sensors collect the vehicle's suspension travel and its rate of change, that is, the suspension tension and extension speed of each wheel. Suspension pressure sensors collect the vehicle's shock absorber pressure, spring force, and dynamic vertical force between the wheels and the body, that is, the real-time force on the suspension system. Vehicle height sensors collect the vehicle's ground clearance and attitude. Steering angle sensors collect the vehicle's steering wheel and wheel angles. The vehicle's stability status and wheel slip ratio are obtained from the Electronic Stability Control (ESC) via the Controller Area Network (CAN). The vehicle's traction control status is obtained from the Traction Control System (TCS) via the CAN. The anti-lock braking system (ABS) status is obtained from the anti-lock braking system (ABS) via the CAN. The ABS status is obtained from the Vehicle Control Unit (VCU) via the CAN. The vehicle's current transmission status is obtained from the transmission controller via the CAN.

[0047] For example, the vehicle's triaxial acceleration and triaxial angular velocity can be collected by an inertial measurement unit (IMU); the vehicle's acceleration can be obtained by an acceleration sensor; the wheel speed of each wheel of the vehicle can be obtained by a wheel speed sensor; and the vertical force on the wheel and the dynamic contact status of the tires can be obtained by a wheel load sensor.

[0048] It should be understood that, due to the inconsistent sampling frequencies of the various sensors, the raw vehicle dynamic state data collected is often accompanied by high noise characteristics, so preprocessing of the raw vehicle dynamic state data is necessary.

[0049] In some embodiments, using the sensor with the highest sampling frequency as the time reference, Kalman filtering and cubic spline interpolation are employed to complete the time alignment and interpolation compensation of the original vehicle dynamic state data; subsequently, Gaussian outlier removal and noise filtering are performed on the original vehicle dynamic state data; finally, data standardization processing is applied to the original vehicle dynamic state data to output the vehicle dynamic state data. Static data of vehicle components is obtained through the vehicle's unique identifier. For example, static data of vehicle components includes information such as wheel information, body center of gravity information, left and right suspension information, front motor information, and rear motor information.

[0050] The second part explains how to determine multiple nodes based on static data of vehicle components at different sampling times.

[0051] At different sampling times, multiple nodes are determined based on static data of vehicle components, aiming to transform the raw, discrete static data of vehicle components into basic units of a spatiotemporal graph structure. Static data of vehicle components is abstracted into an independent node, and this node carries node features that clearly represent the node's origin and time point, laying the foundation for subsequent construction of the spatiotemporal graph structure. For example, at a certain time t, if the wheel is abstracted as a node, then the acceleration data of the wheel from the inertial measurement unit can be one node feature, and the left front wheel speed data from the wheel speed sensor can be another node feature. This mapping method gives each node feature a clear identity and location, facilitating construction and processing within the spatiotemporal graph structure. Mapping the measurement values ​​of vehicle components from different sensors to node features provides clear, discretized data units for subsequent spatiotemporal graph structure processing.

[0052] The third part explains the physical layout and functional association of vehicle components, and how the dynamic state data of each vehicle is mapped to the spatial edges of nodes.

[0053] Based on the physical layout and functional associations of vehicle components, the dynamic state data of each vehicle are mapped to spatial edges of nodes, aiming to capture the spatial correlation between different sensor data within the vehicle at the same moment. The various sensors in a vehicle do not exist in isolation; they have specific positional relationships in physical space (e.g., the left front wheel speed sensor and the right front wheel speed sensor) or are closely related in function (e.g., the steering angle sensor and the vehicle attitude sensor). By analyzing these physical layouts and functional associations, spatial edges can be established between related nodes at the same sampling moment. These spatial edges represent the physical coupling relationships between sensor data. For example, the difference in left and right wheel speeds may reflect the lateral motion state of the vehicle, while the data from the suspension travel sensor and the vehicle attitude sensor together reflect the roll or pitch state of the vehicle.

[0054] In some embodiments, a sensor association matrix or rule base can be predefined, recording which sensors have physical proximity or functional dependencies. When constructing the spatiotemporal graph structure, for all nodes at the same sampling time, this association matrix or rule base is queried. If the sensors corresponding to two nodes satisfy the association condition, a spatial edge is added between them. For example, a spatial edge can be added between all wheel speed sensors of the left front wheel and the right front wheel mounted on the same axle. Alternatively, by analyzing the vehicle's 3D model, nodes with spatial proximity relationships can be automatically identified, and spatial edges can be constructed between these nodes at the same time.

[0055] Part Four explains how to construct a spatiotemporal graph structure based on the node and spatial edge at different sampling times, according to the order of sampling.

[0056] Based on the chronological order of sampling times, a spatiotemporal graph structure is constructed at different sampling times based on the node and the spatial edge, aiming to capture the dynamic characteristics of vehicle dynamic state data evolving over time. Vehicle motion is a continuous process, and there is a necessary causal and evolutionary relationship between the current state and the previous state. By leveraging the explicit chronological order of sampling times in the collected vehicle dynamic state data, temporal edges can be established between nodes corresponding to different sampling times. These temporal edges explicitly represent the state transition relationships across time steps. For example, the wheel node at time t and the wheel node at time t+1 are connected by a temporal edge, reflecting the changes in wheel load, wheel speed, and vertical acceleration over time. This construction of temporal edges enables the spatiotemporal graph structure to encode the dynamic evolution process of the vehicle state, providing crucial information for predicting future states.

[0057] In some embodiments, the node set at each sampling time can be traversed and connected to the corresponding node at the next adjacent sampling time. For example, for the vehicle dynamic state data a of vehicle component A at time t, it can be connected to the data a of vehicle component A at time t+ A time edge is added between the vehicle dynamic state data b at time t. This connection method ensures that the node features of each vehicle component form a sequence in the time dimension. Alternatively, a more complex strategy can be adopted, such as connecting not only nodes at adjacent time points but also nodes spanning multiple time steps to capture longer-term dependencies. For example, a node at time t can be connected to a node at time t+k. The nodes of time step t are defined, where k is a preset step size. Furthermore, the weights of the time edges can be adjusted based on time intervals or vehicle dynamic state data to reflect the dynamic coupling strength between vehicle components. The nodes, time edges, and previously constructed spatial edges of each time step are merged to construct a complete spatiotemporal graph structure.

[0058] Part Five explains how the spatiotemporal diagram structure of multiple consecutive time steps is constructed in a temporal sequence to form dynamic data of multiple vehicle components.

[0059] In some embodiments, after constructing the spatiotemporal graph structure for a single time step, the spatiotemporal graph structures for multiple consecutive time steps are further stacked and concatenated in chronological order. This temporal splicing transforms the discrete spatiotemporal graph structure into a dynamic data stream with continuous evolutionary characteristics, ultimately forming dynamic data of vehicle components containing rich spatiotemporal information, i.e., a time-series graph sequence. This time-series graph sequence can completely record the dynamic trajectory of the state changes of each vehicle component over time.

[0060] Step 302: Based on the dynamic data of the multiple vehicle components, the driver assistance controller determines the risk trend of at least one wheel of the vehicle being off the ground in a future time step.

[0061] Among them, "lifting" refers to the situation where the normal contact force between the wheel and the ground drops to near zero, causing the wheel to lose its load-bearing capacity and ability to transmit driving force. The lifting risk trend characterizes the dynamic evolution of the probability of a wheel lifting off the ground over a future period; that is, how the probability of a wheel lifting off the ground changes over time.

[0062] In one possible implementation, for any node in the spatiotemporal graph structure of the multiple vehicle component dynamic data, an importance weight is determined based on the importance influence between the node and its neighboring nodes, and the node features of the node and its neighboring nodes are weighted and aggregated based on the importance weight to update the node features; feature fusion and global pooling are performed on the node features of multiple nodes in the updated spatiotemporal graph structure to obtain the first fused state features of the vehicle; and the suspension risk trend of at least one wheel in the future time step is determined based on the first fused state features.

[0063] The node feature update mechanism aims to dynamically evaluate the correlation strength between a node and its neighboring nodes through an attention mechanism, and then weight and aggregate the features of neighboring nodes based on the correlation strength. This more effectively captures the local dependencies between nodes and their neighbors and updates the node representation. This allows each node to incorporate the node features of its most relevant neighboring nodes, rather than simply averaging them.

[0064] The first fused state feature is a high-level abstract representation obtained by spatiotemporal attention encoding of dynamic data from multiple vehicle components. It comprehensively reflects the dynamic changes of vehicle components in time and space, including key information about the interactions between vehicle components at different times and between different sensors. Specifically, the first fused state feature can be a high-dimensional vector, where each dimension encodes the importance of a specific spatiotemporal pattern in the dynamic data of multiple vehicle components. In other words, the first fused state feature can also be a feature matrix, capable of capturing the evolution of vehicle components with greater precision.

[0065] In this implementation, by converting vehicle dynamic state data and vehicle component static data into a structured spatiotemporal graph, and utilizing an attention mechanism to dynamically capture the physical coupling and temporal evolution relationships between nodes, accurate modeling of the vehicle's spatiotemporal dynamics is achieved. This refined feature extraction method enables the obtained spatiotemporal fusion state features to more comprehensively and accurately characterize the vehicle's complex motion state, thereby providing high-quality input for determining the subsequent wheel suspension risk trend and improving the accuracy and reliability of the suspension risk trend.

[0066] To provide a clearer explanation of the above embodiments, the following description is divided into several parts.

[0067] The first part describes how, for any node in the spatiotemporal graph structure of the multiple vehicle component dynamic data, the importance weight is determined based on the importance influence between the node and its neighboring nodes, and the node features of the node and its neighboring nodes are weighted and aggregated based on the importance weight to update the content of the node features.

[0068] In some embodiments, a Graph Attention Network (GAT) mechanism can be used to weighted aggregate the node features of a node and its neighbors. Specifically, for a node, its importance weight can be obtained by calculating the importance influence of the node features of each neighbor node on the node features of that node, and then normalizing it using a softmax function. The updated node features are a weighted sum of its own features and the features of all neighbor nodes.

[0069] In some embodiments, the formula for determining importance weights is shown below.

[0070]

[0071] in, Represents a node For neighboring nodes Importance weights A larger value indicates that the neighboring nodes are more important to this node. All weights sum to 1; For nodes with neighboring nodes The original importance scores between them This is used to turn importance scores into positive numbers, amplify the differences in importance scores, and give higher weight to more important neighboring nodes; For nodes The set of all neighboring nodes; For nodes The summation of the exponential attention scores of all neighbors.

[0072] In some embodiments, the calculation formula for weighted aggregation of the node features of the node and its neighboring nodes based on the importance weight to update the node features of the node is shown below.

[0073]

[0074] in, Represents a node The new feature at time t, that is, The updated node features; Represents a nonlinear activation function; This indicates that the node features of the node and its neighboring nodes are weighted and aggregated based on this importance weight; This means that the node features of neighboring nodes are first subjected to a linear transformation to obtain the transformed node features; Represents the original characteristics of neighboring nodes. A linear transformation is performed to map the node features to a new dimensional space, resulting in a learnable weight matrix. Representing neighboring nodes The original eigenvector at time t; Indicates traversing nodes All neighboring nodes .

[0075] The second part explains the content of the first fused state features of the vehicle obtained by performing feature fusion and global pooling on the node features of multiple nodes in the updated spatiotemporal graph structure.

[0076] Feature fusion and global pooling of node features aim to integrate all node features updated by the local attention mechanism into a global feature vector that represents the current spatiotemporal state of the vehicle components. In some embodiments, node features are represented in the form of feature vectors. Feature fusion can further refine information, while global pooling compresses the variable-length spatiotemporal graph structure information into a fixed-length representation, facilitating subsequent model processing.

[0077] In some embodiments, feature fusion can employ a multilayer perceptron (MLP) to perform a nonlinear transformation on all updated node features to extract higher-level abstract features. Global pooling can use average pooling or max pooling to aggregate the node features of all nodes along a certain dimension, such as averaging or maximizing the corresponding dimension of all node features. Alternatively, more complex pooling strategies can be used, such as learning-based pooling methods (e.g., DiffPool, SAGPool), which can adaptively select and aggregate important information based on the structure of the spatiotemporal graph and the node features. The fusion process can also be a gating mechanism, such as using a gated recurrent unit (GRU) or a long short-term memory network (LSTM). Term Memory (LSTM) is used to fuse node features.

[0078] The third part explains the trend of the suspension risk of at least one wheel in the future time step based on the characteristics of the first fusion state.

[0079] In one possible implementation, the state evolution characteristics of the vehicle components of the vehicle in a future time step are determined based on the graph evolution neural network and the first fused state characteristics. The state evolution characteristics are used to characterize the state trend of the vehicle components of the vehicle over time. The air suspension risk trend of at least one wheel in a future time step is determined based on the state evolution characteristics and the first fused state characteristics.

[0080] Graph evolutionary neural networks are neural network architectures that can simultaneously handle graph structures (i.e., the coupling relationships between vehicle components) and temporal evolution (the dynamic data of vehicle components changing over time). In graph evolutionary neural networks, the update of the node state in each layer depends on both the spatial aggregation of neighboring nodes and the historical state of the current node.

[0081] The first fused state feature is the node representation at the current time step, obtained after processing by a graph neural network, which integrates the node's own node features and the features of its neighboring nodes. However, the first fused state feature does not include the evolution of node features over time. In this case, time-dimensional modeling of the first fused state feature yields the state evolution feature of the vehicle component, which describes how node features change over time. This state evolution feature is typically a time-series sequence or recursive state, capturing the changing patterns of node features. The levitation risk trend is the probability of a wheel being levitation in a future time step, predicted based on the state evolution feature and the current first fused state feature. For example, both wheels of a vehicle currently have a coupling strength of 0.6, but one wheel's current coupling strength decreases from 0.9 to 0.6, while the other wheel's current coupling strength increases from 0.4 to 0.6. The changes in the two wheels of the vehicle in a future time step are completely different. The current state (first fused state feature) cannot distinguish between these two situations, nor can it determine which wheel is about to be levitation. The current state evolution characteristics provide the changes of vehicle components within a historical window, so the future risks of vehicle components can be judged by the state evolution characteristics in future time steps.

[0082] For example, for a vehicle at time t1, based on the vehicle's first fusion state characteristics, it can be determined that the coupling strength between the right front wheel and the right front suspension is 0.92, meaning that the right front wheel and the right front suspension are in a high-strength coupling; the coupling strength between the right front suspension and the vehicle body is 0.88, meaning that the right front suspension and the vehicle body are in a high-strength coupling; and the coupling strength between the right front wheel and the left rear wheel is 0.15, meaning that the right front wheel and the left rear wheel are in a weak coupling. Based on the above first fusion state characteristics, it can be determined that the vehicle body is in a stable state. Therefore, based on these first fusion state characteristics, it can be determined that at time t1+ The state evolution characteristics after time t indicate that the vehicle is in a stable state, thus the risk of the vehicle becoming airborne is low. For the vehicle at time t1+3... In case t, based on the vehicle's first fusion state characteristics, the coupling strength between the vehicle's right front wheel and right front suspension can be determined to be 0.78. That is, compared to time t1, the vehicle at time t1+3... At time t, the coupling strength between the right front wheel and the right front suspension gradually decreases; the coupling strength between the vehicle's right front suspension and the body is 0.72, meaning that compared to time t1, the vehicle at time t1+3... At time t, the coupling strength between the right front suspension and the vehicle body gradually decreases; the coupling strength between the right front wheel and the left rear wheel is 0.48, meaning that compared to time t1, the coupling strength between the right front wheel and the left rear wheel at time t1+3 is 0.48. At time t, the coupling strength between the right front wheel and the left rear wheel gradually increases; therefore, based on the characteristics of this first fusion state, it can be determined that at time t1+4... The state evolution characteristics after t indicate that the coupling strength between vehicle components begins to change, thus increasing the risk of the vehicle becoming airborne.

[0083] In this implementation, based on the graph evolution neural network and the first fusion state features, the state evolution characteristics of vehicle components over time are accurately determined, and the trend of wheel suspension risk in future time steps is inferred accordingly. That is, the temporal evolution law of vehicle components is determined by the graph evolution neural network and the first fusion state features. Instead of making an instantaneous judgment on the current state, it makes a forward-looking prediction on the time axis by combining historical trends, the first fusion state features, and state evolution characteristics. Thus, based on the state evolution characteristics in future time steps, it makes an advanced prediction and trend forecast of wheel suspension risk, breaking through the lag limitation of traditional judgment based on instantaneous signals, and significantly improving the timeliness and accuracy of risk identification.

[0084] To provide a clearer explanation of the above embodiments, the following description is divided into several parts.

[0085] The first part explains the content of determining the state evolution characteristics of the vehicle components in future time steps based on the graph evolution neural network and the first fused state features.

[0086] In some embodiments, historical test data of the vehicle under various target conditions are collected; for each target condition, the type of disturbance experienced by the vehicle in the historical test data is labeled; a training set is constructed based on the disturbance type, the historical first fusion state features extracted from the historical test data, and the suspended timestamp; a multi-task loss function is used to iteratively optimize the graph evolution neural network based on the training set to obtain the trained graph evolution neural network.

[0087] The target operating condition refers to the condition under which the vehicle's wheels may tend to become airborne. For example, the target operating condition could be a jump condition, a steep slope crossing condition, a curb impact condition, or a unilateral road surface collapse condition. Historical test data refers to the dynamic data of vehicle components collected within a historical time step, as well as the fused historical first-stage state characteristics. The airborne timestamp is used to record the precise moment when each wheel of the vehicle becomes airborne under the target operating condition.

[0088] In some embodiments, the current first fused state features are input into the trained graph evolution neural network to deduce the state evolution features of the vehicle component within future time steps. These state evolution features are used to predict which wheels in the vehicle will soon be suspended in the air, rather than representing an instantaneous state of the current state.

[0089] In some embodiments, the formula for deriving the state evolution characteristics of the vehicle component within a future time step is shown below.

[0090]

[0091] in, Used to characterize from time step Time to step A continuous graph sequence, total Each time step. It is the first fusion state feature of the current time step. Used to characterize the length of the history window. Used to characterize future time steps. Used to characterize the future The state evolution characteristics at time step F( The state evolution mapping function is used to characterize the state evolution mapping function learned in a graph evolution neural network. It is a non-linear transformation that maps the historical graph sequence to the future graph state.

[0092] Secondly, graph evolutionary neural networks use continuous graph sequences As input, a state evolution mapping is learned. In the above deduction process, the state representation error between mapping historical graph sequences to future graph states is converged. For example, from... convergence to ,in, .

[0093] The second part explains the trend of the suspension risk of at least one wheel in the future time step based on the state evolution characteristics and the first fusion state characteristics.

[0094] In one possible implementation, the state evolution feature and the first fused state feature are mapped to a unified feature space, and the state evolution feature and the first fused state feature are fused to obtain a second fused state feature; the suspension risk trend of the at least one wheel in the future time step is determined based on the second fused state feature.

[0095] The second fused state feature is a feature representation obtained by combining the first fused state feature with dynamic state information related to vehicle dynamics. It is used to provide comprehensive information about the current state and historical trends of the vehicle for predicting the risk of air suspension. For example, the second fused state feature can be a high-dimensional vector containing dynamic parameters such as wheel load information of vehicle components, wheel speed, body suspension, roll angle, and yaw rate, as well as node evolution process information extracted through a graph evolution neural network. The state evolution feature and the first fused state feature are embedded into feature vectors of uniform length through a multi-channel structure mapping network. Fusion is the process of combining two or more information sources to form a more comprehensive information representation, used to integrate different types of information to provide richer and more comprehensive input to subsequent processing modules. In this embodiment, splicing, cross-attention matching, and high-order feature modeling are performed in the intermediate fusion layer. This fusion layer uses a residual structure and a multi-head cross-attention mechanism, enabling the state evolution feature to provide guided correction to the first fused state feature. A simple splicing operation can also be used to connect the first fused state feature and the state evolution feature in terms of dimension. Alternatively, weighted summation or neural network layers (such as fully connected layers) can be used for feature interaction and fusion.

[0096] In this implementation, the state evolution features and the first fused state features are mapped to a unified feature space and fused to obtain the second fused state features. Based on this, the wheel suspension risk trend is determined, thereby realizing the deep coupling between the real-time dynamic state of the vehicle components and the spatiotemporal evolution law of the vehicle components. This significantly improves the completeness and robustness of feature expression. Based on this second fused state feature, risk judgment is performed, which effectively enhances the perception accuracy of complex dynamic scenes and ensures the accuracy and reliability of suspension risk trend prediction.

[0097] To provide a clearer explanation of the above embodiments, the following description is divided into several parts.

[0098] The first part explains the content of obtaining the vehicle's current power status data.

[0099] It is understood that the specific implementation method for obtaining the current power status data of the vehicle can be found in the relevant description of step 301 above, and will not be repeated here.

[0100] The second part explains the process of mapping the state evolution feature and the first fused state feature to a unified feature space, and performing feature fusion on the state evolution feature and the first fused state feature to obtain the second fused state feature.

[0101] Normalizing and embedding the state evolution features and the first fused state features involves mapping three types of features—those from different sources, with different physical meanings, different numerical dimensions, and different original dimensions—to the same high-dimensional continuous vector space through a learnable nonlinear transformation. Each fused feature is then given a clear semantic context, enabling the posterior sampling reinforcement learning (PSRL) model to understand its physical meaning. This provides a unified input representation for subsequent feature fusion, essentially providing the second fused state features to the PSRL model. The PSRL model can integrate reinforcement learning algorithms with Bayesian optimization principles within each training iteration, thus balancing sampling efficiency and policy iteration efficiency. This makes it suitable for application scenarios with uncertain perturbations and for action planning and decision-making under risk preference constraints.

[0102] The third part explains the trend of the suspension risk of at least one wheel in the future time step based on the characteristics of the second fusion state.

[0103] In one possible implementation, it is determined whether the wheel load of each wheel in the second fusion state feature is less than a first threshold within a first future time window, the first future time window corresponding to a recent time period requiring immediate response; an abnormal increase rate of wheel speed of each wheel within the first future time window is determined; an abnormal elongation rate of the vehicle suspension in the second fusion state feature within the first future time window is determined; a decreasing trend of wheel load in the second fusion state feature within the first future time window is determined; and a suspension risk trend of at least one wheel is determined based on whether the wheel load is less than the first threshold, the abnormal increase rate, the abnormal elongation rate, and the decreasing trend.

[0104] The first future time window is the time period requiring immediate response, used to identify and address impending risks. The second fused state feature is a comprehensive feature vector that integrates the first fused state feature, state evolution feature, and current dynamic state data. Wheel load is the normal contact force (vertical force) between the wheel and the ground. The first threshold is a threshold automatically determined by the driver assistance controller; this embodiment does not limit this. If the wheel load is less than the first threshold, it indicates that the wheel is very close to being suspended in the air. Wheel load is usually the most direct evidence of wheel suspension, but it can only detect wheels that are about to be suspended and cannot predict them in advance. Therefore, it is necessary to combine other parameters to determine the risk trend of wheel suspension. Wheel speed represents the angular velocity or linear velocity of wheel rotation. When the vehicle is driving normally, wheel speed is equivalent to the vehicle speed. Abnormal increase in speed indicates that the wheel speed suddenly increases significantly in a very short time (e.g., 50ms) and exceeds the vehicle speed by a certain proportion (e.g., more than 20%). Wheel speed greater than vehicle speed indicates that the wheel is slipping or spinning freely, and a sudden increase in wheel speed indicates that the wheel is losing ground adhesion. The vehicle suspension characterizes the elastic / damping elements connecting the wheels to the vehicle body. Suspension extension or compression reflects the wheel's motion relative to the vehicle body. Abnormal extension rate characterizes the speed at which the suspension changes from a compressed to an extended state. Extension indicates that the suspension is lengthening, and the wheel is moving downwards relative to the vehicle body. A sudden, rapid extension of the suspension indicates that the wheel is rapidly losing contact with the vehicle body; that is, a direct mechanical manifestation before wheel lift-off. The wheel load decline trend characterizes a continuous decrease in wheel load over time at a negative rate. When the wheel load exceeds a first threshold, the wheel load decline trend can indicate the risk of wheel lift-off. For example, if the wheel load decline trend is a brief decrease followed by a rapid rebound, it indicates the vehicle is under normal bumpy conditions. If the wheel load decline trend is a continuous decrease at an increasingly faster rate, it indicates that at least one wheel is at risk of lift-off.

[0105] In some embodiments, when the wheel load is less than a first threshold, the abnormal increase rate of the wheel speed within a first future time window is greater than a second threshold, the abnormal elongation rate of the vehicle suspension is greater than a third threshold, and the decreasing trend of the wheel load is greater than a fourth threshold, the wheel's suspension risk trend is determined to be a high suspension trend.

[0106] In some embodiments, the wheel suspension risk trend is determined to be a low suspension trend when the wheel load is less than or equal to a first threshold, the abnormal increase rate of the wheel speed within a first future time window is less than or equal to a second threshold, the abnormal elongation rate of the vehicle suspension is less than or equal to a third threshold, and the decreasing trend of the wheel load is less than or equal to a fourth threshold.

[0107] In this implementation, by comprehensively monitoring multi-dimensional dynamic indicators such as whether the wheel load is below a threshold, the abnormal increase rate of wheel speed, the abnormal elongation rate of the suspension, and the decreasing trend of wheel load within the first future time window corresponding to the immediate response period, the risk trend of wheel suspension can be accurately determined. This avoids determining the risk trend of wheel suspension based on a single static threshold. By utilizing multi-physical quantity coupling analysis, it can effectively identify transient suspension symptoms of the vehicle at the moment of severe bumps or landing, significantly reducing the false alarm rate and missed alarm rate under complex road conditions. This allows for earlier prediction of wheel suspension risk and greatly improves the sensitivity and accuracy of suspension risk identification, providing response time for the vehicle chassis system's active safety intervention and ensuring driving safety.

[0108] Step 303: Based on the suspension risk trend and the vehicle's current power status data, the driver assistance controller generates torque adjustment commands for the corresponding wheels.

[0109] The hovering risk trend characterizes the risk of a wheel becoming airborne. Current power state data characterizes the current state of the vehicle's powertrain, such as the vehicle's drive mode, torque request, vehicle speed, and estimated tire grip coefficient. Torque adjustment commands are used to increase, maintain, or decrease the drive torque of that wheel. In some embodiments, the driver assistance controller sends torque adjustment commands to the motor controller (MCU) or engine control unit (ECU). A generation representation maps the input hovering risk trend and power state data to specific torque command values ​​using a preset optimization algorithm. The corresponding wheel is the wheel with a hovering risk trend, that is, the wheel that is about to actually become airborne.

[0110] In one possible implementation, a corresponding adjustment factor is determined based on the air suspension risk trend. This adjustment factor is used to determine the distribution ratio of the drive torque of the corresponding wheel. A torque adjustment command for the corresponding wheel is generated based on the air suspension risk trend, the current power state data, and the adjustment factor.

[0111] It should be understood that before the wheel actually becomes airborne, the driver assistance controller needs to actively and appropriately reduce the drive torque to avoid the impact of wheel spin during the moment of airborne suspension. At the same time, after the airborne risk trend subsides, the drive torque should be restored in a timely manner to ensure that the vehicle's power output is not interrupted. In this case, the corresponding adjustment factor is determined based on the airborne risk trend.

[0112] The "levitation risk trend" describes the likelihood and direction of future wheel levitation. The adjustment factor is a coefficient between 0 and 1; for example, a factor of 0 indicates zero drive torque, and a factor of 1 indicates full drive torque. The adjustment factor determines the proportion of drive torque that wheel should receive. The drive torque distribution ratio is the percentage of total drive torque allocated to that wheel. Typically, each wheel in a four-wheel drive vehicle has an independent torque command.

[0113] In this implementation, the adjustment factor is dynamically determined based on the risk of wheel slippage, and a torque adjustment command is generated, enabling active, on-demand distribution of drive torque. This solution can quickly suppress wheel slippage before the wheel actually slips off the ground, significantly improving the vehicle's power response speed and ensuring vehicle stability and safety.

[0114] To provide a clearer explanation of the above embodiments, the following description is divided into several parts.

[0115] The first part explains the content of determining the corresponding adjustment factors based on this suspended risk trend.

[0116] In some embodiments, the corresponding adjustment factor is determined based on the suspending risk trend and the target control mode.

[0117] The wheel lift risk trend is a disturbance intensity score, i.e., a score indicating the intensity of the wheel lift risk. The target control mode characterizes the intensity of the adjustment to the corresponding wheel's drive torque. For example, target modes include no adjustment (no adjustment at all, maintaining full torque), light adjustment (retaining 85% torque, slightly reducing torque), moderate adjustment (retaining 60% torque, significantly reducing torque), heavy adjustment (retaining 35% torque, drastically reducing torque), and emergency adjustment (retaining 10% torque, almost cutting off torque).

[0118] A weighted mapping is performed between the lingering risk trend and the target control mode. This means that a adjustment factor is generated through a weighted mapping function. For example, weighted mapping functions include linear weighted mapping functions, sigmoid weighted mapping, piecewise linear mapping, etc. Generating the adjustment factor through a weighted mapping function, rather than directly relying on a threshold, avoids abrupt changes in the adjustment factor, making the generation of the adjustment factor more continuous and smooth, and allowing for the configuration of different lingering risk trends.

[0119] The second part explains the content of the torque adjustment command for the corresponding wheel generated based on the suspension risk trend, the current dynamic state data, and the adjustment factor.

[0120] In one possible implementation, a reward function is determined based on the current wheel load recovery speed, vehicle acceleration stability, vehicle traction efficiency, and output smoothness of wheel output torque in the current dynamic state data; and a torque adjustment command for the corresponding wheel is generated based on the reward function, the suspension risk trend, and the adjustment factor.

[0121] Among them, the current wheel load recovery speed is used to characterize the rate at which the wheel load recovers from a trough to a normal value; the vehicle acceleration stability is used to characterize the time-domain fluctuation of the vehicle's longitudinal and lateral acceleration; the vehicle traction efficiency is the ratio of the vehicle's actual effective traction force to the theoretically required output traction force; the output smoothness of the wheel output torque is used to characterize the smoothness of the torque command changing over time; and the reward function is used to characterize the control effect of controlling the drive torque, with a higher reward function indicating a more ideal control effect.

[0122] In some embodiments, a reward function is obtained by weighted summing of the current wheel load recovery speed, vehicle acceleration smoothness, vehicle traction efficiency, and the smoothness of wheel output torque. This reward function does not only consider whether the torque adjustment command prevents wheel spin, but also comprehensively considers four dimensions: load recovery speed, acceleration smoothness, traction efficiency, and control smoothness, to find the optimal torque adjustment command. High-speed wheel rotation under no-load conditions can lead to wasted drive torque and even wheel loss of control. In this case, the torque adjustment command prevents wheel spin, and the reward function provides positive feedback to incentivize timely reduction of drive torque when the wheel is about to become airborne. However, if the torque adjustment command reduces drive torque too early, it can cause a delay in wheel power, resulting in unnecessary traction loss. In this case, the reward function provides negative feedback to suppress excessive protection of the wheel's drive torque. Therefore, in practical applications, the driver assistance controller needs to find the optimal balance between timely adjustment of drive torque to prevent wheel spin and avoiding power loss due to erroneous adjustment of drive torque; that is, it needs to find the optimal torque adjustment command through the reward function.

[0123] In this implementation, a reward function is constructed by comprehensively considering wheel load recovery speed, acceleration stability, traction efficiency, and output smoothness. The adjustment factor is dynamically determined based on the suspension risk trend, and a torque adjustment command is generated. This enables the active on-demand allocation of drive torque, thereby effectively suppressing suspension risk while taking into account the smoothness of vehicle driving and power transmission efficiency, avoiding torque abrupt shocks, and significantly improving the vehicle's handling stability and ride comfort.

[0124] To provide a clearer explanation of the above implementation method, the following describes the generation of the torque adjustment command for the corresponding wheel based on the reward function, the suspension risk trend, and the adjustment factor.

[0125] In one possible implementation, the probability distribution of the torque adjustment command for the corresponding wheel at the current time step is determined based on the suspension risk trend; the execution time of the torque adjustment command is optimized with the goal of maximizing the expected future return, and the expected future return is the expected value of the sum of the reward functions obtained from the current moment to all future moments; the optimal torque adjustment command is generated based on the optimal execution time, the suspension risk trend, and the adjustment factor.

[0126] Here, the current time step is the current moment when the torque adjustment command is generated. The torque adjustment command is the final drive torque value applied to the wheel at risk of being airborne. The probability distribution represents the probability of selecting the numerical variable of the torque adjustment command under the airborne risk trend. The predicted airborne risk trend still has uncertainty; the probability distribution can represent this uncertainty, and sampling or taking the expected value from the probability distribution will yield a smoother torque adjustment command. The expected future reward is the expected value of the cumulative sum of rewards from the current moment to all future moments. The optimization objective is to maximize the expected future reward. The execution time of the torque adjustment command is the time required from the issuance of the torque adjustment command to its complete execution. The optimal execution time is the execution time that maximizes the expected future reward. The reward function characterizes the control effect of controlling the drive torque.

[0127] In some embodiments, the formula for generating the optimal torque adjustment command is shown below, which inputs the levitation risk trend into the posterior sampling reinforcement learning model.

[0128]

[0129] in, The levitation risk trend is used to characterize the rate of change of the current levitation risk of the wheel. For discrete-time index (current time step); This is the numerical variable for the torque adjustment command; A numerical variable used to represent the probability of selecting a torque adjustment command under a suspending risk trend; The optimal probability of selecting the torque adjustment command under the risk of suspension is represented by a numerical variable. Expected future return represents the expected value of the sum of all future rewards from the current moment onwards; The goal is to optimize the search for the objective that maximizes the expected cumulative reward. The optimal execution time is used to characterize the execution time parameters included in the torque adjustment command selected under the levitation risk trend. The reward function is used to characterize the reward at time step [step]. The instant reward signal received; Let be the expected value, used to account for randomness in the average cumulative reward. Furthermore, in the posterior sampling reinforcement learning model, the dangling risk trend is not input, where the dangling risk trend is... .

[0130] In this implementation, by determining the probability distribution of torque adjustment commands and optimizing the execution time with the goal of maximizing future expected returns, not only is the output of drive torque suppressed before the wheels actually become airborne, but the optimal balance point is also found between suppressing the risk of airborne failure and maintaining power output. This avoids power loss or response lag caused by excessive intervention, thereby significantly improving the vehicle's adaptability and driving smoothness under complex dynamic conditions, and ensuring the efficiency and safety of power distribution.

[0131] Understandably, the posterior sampling reinforcement learning model also outputs a global confidence score. The global confidence score is used to indicate whether the currently generated torque adjustment command needs the posterior sampling reinforcement learning model to participate in the adjustment update. If the global confidence score is too low, the posterior sampling reinforcement learning model is triggered to regenerate the torque adjustment command in order to avoid the stability risk caused by low-quality torque adjustment command control.

[0132] Step 304: The driver assistance controller adjusts the drive torque of the corresponding wheel based on the torque adjustment command.

[0133] In some embodiments, the driver assistance controller receives torque adjustment commands via the CAN bus, parses the torque adjustment commands, performs security verification on the parsed torque adjustment commands, maps and converts the torque adjustment commands to obtain the torque allocation of the motor corresponding to each wheel; based on the torque allocation, it controls the opening of the proportional valve to control the drive torque of the corresponding wheel, and feeds back the actual torque, execution status, etc. to the upper-level module.

[0134] The torque adjustment command includes fields such as target torque, rate of change, execution mode, and confidence level.

[0135] This application provides a wheel control method. This method can acquire real-time dynamic data of vehicle components through an auxiliary driving controller, accurately predict the risk trend of wheel slippage within a future time step, and generate torque adjustment commands based on this and the current power state, achieving active on-demand distribution of drive torque. That is, it can proactively intervene before the wheel actually slips, effectively suppressing power loss and vehicle instability. By dynamically adjusting the drive torque, the vehicle's passability and driving stability under complex conditions are significantly improved, while avoiding energy waste caused by wheel spin, enhancing the response efficiency and safety of the entire vehicle's power system, and providing users with a smoother and more reliable driving experience.

[0136] After introducing the wheel control method provided in the embodiments of this application, the following is a schematic flowchart of the model construction method provided in the embodiments of this application. (See attached diagram.) Figure 4 Taking the driver assistance controller as the executing entity as an example, the method includes the following steps.

[0137] Step 401: Collect vehicle dynamic status data and vehicle component static data, and construct multiple vehicle component dynamic data.

[0138] It is understood that the specific implementation of step 401 can be found in the relevant descriptions of steps 201 and 301 above, and will not be repeated here.

[0139] After completing step 401, you can proceed directly to step 403 or step 402 depending on the actual situation.

[0140] Step 402: Construct a graph attention network based on the dynamic data of vehicle components.

[0141] It is understood that the specific implementation of step 402 can be found in the relevant description of step 302 above, and will not be repeated here.

[0142] Training labels are generated based on dynamic data of vehicle components under different operating conditions to construct a learning sample set. For example, real-vehicle data under typical operating conditions such as vehicle speed bumps, off-road obstacles, and tire slippage are collected, and classification labels are manually labeled. A weighted loss function is used to set training constraints, such as increasing the loss weight for risk omissions and misjudgments, and prioritizing the identification of potential driving hazards during training. During the training of the graph attention network, some node features of the spatiotemporal graph structure in the dynamic data of vehicle components are randomly masked to simulate scenarios of vehicle sensor failure and data interference, thereby enhancing the fault tolerance of the graph attention network. Input vehicle component dynamic data is processed through attention weighted aggregation operations to iteratively update the parameters of the graph attention network. After training convergence, the node features and attention weight matrices of each node in the spatiotemporal graph structure are output, and the first fused state features are generated based on the fused node features and attention weight matrices.

[0143] Step 403: Construct a graph evolutionary neural network based on the dynamic data of vehicle components.

[0144] It is understood that the specific implementation of step 403 can be found in the relevant description of step 302 above, and will not be repeated here.

[0145] It should be understood that during normal driving, the wheels of a vehicle will not be suspended in the air. However, when the vehicle is subjected to road disturbances, the wheels will actually be suspended in the air. In this case, a dedicated training dataset is built around the phenomenon of wheel suspension, which labels various road disturbances such as jumps, steep slopes, and road impacts, and records the changes in the graph structure state simultaneously. The actual time of suspension is used as the basis for labeling to carry out supervised training.

[0146] Historical test data of vehicles under various target operating conditions were collected. For each target operating condition, the type of disturbance experienced by the vehicle within the historical test data was labeled. A training set was constructed based on the disturbance type, the historical first fusion state features extracted from the historical test data, and the levitation timestamp. A multi-task loss function was used to iteratively optimize the graph evolution neural network based on this training set, resulting in the trained graph evolution neural network. To combat the amplification of disturbance samples, node noise and edge noise in the spatiotemporal graph structure were actively added during the training of the graph evolution neural network. The graph evolution neural network was iteratively trained repeatedly until it could accurately predict the levitation trend of the vehicle under different operating conditions, at which point the training was terminated.

[0147] Step 404: Construct a posterior sampling reinforcement learning model based on the graph evolutionary neural network and the graph attention network.

[0148] The risk of a vehicle's wheels becoming airborne is a gradual, temporal process, also influenced by the structural constraints between vehicle components. A single-dimensional judgment is insufficient. In this case, it is necessary to construct a feature fusion module that takes into account both temporal and spatial structures, that is, to construct a posterior sampling reinforcement learning model.

[0149] The posterior sampling reinforcement learning model is trained using a joint loss function (attention loss + prediction loss + control loss). The training data integrates information from simulations and real-vehicle data, including hovering start time, duration, power response, and slippage rate, ensuring that the model learns effective and optimal torque adjustment commands under realistic dynamic conditions. This model is adaptable to various powertrain architectures, including in-wheel motors, central motors, and hydraulic drives.

[0150] In practical applications, situations where vehicle wheels are suspended in the air are rare, and the probability of such occurrences is low. Resampling every time is extremely costly, making it difficult for posterior sampling reinforcement learning models to learn sufficient experience. Furthermore, continuously collected vehicle driving data is highly correlated, and directly using it for training leads to overfitting and extremely low learning efficiency. Therefore, the collected vehicle component dynamic data, torque adjustment commands, and reward functions are cached in a dynamic memory pool for periodic replay training. This allows for repeated use of rare, critical suspension conditions, avoiding repeated sampling of dangerous scenarios, reducing training costs, and shortening model convergence time. Since the generation of torque adjustment commands varies significantly under different road conditions, a single posterior sampling reinforcement learning model cannot adapt to all different road conditions and has poor generalization ability. Therefore, multiple posterior sampling reinforcement learning models are trained in parallel using vehicle component dynamic data under different road conditions. These models are then fused to obtain a unified posterior sampling reinforcement learning model. The trained posterior sampling reinforcement learning model, combined with the prediction of the suspension trend by graph modeling, can make actions to adjust the drive torque in advance, rather than waiting for the fault to occur before taking remedial action, thus achieving predictive active control.

[0151] During the training of the posterior sampling reinforcement learning model, an intermittent adversarial joint training method is used. First, a confidence score is learned, which is used to determine whether vehicle control is possible. For example, if the vehicle is traveling on a flat road, the posterior sampling reinforcement learning model is very confident in its predictions, with a high confidence score (e.g., 0.95). In this case, the driving torque of the wheels can be adjusted via torque adjustment commands. However, if a suspended situation occurs and the right front wheel enters a pothole, the posterior sampling reinforcement learning model predicts the risk, but it is still within the range of known experience. The driving torque of the wheel can still be adjusted via torque adjustment commands. However, if the confidence score drops to the lowest confidence threshold, it is determined that adjusting the driving torque of the wheel via torque adjustment commands would be dangerous, and the torque adjustment commands are regenerated.

[0152] Step 405: In response to the torque adjustment command output by the posterior sampling reinforcement learning model, control the vehicle to perform the adjustment operation.

[0153] It is understood that the specific implementation of step 405 can be found in the relevant description of step 304 above, and will not be repeated here.

[0154] Figure 5 This is a schematic diagram of the structure of a wheel control device provided in an embodiment of this application.

[0155] For example, the device 500 includes: The acquisition module 501 is used to acquire dynamic data of multiple vehicle components of the vehicle during the vehicle's operation. The dynamic data of the vehicle components is obtained by constructing a correlation based on the vehicle's dynamic state data and the static data of the vehicle components. The dynamic data of the vehicle components is used to represent the dynamic coupling strength between the vehicle components. The multiple dynamic data of the vehicle components are generated continuously. The determination module 502 is used to determine the air suspension risk trend of at least one wheel of the vehicle within a future time step based on the dynamic data of the multiple vehicle components. The generation module 503 is used to generate torque adjustment commands for the corresponding wheels based on the risk of suspension and the current power status data of the vehicle, so as to adjust the drive torque in advance before the wheel actually suspends.

[0156] In some embodiments, the acquisition module 501 is specifically used for: acquiring vehicle dynamic state data and vehicle component static data of the vehicle; determining multiple nodes based on the vehicle component static data at different sampling times, wherein the nodes are used to characterize the attributes of the vehicle components of the vehicle; mapping each vehicle dynamic state data to a spatial edge of the node based on the physical layout and functional association of the vehicle components, wherein the spatial edge is used to characterize the physical coupling relationship between the nodes; constructing a spatiotemporal graph structure based on the node and the spatial edge at different sampling times based on the chronological order of the sampling times, wherein the edges of the spatiotemporal graph structure are used to characterize the state transition relationship between the nodes across time steps; and constructing the spatiotemporal graph structure of multiple consecutive time steps in chronological order to form multiple vehicle component dynamic data.

[0157] In some embodiments, the determining module 502 is specifically configured to: for any node in the spatiotemporal graph structure of the multiple vehicle component dynamic data, determine an importance weight based on the importance influence between the node and its neighboring nodes, and perform weighted aggregation of the node features of the node and its neighboring nodes based on the importance weight to update the node features of the node; perform feature fusion and global pooling on the node features of multiple nodes in the updated spatiotemporal graph structure to obtain the first fused state features of the vehicle; and determine the suspension risk trend of at least one wheel in a future time step based on the first fused state features.

[0158] In some embodiments, the determining module 502 is specifically configured to: determine the state evolution characteristics of the vehicle components of the vehicle in a future time step based on the graph evolution neural network and the first fused state characteristics, the state evolution characteristics being used to characterize the state trend of the vehicle components of the vehicle over time; and determine the air suspension risk trend of the at least one wheel in a future time step based on the state evolution characteristics and the first fused state characteristics.

[0159] In some embodiments, the determining module 502 is specifically used to: map the state evolution feature and the first fused state feature to a unified feature space, and fuse the state evolution feature and the first fused state feature to obtain a second fused state feature; and determine the suspension risk trend of the at least one wheel in a future time step based on the second fused state feature.

[0160] In some embodiments, the determining module 502 is specifically configured to: determine whether the wheel load of each wheel in the second fusion state feature is less than a first threshold within a first future time window, the first future time window corresponding to a recent time period requiring immediate response; determine the abnormal increase rate of the wheel speed of each wheel within the first future time window; determine the abnormal elongation rate of the vehicle suspension in the second fusion state feature within the first future time window; determine the decreasing trend of the wheel load in the second fusion state feature within the first future time window; and determine the suspension risk trend of the at least one wheel based on whether the wheel load is less than the first threshold, the abnormal increase rate, the abnormal elongation rate, and the decreasing trend.

[0161] In some embodiments, the generation module 503 is specifically used to: determine a corresponding adjustment factor based on the air suspension risk trend, the adjustment factor being used to determine the distribution ratio of the driving torque of the corresponding wheel; and generate a torque adjustment command for the corresponding wheel based on the air suspension risk trend, the current power state data, and the adjustment factor.

[0162] In some embodiments, the generation module 503 is specifically used to: determine a reward function based on the current wheel load recovery speed, vehicle acceleration stability, vehicle traction efficiency, and output smoothness of wheel output torque in the current dynamic state data; and generate a torque adjustment command for the corresponding wheel based on the reward function, the suspension risk trend, and the adjustment factor.

[0163] In some embodiments, the generation module 503 is specifically used to: determine the probability distribution of the torque adjustment command for the corresponding wheel at the current time step based on the suspension risk trend; optimize the execution time of the torque adjustment command with the goal of maximizing the expected future return, to obtain the optimal execution time, wherein the expected future return is the expected value of the sum of the reward functions obtained from the current moment to all future moments; and generate the optimal torque adjustment command based on the optimal execution time, the suspension risk trend, and the adjustment factor.

[0164] Figure 6 This is a schematic diagram of the structure of a vehicle provided in an embodiment of this application.

[0165] For example, such as Figure 6As shown, the vehicle 600 includes a memory 601 and a processor 602. The memory 601 stores executable program code 603, and the processor 602 is used to call and execute the executable program code 603 to perform a wheel control method.

[0166] 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 wheel control method provided in embodiments of this application.

[0167] 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.

[0168] It should be understood that the device provided in this embodiment is used to execute the above-described wheel control method, and therefore can achieve the same effect as the above-described implementation method.

[0169] 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.

[0170] 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 microprocessors, etc., and the storage module may be a memory.

[0171] 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 wheel control method provided in the above embodiments.

[0172] 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 wheel control method provided in the above embodiment.

[0173] This embodiment also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned steps to implement a wheel control method provided in the above embodiment.

[0174] 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.

[0175] 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.

[0176] 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.

[0177] 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 method for controlling a wheel, characterized in that, The method includes: During vehicle operation, dynamic data of multiple vehicle components are acquired. The dynamic data of vehicle components is obtained by constructing a correlation based on the vehicle's dynamic state data and the static data of the vehicle components. The dynamic data of vehicle components is used to represent the dynamic coupling strength between the vehicle components. Based on the dynamic data of the multiple vehicle components, the risk trend of at least one wheel of the vehicle being suspended in the air is determined in a future time step. Based on the aforementioned risk of air suspension and the vehicle's current power status data, torque adjustment commands are generated for the corresponding wheels to adjust the drive torque in advance before the wheels actually become air suspension.

2. The method according to claim 1, characterized in that, The acquisition of dynamic data of multiple vehicle components of the vehicle includes: Acquire the vehicle's dynamic status data and the vehicle's component static data; At different sampling times, multiple nodes are determined based on the static data of the vehicle components, and the nodes are used to characterize the attributes of the vehicle components of the vehicle. Based on the physical layout and functional association of the vehicle components, the dynamic state data of each vehicle are mapped to the spatial edges of the nodes, and the spatial edges are used to characterize the physical coupling relationship between the nodes. Based on the order of sampling times, a spatiotemporal graph structure is constructed at different sampling times based on the nodes and the spatial edges. The edges of the spatiotemporal graph structure are used to characterize the state transition relationship between the nodes across time steps. The spatiotemporal graph structure of multiple consecutive time steps is constructed in a temporal sequence to form dynamic data of multiple vehicle components.

3. The method according to claim 2, characterized in that, The step of determining the air suspension risk trend of at least one wheel of the vehicle within a future time step based on the dynamic data of the multiple vehicle components includes: For any node in the spatiotemporal graph structure of the multiple vehicle component dynamic data, an importance weight is determined based on the importance influence between the node and its neighboring nodes, and the node features of the node and its neighboring nodes are weighted and aggregated based on the importance weight to update the node features of the node. The node features of multiple nodes in the updated spatiotemporal graph structure are fused and global pooled to obtain the first fused state features of the vehicle. Based on the characteristics of the first fusion state, the air suspension risk trend of at least one wheel within a future time step is determined.

4. The method according to claim 3, characterized in that, The step of determining the air suspension risk trend of at least one wheel within a future time step based on the first fusion state characteristics includes: Based on the graph evolution neural network and the first fused state features, the state evolution features of the vehicle components of the vehicle in the future time step are determined, and the state evolution features are used to characterize the state trend of the vehicle components of the vehicle over time. Based on the state evolution characteristics and the first fused state characteristics, the air suspension risk trend of at least one wheel within a future time step is determined.

5. The method according to claim 4, characterized in that, The determination of the air suspension risk trend of at least one wheel within a future time step based on the state evolution characteristics and the first fused state characteristics includes: The state evolution features and the first fused state features are mapped to a unified feature space, and the state evolution features and the first fused state features are fused to obtain the second fused state features. The air suspension risk trend of at least one wheel within a future time step is determined based on the characteristics of the second fusion state.

6. The method according to claim 5, characterized in that, The step of determining the air suspension risk trend of at least one wheel within a future time step based on the second fusion state characteristics includes: Determine whether the wheel load of each wheel in the second fusion state feature is less than a first threshold within a first future time window, where the first future time window corresponds to a recent time period that requires an immediate response; Determine the abnormal increase rate of wheel speed of each wheel within a first future time window; Determine the abnormal elongation rate of the vehicle suspension in the second fusion state feature within the first future time window; Determine the decreasing trend of wheel load in the second fusion state characteristics within the first future time window; The risk trend of at least one wheel being suspended in mid-air is determined based on whether the wheel load is less than a first threshold, the abnormal rise rate, the descent trend, and the abnormal elongation rate.

7. The method according to claim 1, characterized in that, The step of generating torque adjustment commands for the corresponding wheels based on the suspension risk trend and the vehicle's current power state data includes: Based on the aforementioned risk of suspension, a corresponding adjustment factor is determined, which is used to determine the distribution ratio of the driving torque of the corresponding wheel. Based on the suspension risk trend, the current dynamic state data, and the adjustment factor, a torque adjustment command is generated for the corresponding wheel.

8. The method according to claim 7, characterized in that, The process of generating the torque adjustment command for the corresponding wheel based on the suspension risk trend, the current dynamic state data, and the adjustment factor includes: The reward function is determined based on the current wheel load recovery speed, vehicle acceleration stability, vehicle traction efficiency, and output smoothness of wheel output torque in the current dynamic state data. The torque adjustment command for the corresponding wheel is generated based on the reward function, the suspension risk trend, and the adjustment factor.

9. The method according to claim 8, characterized in that, The process of generating the torque adjustment command for the corresponding wheel based on the reward function, the suspension risk trend, and the adjustment factor includes: Based on the aforementioned suspension risk trend, the probability distribution of the torque adjustment command for the corresponding wheel at the current time step is determined; With maximizing the expected future return as the optimization objective, the execution time of the torque adjustment command is optimized to obtain the optimal execution time. The expected future return is the expected value of the sum of the reward functions obtained from the current moment to all future moments. The optimal torque adjustment command is generated based on the optimal execution time, the suspending risk trend, and the adjustment factor.

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