Wheel torque distribution method, device and equipment and vehicle

Abstract

The invention provides a wheel torque distribution method, device and equipment and a vehicle, and is applied to the technical field of vehicle control, and the method comprises the steps: determining the structure perception characteristics representing the torque control priority of each wheel based on the current wheel state data, and determining the torque control priority of each wheel based on the wheel state data of a first preset time period until the current moment; and predicting the state of each wheel in a second preset time period after the current moment to obtain state prediction information of each wheel, performing torque distribution on each wheel based on the structure perception characteristics and the state prediction information, and determining the target torque of each wheel. The key driving wheel, namely the wheel with the highest torque control priority, can be effectively identified, so that the wheel torque is reasonably distributed in time, the intelligent distribution of the wheel torque is realized, the vehicle escape time and energy consumption are reduced, and the control precision is improved.

Description

Technical Field

[0001] This application relates to the field of vehicle control technology, specifically to a wheel torque distribution method, device, equipment, and vehicle. Background Technology

[0002] In existing technologies, the inter-wheel torque of a vehicle, i.e., the torque of each wheel, is typically distributed and controlled based on fixed rules or empirical parameters. In unstructured terrain with complex road conditions (i.e., environments without obvious, fixed, predictable rules, paths, or geometries, such as mud, gravel, cross-axle situations, steep slopes, etc.), some or all of the vehicle's wheels are prone to slippage, suspension, and other abnormalities. In such cases, distributing and controlling the inter-wheel torque based on fixed rules or empirical parameters may result in longer vehicle recovery times (overcoming slippage, suspension, and other abnormalities), higher energy consumption, and lower accuracy in controlling the inter-wheel torque. Summary of the Invention

[0003] Based on the defects and shortcomings of the prior art, this application proposes a wheel torque distribution method, device, equipment and vehicle, which can determine the torque control priority of each wheel and predict the state of each wheel in a second time period based on wheel state data in a first preset time period, so as to distribute the torque of each wheel in a timely and reasonable manner and solve the problems of long vehicle traction time, high energy consumption and low control accuracy.

[0004] According to a first aspect of the embodiments of this application, a wheel torque distribution method is provided, comprising: Based on the current wheel status data, the structural perception features of each wheel are determined, and the structural perception features characterize the torque control priority of the wheel. Based on the wheel state data within a first preset time period, the state of each wheel within a second preset time period is predicted to obtain the state prediction information of each wheel. The first preset time period is a time period lasting a first preset duration up to the current moment, and the second preset time period is a time period lasting a second preset duration after the current moment. Based on the structural perception features and the state prediction information, torque is allocated to each wheel to determine the target torque for each wheel.

[0005] In this way, since wheel status data can accurately reflect whether each wheel is stuck, such as locked or suspended, the torque control priority of each wheel can be accurately determined, and the key drive wheels can be effectively identified. Based on this, combined with the predicted state of the vehicle in the second preset time period after the current moment, i.e., the state prediction information, torque is allocated to each wheel. Compared with torque allocation based on fixed rules or experience data, it can ensure the timeliness and rationality of wheel torque allocation, realize intelligent wheel torque allocation, and reduce vehicle extrication time and energy consumption.

[0006] According to a second aspect of the embodiments of this application, a wheel torque distribution device is provided, comprising: The determination module is used to determine the structural perception features of each wheel based on the current wheel state data, wherein the structural perception features characterize the torque control priority of the wheel. The prediction module is used to predict the state of each wheel in a second preset time period based on the wheel state data in a first preset time period, and to obtain the state prediction information of each wheel. The first preset time period is a time period lasting a first preset duration up to the current moment, and the second preset time period is a time period lasting a second preset duration after the current moment. The allocation module is used to allocate torque to each wheel based on the structural perception features and the state prediction information, and to determine the target torque of each wheel.

[0007] According to a third aspect of the embodiments of this application, an electronic device is provided, including a memory and a processor; The memory is connected to the processor and is used to store programs; The processor is used to implement the wheel torque distribution method as described in the first aspect by running a program in the memory.

[0008] According to a fourth aspect of the embodiments of this application, a storage medium is provided, on which a computer program is stored, and when the computer program is run by a processor, it implements the wheel torque distribution method as described in the first aspect.

[0009] According to a fifth aspect of the present application, a computer program product is provided, the computer program product including computer program instructions, which, when executed by a processor, cause the processor to perform the wheel torque distribution method as described in the first aspect.

[0010] According to a sixth aspect of the embodiments of this application, a vehicle is provided, wherein the vehicle is provided with a wheel torque distribution method as described in the second aspect, or an electronic device as described in the third aspect.

[0011] In the aforementioned wheel torque distribution method, device, equipment, and vehicle, structural sensing features characterizing the torque control priority of each wheel can be determined based on current wheel state data. Wheel state data from a first preset time period up to the current moment is input into a target prediction model to predict the state of each wheel in a second preset time period after the current moment, obtaining state prediction information for each wheel. Based on the structural sensing features and state prediction information, torque is distributed to each wheel to determine the target torque for each wheel. Since wheel state data accurately reflects whether each wheel is stuck (e.g., locked or suspended), the torque control priority of each wheel at the current moment can be accurately determined based on the current wheel state data, effectively identifying the key drive wheel (i.e., the wheel with the highest torque control priority) at the current moment. Based on this, combined with the predicted state of the vehicle in the second preset time period after the current moment (i.e., state prediction information), wheel torque is distributed. Compared to torque distribution based on fixed rules or empirical data, this ensures the timeliness and rationality of wheel torque distribution, reduces vehicle escaping time and energy consumption, and is applicable to various terrains, not limited to fixed rules or empirical data. This improves the control accuracy of torque distribution and achieves intelligent and flexible wheel torque distribution. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0013] Figure 1 This is a schematic flowchart illustrating a wheel torque distribution method according to an embodiment of this application.

[0014] Figure 2 This is a schematic diagram of a wheel torque distribution process provided in an embodiment of this application.

[0015] Figure 3 This is a schematic diagram of the structure of a wheel torque distribution device according to an embodiment of this application.

[0016] Figure 4 This is a schematic diagram of the structure of an electronic device proposed in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] Overview As described in the background section, existing technologies typically distribute and control the inter-wheel torque (i.e., the torque of each wheel) of a vehicle based on fixed rules or empirical parameters. In unstructured terrain with complex road conditions (i.e., environments without obvious, fixed, predictable rules, paths, or geometries, such as mud, gravel, cross-axle situations, steep slopes, etc.), some or all of the vehicle's wheels are prone to slippage, suspension, or other abnormalities. In such cases, distributing and controlling the inter-wheel torque based on fixed rules or empirical parameters may result in longer vehicle recovery times (overcoming slippage, suspension, or other abnormalities), higher energy consumption, and lower accuracy in controlling the inter-wheel torque.

[0019] Further research by the inventors revealed that when wheel torque distribution is based on fixed rules or empirical data, it usually relies on historical data and lacks an understanding of the coupling relationship between wheels. It is difficult to perceive the dynamic changes in complex road conditions in real time and effectively predict the changing trend of wheel status. Therefore, when a vehicle is driving, especially on complex road conditions such as the aforementioned unstructured terrain, when some or all wheels slip or become suspended, the wheel torque distribution may not match the actual needs, the torque distribution response may be lagging, the torque may be continuously output to ineffective wheels, resulting in energy waste and failure to form effective driving force, or the torque output to effective wheels may not meet the needs. This leads to problems such as long vehicle extrication time, high energy consumption, and low control precision.

[0020] Based on this, since wheel state data can characterize abnormal vehicle conditions, the structural perception features that characterize the torque control priority of each wheel can be determined relatively accurately based on the current wheel state data. This effectively identifies the key drive wheel, i.e., the wheel with the highest torque control priority. Based on the wheel state data within the first preset time period up to the current moment, the state of each wheel within the second preset time period after the current moment is predicted, obtaining the state prediction information of each wheel. Based on the structural perception features and state prediction information, torque is allocated to each wheel, and the target torque of each wheel is determined. Compared with torque allocation based on fixed rules or empirical data, this method can allocate wheel torque in a timely and reasonable manner, reducing vehicle extrication time and energy consumption. It is also applicable to various terrains and is not limited to fixed rules or empirical data, thereby improving the control accuracy of torque allocation and realizing the intelligence and flexibility of wheel torque allocation.

[0021] Based on the above concept, this specification provides a wheel torque distribution method, which will be described exemplarily below with reference to the accompanying drawings.

[0022] Exemplary methods Please see Figure 1 In one exemplary embodiment, a wheel torque distribution method is provided, applied to any electronic device capable of communicating with the vehicle controller. This electronic device can be located either inside or outside the vehicle. Specifically, the vehicle is a distributed electric drive vehicle, where each wheel is independently equipped with a complete drive motor, inverter, and reducer, making each wheel an independent drive unit capable of independently controlling its driving and braking forces. Figure 1 As shown, the wheel torque distribution method includes steps S101-S103: S101: Based on the current wheel status data, determine the structural perception features of each wheel.

[0023] The wheel status data includes wheel speed, motor output torque, motor current, motor voltage, wheel slip ratio, tire ground load, and motor temperature.

[0024] Specifically, wheel status data can be obtained directly by measuring or indirectly by sensors equipped on the vehicle, such as wheel speed sensors, torque sensors, current / voltage sensors, temperature sensors, inertial measurement units (IMUs), attitude sensors, etc.

[0025] These sensors are sensors in a high-frequency data acquisition system, acquiring data at a relatively high frequency. For example, the data sampling frequency of each sensor is 100Hz.

[0026] Specifically, the high-frequency data acquisition system also includes a storage module, through which all data acquired by the sensors are synchronously recorded. Similarly, other data estimated based on the data acquired by the sensors can also be synchronously recorded in the storage module of the high-frequency data acquisition system.

[0027] Specifically, after data is collected by sensors, a data preprocessing process is executed. This process can be standardized for data collected from different sensors. For example, the preprocessing process may include outlier removal, synchronization alignment, interpolation, and normalization. This effectively enhances the accuracy and reliability of the obtained wheel state data. Based on this, intelligent wheel torque distribution can be effectively achieved, ensuring the reliability of wheel torque allocation.

[0028] In wheel condition data, tire contact load, also known as tire contact normal load, refers to the force perpendicular to the contact surface experienced by the tire in the contact area with the ground. Specifically, the tire contact normal load can be estimated by calculating the suspension deformation caused by load transfer using the vehicle's lateral / longitudinal acceleration measured by an IMU and combined with a dynamic model, thereby estimating the suspension travel and further estimating the tire contact load. The suspension travel can also be directly measured using other dedicated sensors.

[0029] Wheel slip ratio quantifies the degree of wheel slippage relative to the ground during driving. In braking conditions (wheel deceleration), the wheel slip ratio is the ratio of the difference between the vehicle speed and the wheel's linear velocity to the vehicle speed. In driving conditions (wheel acceleration), the wheel slip ratio is the ratio of the difference between the wheel's linear velocity and the vehicle speed to the vehicle speed. In other words, the wheel slip ratio is calculated from the wheel speed and the vehicle speed / total vehicle speed. The wheel slip ratio can also be called the wheel slip ratio. The wheel linear velocity is the product of the wheel angular velocity (i.e., the wheel speed mentioned above) collected by the wheel speed sensor and the wheel radius.

[0030] In addition, motor temperature can be directly detected by a temperature sensor, or it can be calculated using a thermal model based on motor torque and speed. A thermal model describes the dynamic process of heat generation, storage, transfer, and dissipation in a system, and based on this model, accurate predictions of motor temperature changes can be made.

[0031] The aforementioned current wheel status data refers to the wheel status data of each wheel acquired at the current moment. The number of wheels may vary depending on the actual situation. For example, a typical sedan has four wheels: a left front wheel, a right front wheel, a left rear wheel, and a right rear wheel.

[0032] In addition, the structural perception characteristics of each wheel represent the control importance / degree of importance of each wheel under the current working conditions, which can characterize the torque control priority of each wheel at the current moment.

[0033] Specifically, structural perception features can be represented by scores, which are then called structural perception scores. The higher the score of a structural perception feature, the higher the torque control priority of the corresponding wheel. Conversely, the lower the score of a structural perception feature, the lower the torque control priority of the wheel corresponding to that feature.

[0034] For example, the score of the structure-aware feature ranges from [0,1].

[0035] More specifically, the importance of each item in the current wheel state data to the control of the wheel under the current working condition is evaluated / scored, and the final structural perception characteristics of the wheel are determined based on the scores corresponding to each item, such as the mean, sum or weighted sum of each item's score.

[0036] Alternatively, more specifically, structural sensing features include load concentration and / or torque transmission coupling. In this case, based on the current wheel state data, i.e., the state data of each wheel at the current moment, the load concentration and / or torque transmission coupling of each wheel are determined, and based on the score of the load concentration and / or torque transmission coupling, the structural sensing features of each wheel are determined.

[0037] Load concentration refers to the ratio or relative magnitude of tire contact load to the total vehicle load. It reflects the degree of concentration of load distribution among wheels. The higher the load concentration, the higher the tire contact load of that wheel relative to other wheels. In this case, the corresponding score is higher, and the torque control priority of that wheel is higher.

[0038] Torque transmission coupling describes the tightness of torque transmission between the wheels of a drivetrain. It refers to the strength and degree of correlation between the torque output of each wheel and the influence of the vehicle structure on other wheels or the overall vehicle state. A higher torque transmission coupling indicates better power transmission performance for that wheel, resulting in a higher score and a higher priority for torque control of that wheel.

[0039] Alternatively, more specifically, there is a correspondence between the load concentration and / or torque transmission coupling of the wheel and the score of the structural perception feature. The score corresponding to the load concentration and / or torque transmission coupling of the wheel is directly determined as the value of the structural perception feature corresponding to the load concentration and / or torque transmission coupling.

[0040] Alternatively, specifically, the wheel status data may also include the control command mode. The control command mode refers to the torque control method used by the wheel motor. Generally, the control command mode is drive control, slip control, or thermal protection control. Drive control refers to controlling the motor to output corresponding torque based on actual needs, such as driver instructions; slip control refers to dynamically adjusting the motor torque based on wheel slip ratio; and thermal protection control refers to dynamically limiting torque output based on monitored motor temperature.

[0041] At this point, based on the control command mode, at least one of the other wheel state data is analyzed to determine the structural perception characteristics of each wheel.

[0042] For example, in slip control mode, the corresponding score is determined based on the slip ratio / slip ratio of each wheel, thereby obtaining the structural perception characteristics of each wheel.

[0043] For example, if the difference between the slip rate of the vehicle's front right wheel and the preset slip rate in the current wheel state data exceeds the preset difference, and the difference between the slip rate of the vehicle's front right wheel and the preset slip rate is greater than that of other wheels, then the torque control priority of the vehicle's front right wheel is determined to be higher than that of other wheels, that is, the value of the structural perception feature of the front right wheel is greater than that of other wheels.

[0044] Understandably, the wheel control command pattern directly affects which wheel's needs should be prioritized when total torque / traction is limited. Therefore, combining the wheel control command patterns allows for a more accurate determination of the torque control priority for each wheel.

[0045] For specific implementation details, please refer to existing technologies, which will not be elaborated here.

[0046] Alternatively, a feature extraction model can be used to directly analyze and extract features from the current wheel state data of each wheel to obtain the structural perception features of each wheel.

[0047] Understandably, the high priority of torque control for a certain wheel only indicates that the torque of that wheel needs to be adjusted first relative to the torque of other wheels, and does not limit whether the torque of that wheel is increased or decreased.

[0048] S102: Based on the wheel state data within the first preset time period, predict the state of each wheel within the second preset time period to obtain the state prediction information of each wheel.

[0049] The first preset time period is the time period lasting for a first preset duration up to the current moment, and the second preset duration is the time period lasting for a second preset duration after the current moment.

[0050] Generally, the first preset duration is longer than the second preset duration. In this way, since data in a short time window is easily interfered with by instantaneous noise, while data in a long time window contains more historical information, predicting the wheel state for a shorter second preset duration based on wheel state data with a longer first preset duration can better avoid interference and ensure prediction accuracy.

[0051] For example, the first preset duration can be, for instance, 5 seconds. In this case, taking a time interval of 100ms for acquiring wheel status data as an example, the first preset time period typically contains 50 sets / frames of wheel status data.

[0052] Wheel state data within the first preset time period, i.e., the sequence of wheel state changes within the first preset time period.

[0053] Specifically, the first preset time period is the duration of multiple torque control cycles, and the second preset time period is the duration of one torque control cycle. Here, the torque control cycle refers to the time interval for adjusting the wheel torque. For example, the duration of the torque control cycle is 100ms.

[0054] The process of determining the state prediction information for each wheel can be carried out in parallel.

[0055] Specifically, the state prediction information for each wheel is the wheel state data for each wheel within a second preset time period.

[0056] More specifically, based on the vehicle state data of the wheel during a first preset time period, the state change trend of the wheel is determined, and based on the state change trend of the wheel, the vehicle state data of the wheel during a second preset time period is predicted.

[0057] Alternatively, specifically, the state prediction information may also include a state category, that is, the category of the wheel's state during a second preset time period. The state category includes normal state and abnormal state.

[0058] At this point, for each wheel, based on the wheel state data of the wheel in the first preset time period, the wheel state data of the wheel in the second preset time period is predicted, and based on the predicted wheel state data of the wheel in the second preset time period, it is determined whether the wheel is in an abnormal state, and the state category of the wheel is obtained, that is, the state category of the wheel in the second preset time period.

[0059] More specifically, based on the differences between various wheel state data and their corresponding target values ​​within a second preset time period, the anomaly probability for each item is determined; the larger the difference (in absolute value), the greater the anomaly probability for that item. Then, based on whether the anomaly probability for each item exceeds the corresponding preset anomaly probability, it is determined whether the wheel state within the second preset time period is abnormal.

[0060] If the probability of any anomaly does not exceed the corresponding preset probability, the wheel is determined to be in a normal state during the second preset time period. If the probability of at least one anomaly exceeds the corresponding preset probability, the wheel is determined to be in an abnormal state during the second preset time period.

[0061] More specifically, abnormal states also include instability states, thermal protection states, etc. Instability states specifically include slippage, suspension, torque overload, etc.

[0062] At this point, after determining that the wheel is in an abnormal state during the second preset time period, the state category of the abnormal state of the wheel can be further determined based on the predicted value of the item whose abnormal probability exceeds the corresponding preset abnormal probability.

[0063] For example, when determining a specific abnormal state, if the motor temperature of the left front wheel of the vehicle is higher than the preset thermal protection temperature threshold in the wheel state data during the second preset time period, then the abnormal state of the wheel during the second preset time period is determined to be a thermal protection state.

[0064] Alternatively, specifically, the state prediction information also includes a power response delay estimate, i.e., the time from the moment the torque command is issued to the moment the wheel motor actually outputs and transmits the torque to the tire contact surface to form effective traction or braking force. Correspondingly, the wheel state data within the first preset time period also includes the moment the command for controlling the wheel motor torque is issued, and the correspondence between the tire contact load and time. In this way, when subsequently distributing torque to each vehicle, the wheel's power response delay can be taken into account, allowing for advance adjustment of the wheel torque when a wheel is predicted to be suspended or stuck, thereby compensating for the impact of the power response delay and improving wheel traction efficiency.

[0065] S103: Based on the structural perception features and the state prediction information, torque is allocated to each wheel to determine the target torque for each wheel.

[0066] Specifically, based on structural perception features, the control priority of each wheel is determined. Combined with state prediction information, torque is allocated to wheels with high control priority, and then the torque of other lower priority wheels is allocated sequentially to obtain the target torque for each wheel. It is understandable that the sequence described here is only at the logical decision-making level; the actual execution is completed simultaneously. That is, given limited torque resources, priority is given to ensuring that the torque allocation meets the torque allocation needs of the higher priority wheels.

[0067] For example, if the total torque provided by the vehicle is N, and the torque is sufficient, the torques allocated to wheel a and wheel b are n1 and n2, respectively, where N = n1 + n2. If the torque is insufficient, for example, n2 < n1 < N < n1 + n2, the torque control priority of wheel a is higher than that of wheel b. Therefore, the torque allocated to wheel a is determined to be n1, and the torque allocated to wheel b is N - n1. It is understood that this example is only used to illustrate how torque is preferentially allocated based on torque control priority and does not represent the actual torque relationship.

[0068] More specifically, based on the state prediction information of each wheel, including wheel state data, state category, power response delay, etc., the torque of each wheel is allocated to obtain the target torque of each wheel, that is, the torque of each wheel in the next control cycle.

[0069] Furthermore, based on different driving experience needs, when distributing torque to each wheel, the torque allocated to each wheel is adaptively adjusted according to the response strength corresponding to the driving experience needs, so as to obtain the target torque for each wheel.

[0070] For example, if the torque allocated to wheel a is 100 Nm, but its current torque is 20 Nm, in order to ensure a smoother driving experience and avoid passengers feeling sudden acceleration, the torque allocated to wheel a can be adaptively reduced to, for example, 40 Nm, and the torque of wheel a can be adjusted in multiple control cycles.

[0071] The response strength refers to the rate of change of torque; a larger response strength results in a larger rate of change of torque, and a smaller response strength results in a smaller rate of change of torque. Specifically, the rate of change of torque here refers to the amount of torque change in each control cycle.

[0072] In addition to the driving experience requirements mentioned above, the responsiveness also needs to take into account the vehicle's environment, the properties of the wheel motors, and other information.

[0073] After determining the target torque for each wheel, the target torque for each wheel is sent to the vehicle controller, and then sent to the wheel-end motor controller. The wheel-end motor controller then controls the wheel motor torque to the target torque.

[0074] Alternatively, for each wheel, a control command is generated based on the target torque of the wheel, and this command is sent to the wheel-end motor controller via the vehicle controller. The wheel-end motor controller then executes the control command, adjusting the wheel motor torque to the corresponding target torque, thus achieving closed-loop control of the wheel torque. With a short control cycle, high-frequency closed-loop control of the wheel torque can be achieved.

[0075] Specifically, the target torque or control command is sent to the vehicle controller in the form of a standard bus message.

[0076] Depending on the specific needs, for each wheel, along with the target torque or control command, the information sent to the vehicle controller also includes the wheel's control priority tag. This way, in the event of certain faults, a redundancy strategy can be implemented, prioritizing torque adjustment for the wheel with the highest control priority based on the received control priority tag, thus preventing the vehicle from losing control and ensuring timely vehicle recovery.

[0077] In addition, the information sent to the vehicle controller also includes the model fusion confidence score. The vehicle controller can use this model fusion confidence score to perform online debugging to locate problems and monitor model performance, or to select strategies. For example, if the confidence score is low, the traditional method of rule-based torque distribution can be used.

[0078] In this embodiment, since wheel status data can accurately reflect whether each wheel is stuck, such as locked or suspended, the torque control priority of each wheel can be accurately determined, and the key drive wheel can be effectively identified. Based on this, the wheel torque is allocated by combining the predicted state of the vehicle in the second preset time period after the current moment, i.e., the state prediction information. Compared with torque allocation based on fixed rules or experience data, the timeliness and rationality of wheel torque allocation can be guaranteed, realizing intelligent wheel torque allocation and reducing vehicle extrication time and energy consumption.

[0079] During vehicle operation, the above steps can be executed cyclically based on the control cycle. In each control cycle, torque is redistributed based on new current wheel status data, thereby ensuring the timeliness and rationality of wheel torque distribution in each control cycle, and ensuring the control accuracy of wheel torque distribution in each control cycle. This adapts to various working conditions and continuously and intelligently distributes wheel torque.

[0080] In order to accurately determine the structural perception features of each wheel, and then rationally distribute the wheel torque to reduce the vehicle's extrication time and energy consumption, and improve the torque control accuracy, in some embodiments, when determining the structural perception features of each wheel based on the current wheel state data, an inter-wheel graph structure is constructed based on the current wheel state data, and the inter-wheel graph structure is input into the target graph neural network (GNN) model for feature extraction to obtain the structural perception features of each wheel.

[0081] Among them, the wheel diagram structure represents the state of each wheel and the coupling strength between wheels.

[0082] Wheel coupling strength refers to the degree to which a change in the state of one wheel affects the other wheels.

[0083] Specifically, an inter-wheel structure diagram is constructed based on the current wheel state data and the basic diagram structure.

[0084] The basic graph structure can be constructed in real time or obtained directly from the storage area.

[0085] Specifically, using wheels as the nodes of the basic graph structure, edge connections are established between nodes based on the physical connections between the wheels, thus constructing the basic graph structure for all wheels. It is understandable that the resulting basic graph structure will differ depending on the number of wheels and the physical connections between them.

[0086] More specifically, each wheel drive unit, namely the wheel motor, tire and reduction structure, is selected as the basic node unit / node of the basic diagram structure or the wheel-to-wheel diagram structure.

[0087] Alternatively, since the number of wheels and the physical connections between wheels are usually related to the vehicle model, the base graph structure is also related to the vehicle model. The vehicle model of the vehicle where the wheel is located can be obtained, and the relevant base graph structure can be obtained from the storage area based on the vehicle model.

[0088] The physical connection between wheels refers to a broad physical connection, that is, a connection through the dynamics of the chassis, rather than a narrow physical connection, such as a rigid mechanical connection.

[0089] For example, in a typical four-wheeled car, the physical connections between wheels include the connections between any two wheels. In this case, the basic graph structure constructed for a typical four-wheeled car includes front axle wheel connections (FL↔FR), rear axle wheel connections (RL↔RR), left wheel connections (FL↔RL), right wheel connections (FR↔RR), and diagonal connections (FL↔RR, FR↔RL). All wheels are coupled, and all edge connections are bidirectional. By configuring these coupling relationships along the edges, a dense network structure is formed, i.e., the wheel-to-wheel structure diagram. Furthermore, FL refers to the left front wheel, FR to the right front wheel, RL to the left rear wheel, and RR to the right rear wheel.

[0090] Then, based on the current wheel state data, the state of each wheel and the inter-wheel coupling strength are determined, and an inter-wheel structure diagram is constructed based on the state of each wheel and the inter-wheel coupling strength.

[0091] More specifically, based on the current wheel state data of each wheel, the comprehensive state value of each wheel and the inter-wheel coupling strength are calculated, and the comprehensive state value of each wheel is configured as the feature of the corresponding node, and the inter-wheel coupling strength is configured as the feature of the corresponding edge, thereby obtaining the inter-wheel structure diagram.

[0092] The overall state value of a wheel is obtained by weighted summation of all items in the current wheel state data. The weight of each item in the wheel state data is determined based on its importance to the wheel torque distribution; the higher the importance, the greater the weight.

[0093] Alternatively, the overall state value of a wheel can be obtained by performing fuzzy logic processing on the values ​​of various data points in the current wheel state data.

[0094] The inter-wheel coupling strength can be calculated based on the vehicle dynamics model and the current wheel state data of each wheel.

[0095] Alternatively, the correlation coefficient between time series of any item in the wheel state data of each wheel (such as wheel speed or slip ratio), such as the Pearson correlation coefficient, can be calculated as the inter-wheel coupling degree between the wheels.

[0096] For example, graph computing libraries such as the Deep Graph Library (DGL) or the PyTorch Geometric Library (PyG) can be used to construct inter-wheel graph structures.

[0097] In addition, the wheel graph structure is bound to the current time, that is, the timestamp bound to the wheel graph structure built based on the current wheel state data is the current time.

[0098] Specifically, the wheel-to-wheel diagram structure is updated once per control cycle as it changes over time. Correspondingly, the target graph neural network model acquires a new wheel-to-wheel diagram structure once per control cycle. The new wheel-to-wheel diagram structure is generated in the same way as the original one. This allows for better dynamic tracking of the wheel-to-wheel coupling strength and the structural perception characteristics of each wheel, thereby achieving continuous, timely, and reasonable wheel torque distribution.

[0099] Alternatively, the wheel-to-wheel diagram structure can be updated over time at a period slightly longer than the control cycle. For example, this period, slightly longer than the control cycle, is 1 second. This longer update period for the resource-intensive wheel-to-wheel diagram structure optimizes efficiency and resource allocation, allowing the entire system to operate stably with limited hardware resources, meeting the real-time requirements of vehicle control. Furthermore, it avoids overly frequent changes in upper-level control strategies, ensuring overall stability.

[0100] When the features of all nodes are used as input data for the target graph neural network model, they are stored in tensor form to facilitate subsequent input into the target graph neural network model.

[0101] After data collection from actual vehicles is completed, the sequence of wheel-to-wheel graph structures corresponding to the data collected throughout the entire sampling period is saved on the local training platform. This facilitates the training of the target graph neural network model to obtain a more accurate model. The entire sampling period refers to the entire driving process of the actual vehicle, or the entire time period during which data is collected from the actual vehicle. Furthermore, the local training platform can be a private platform, such as a company's proprietary platform. Depending on the specific needs, the training platform can also be public.

[0102] After obtaining the inter-wheel graph structure, the data format of the nodes and edges of the inter-wheel graph structure is rectified so that the data format of the nodes and edges of the inter-wheel graph structure is in the standard graph input format. Then, the inter-wheel structure graph in the standard graph input format is input into the target graph neural network model for feature extraction.

[0103] For example, the data format of nodes and edges in the wheel graph structure can be standard formats such as JSON, HDF5, and Pickle.

[0104] For example, the data format of nodes and edges in the inter-wheel graph structure, i.e., the standard graph input format, includes a node feature matrix, an edge index list, and an edge feature matrix.

[0105] The node feature matrix has a tensor shape of shape=[4,N], allowing for precise definition of its shape. Specifically, it's a 4xN two-dimensional matrix. The 4 rows represent the 4 nodes in the wheel diagram structure, corresponding to the 4 wheels of the vehicle. The N columns represent N features for each node, describing its state. The specific value of N can be adjusted according to actual needs.

[0106] The edge index list includes the index number of each edge in the wheel graph structure, such as [0,1], [1,0], [2,3], etc. Here, 0, 1, 2, and 3 represent the four wheels of the vehicle.

[0107] The tensor shape of the edge feature matrix is ​​shape=[E,M], which allows for a precise definition of the shape of the edge feature matrix. That is, the edge feature matrix is ​​a two-dimensional matrix with E rows and M columns. Here, E represents the number of edges in the wheel graph structure, corresponding to the number of edges between the wheels, and M columns represent the number of dimensions of each edge feature.

[0108] This facilitates subsequent offline training and model fine-tuning based on the wheel graph structure stored in the training platform, and also ensures that the target graph neural network model accurately extracts features from the wheel graph structure, avoiding feature extraction errors caused by format errors.

[0109] For example, DGL or PyG can be used to batch process and package the sequence of wheel graph structures, and then the packaged data can be stored in the training platform.

[0110] Additionally, the model outputs a vector at each inference iteration, containing a structure perception score for each wheel. This score assesses whether the wheel has the highest torque control priority in the current control cycle. The vector has four dimensions: four wheels, with one corresponding structure perception score for each wheel.

[0111] In this way, since the wheel-to-wheel diagram structure represents the state of each wheel and the inter-wheel coupling strength, and the inter-wheel coupling strength represents the degree of influence between wheels, based on the wheel-to-wheel diagram structure constructed based on the current wheel state data, the target graph neural network model is used to extract features from the wheel-to-wheel diagram structure, which can obtain structural perception features that can accurately represent the wheel state. This allows for accurate determination of the torque control priority of each wheel, which helps to reduce the time required to get out of trouble when the wheel is abnormal and reduce energy consumption.

[0112] It should be noted that the target graph neural network model can be flexibly adjusted according to the electric drive layout of different vehicle models, such as electric axles, wheel-side motors, dual-motor axles, etc., to improve the model's adaptability to different vehicle models.

[0113] To effectively ensure the accuracy of structure-aware features, in some embodiments, when constructing the inter-wheel graph structure based on the current wheel state data, the wheels are used as nodes in the basic graph structure, and edge connections between nodes are established according to the physical connections between the wheels to construct the basic graph structure. Then, multi-dimensional state features are configured for the nodes of the basic graph structure based on the current wheel state data, the inter-wheel state differences are determined based on the current wheel state data, and multi-dimensional weight features are configured for the edges of the basic graph structure based on the inter-wheel state differences.

[0114] Among them, the multidimensional state features and the multidimensional weight features correspond one-to-one.

[0115] For details on how to construct the basic graph structure, please refer to the above content, which will not be repeated here.

[0116] Specifically, for each wheel, based on the current wheel state data, multi-dimensional state features are configured for the node with that wheel in the basic graph structure.

[0117] Based on the current wheel state data of each wheel, the state difference between any two wheels is determined, i.e., the wheel-to-wheel state difference, and the wheel-to-wheel state difference is used as the weight feature between the two corresponding wheels.

[0118] For example, the current wheel status data includes: the current output torque of the motor, current, voltage, wheel speed, slip ratio, tire ground contact normal load, motor temperature or thermal model calculated value, current control command mode, etc.

[0119] For example, the multi-dimensional state features configured for the nodes corresponding to the wheels are packaged into a fixed-dimensional vector. The fixed dimensions of this vector include torque, current, voltage, wheel speed, slip ratio, load, and temperature rise. That is, the number of fixed dimensions in this vector is 7, forming the input vector. Here, the input vector can be represented by X ∈ Rⁿ, where n is the number of dimensions, X represents an input vector with n feature dimensions, and R is a real number, that is, the value of each of the n dimensions of the input vector X is a real number.

[0120] Specifically, the number of edge feature dimensions of different edge connections in the wheel graph structure is consistent, so that the features of different edge connections in the wheel graph structure can be uniformly input into the target graph neural network model for joint calculation to obtain accurate structural perception features of the wheel.

[0121] For example, the edge feature dimension includes the difference in wheel slip rate, the difference in motor torque response, the difference in load, and the difference in terrain adhesion. The difference in motor torque response is the power response delay between wheels. Additionally, the difference in terrain adhesion can be input or estimated by the terrain perception system.

[0122] Similarly, the number of dimensions of different nodes in the inter-wheel structure diagram is the same.

[0123] Specifically, the target graph neural network model can be flexibly adjusted based on the electric drive layout of different vehicle models, such as electric axles, wheel-side motors, dual-motor axles, etc., to improve the model's adaptability.

[0124] In this embodiment, based on state data and the state differences between wheels, node features and wheel edge features can be accurately determined, and the number of dimensions of wheel features and node features can be kept consistent, which facilitates model processing and effectively ensures the accuracy of structure-aware features.

[0125] In training the target graph neural network model, each data point in the training dataset includes the model input sample and its corresponding label, and is divided into training data, validation data, and test data according to a preset ratio. After training the model using the training data, the model is validated using the validation data and tested using the test data to ensure the accuracy of the final target prediction model.

[0126] For example, the proportions of training data, validation data, and test data are 85%, 15%, and 15%, respectively.

[0127] The model input samples are wheel structure diagrams determined based on historical wheel state data, and the corresponding sample labels are the structure-aware features of the wheel structure diagrams determined based on historical wheel state data.

[0128] Specifically, based on the trigger records, traction control system / anti-lock braking system trigger records, traction markers, wheel spin times, etc., corresponding to the historical wheel state data, the sample labels corresponding to the wheel structure diagram determined based on the historical wheel state data are determined.

[0129] Alternatively, the wheel structure diagram corresponding to the historical wheel state data can be stored in a preset area. When model training is required, the wheel structure diagram corresponding to the required historical wheel state data can be directly read from the preset area.

[0130] After obtaining the wheel-to-wheel structure diagram corresponding to the historical wheel state data, the time series of the historical wheel state data can be used to assess whether the wheel-to-wheel structure diagram accurately reflects the trend of wheel-to-wheel load changes, correctly captures the state boundary between the slippery wheel and the normal wheel, and / or adapts to the adjustment process of motor output changes. If the verification is successful, meaning that the wheel-to-wheel structure diagram accurately reflects the trend of wheel-to-wheel load changes, correctly captures the state boundary between the slippery wheel and the normal wheel, and adapts to the adjustment process of motor output changes, then it can be used as a sample in the training data of the target graph neural network model.

[0131] During training, noise perturbation can be added to the weight features of the edges in the inter-wheel structure graph to improve the robustness and generalization ability of the trained target graph neural network model.

[0132] For example, depending on the different ways of expressing the output content of the target graph neural network model, i.e., the sample labels, different training methods can be adopted, such as using the cross-entropy function or the weighted mean square error function as the loss function in the training process.

[0133] For example, minimizing the difference between the structure weights / structure-aware features output by the model and the labels of the actual structure-aware features is used as the training objective of the target graph neural network model.

[0134] For example, the Adam optimizer is used for optimization, with an initial learning rate of 0.001, and cosine annealing is used for dynamic parameter tuning during training.

[0135] For example, the number of training rounds is set to 200 rounds. The number of training rounds can be adjusted according to the convergence in the validation set. For example, the training can be stopped early, i.e., the training process can be stopped before the number of training rounds reaches 200 rounds.

[0136] For example, parallel training can be performed based on the inter-round graph structure in the training data.

[0137] For example, the aforementioned historical wheel status data can be obtained by analyzing data from real-vehicle tests conducted on a pre-set electric off-road platform, or by analyzing data from the actual operation of the vehicle.

[0138] This can significantly improve the training effect of the target graph neural network model.

[0139] In order to achieve intelligent wheel torque distribution, reduce vehicle extrication time and energy consumption, and improve the control accuracy of torque distribution, in some embodiments, the above wheel torque distribution method further includes: inputting wheel state data within a first preset time period into a target prediction model, so that the target prediction model can predict the state of each wheel within a second preset time period, and obtain state prediction information of each wheel.

[0140] For details regarding the first preset time period, the second preset time period, and the state prediction information, please refer to the above content, which will not be repeated here.

[0141] The first preset time period can be determined based on a sliding window method.

[0142] For example, the target prediction model employs a standard two-layer Long Short-Term Memory (LSTM) network structure, with 128 units per layer. The input is a multidimensional time-series tensor for each wheel, with the input parameters having dimensions of the number of time steps and the number of features. The model output is the state label or state vector of the current wheel in the next control cycle.

[0143] The output of the target prediction model can be set as classification (such as normal state, slip state, escape state, etc.) or regression (slip rate prediction value, torque abnormal probability, etc.) depending on the training method.

[0144] For example, in classification tasks, the cross-entropy loss function can be used for model optimization, while in regression tasks, the mean squared error (MSE) function can be used for model optimization.

[0145] For example, the optimization tool can be the Adam optimizer, with an initial learning rate of 0.001, combined with an Early Stopping mechanism to control the number of training epochs and avoid overfitting. Key parameters such as the number of LSTM layers, hidden units, and time steps are optimized using a grid search method to improve prediction accuracy and generalization ability, thereby obtaining the desired target prediction model.

[0146] In addition, the target prediction model supports four-wheel parallel processing, has a lightweight structure, and is easy to deploy on vehicle control platforms.

[0147] For example, the target prediction model trained based on the LSTM model is converted into the Open Neural Network Exchange (ONNX) format for actual deployment.

[0148] Specifically, the update frequency of the model's prediction results, i.e., the wheel state prediction information, is consistent with the control cycle, which can better ensure the real-time performance and synchronization of the data.

[0149] In this embodiment, since the wheel state data within the first preset time period can accurately characterize the changes in wheel state, the state of each wheel within the second preset time period is predicted by the target prediction model to obtain the state prediction information of each wheel. Based on the wheel state prediction information, combined with structural perception features, the wheel torque control in the vehicle can be transformed from passive to active, improving the safety and rationality of wheel torque distribution, thereby realizing intelligent wheel torque distribution, reducing vehicle extrication time and energy consumption, and improving the control accuracy of torque distribution.

[0150] Furthermore, during the training process of the target prediction model, each data point in the training dataset includes the model input sample and its corresponding label, and is divided into training data, validation data, and test data according to a preset ratio. After training the model using the training data, the model is validated using the validation data and tested using the test data to ensure the accuracy of the final trained target prediction model.

[0151] For example, the ratio of training data, validation data, and test data can be 8:1:1.

[0152] It is understandable that the samples in the dataset used to train the target prediction model can be time series extracted from the aforementioned historical wheel state data.

[0153] When validating the model, accuracy, recall, and latency response capabilities are evaluated on a validation set, with a focus on its ability to identify slip trends in advance. For example, before a slip rate increases, the model is assessed to see if it can successfully predict slip risk within 1-2 attempts and prompt the controller to intervene. Simultaneously, the prediction error distribution is analyzed. If the error distribution meets preset requirements, the model is considered validated to ensure the stability and robustness of the target prediction model's output. If classification output is used, the prediction accuracy must be at least 90%, especially under low-adhesion conditions, before the model is considered validated.

[0154] To reduce the time required to get out of trouble after wheel malfunctions and to reduce the energy consumption required for getting out of trouble, in some embodiments, torque is allocated to each wheel based on structural perception features and state prediction information. When determining the target torque of each wheel, the structural perception features and state prediction information are fused to obtain fused features. The fused features are input into the target control strategy model for analysis to determine the torque allocation coefficient of each wheel. Through the target control strategy model, the target torque of each wheel is determined based on the torque allocation coefficient and the vehicle power request.

[0155] For each wheel, the structural perception features and state prediction information of that wheel are spliced ​​together, and the spliced ​​features are used as fused features.

[0156] There is no specific limitation on the order of structural sensing features and state prediction information; it can be either structural sensing features first and then state prediction information, or state prediction information first and then structural sensing features.

[0157] Specifically, the weights corresponding to the structural perception features and the state prediction information can be set based on their respective importance, i.e., the degree of influence on the vehicle torque distribution.

[0158] In structural perception features and state prediction information, the greater the influence on vehicle torque distribution, the higher the importance and the greater the corresponding weight.

[0159] Specifically, the features obtained by weighted concatenation of structurally perceived features and state prediction information are used as fused features.

[0160] Since structural perception features are used to characterize the torque control priority of wheels, they are mainly used to prioritize the torque adjustment of high-priority wheels when torque resources are limited. When torque resources are sufficient, the level of torque control priority characterized by structural perception features has little impact on wheel torque distribution. When torque resources are insufficient, the vehicle's speed of getting out of trouble may be slow when torque distribution is based solely on state prediction information.

[0161] More specifically, based on the vehicle's power request and the total torque that the vehicle can provide, it can be determined whether the torque resources are sufficient. If the torque resources are sufficient, the weight of the state prediction information is determined to be greater than the weight of the structural perception features. If the torque resources are insufficient, the weight of the structural perception features is determined to be greater than the weight of the state prediction information.

[0162] Specifically, the target control strategy model can also be described as a reinforcement learning controller based on Deep Deterministic Policy Gradient (DDPG). This target controller strategy is used to generate the torque distribution ratio of each wheel in real time, or the torque distribution coefficient of each wheel.

[0163] The torque distribution ratio ranges from [0,1].

[0164] The input vector of the target control strategy model consists of two parts: first, the structure-aware features output by the target GNN model (representing the control priority of each round); and second, the state prediction vector output by the target prediction model trained by the LSTM model (i.e., the above-mentioned state prediction information, which may also include the slip trend and dynamic risk assessment of each round).

[0165] The output vector of the target control strategy model is a four-dimensional continuous control vector, corresponding to the torque distribution coefficient of the four wheels, ranging from 0 to 1. The controller calculates the final torque command based on the actual power request and the coefficient ratio.

[0166] In this embodiment, since priority reflects the key drive wheel / wheel and structural perception features can characterize priority, the structural perception features of the wheel and the state prediction information are combined to obtain fused features. The fused features are then input into the target control strategy model for analysis. This allows for the timely and reasonable determination of the torque distribution coefficients of each wheel that are suitable for the current working condition. Based on this, the target control strategy model determines the target torque of each wheel based on the torque distribution coefficients and the actual power demand of the vehicle, i.e., the vehicle's power request. This may better determine the wheel torque suitable for the current working condition, enabling the wheel to reach the target torque through the adjuster, thereby quickly achieving wheel extrication from abnormal conditions, reducing the energy consumption required for extrication, and improving control accuracy.

[0167] To ensure the vehicle's speed of extrication from trouble, reduce the time required to extricate itself after wheel malfunctions, and minimize energy consumption for extrication, in some embodiments, structural perception features and state prediction information are input into the target fusion model, adaptive weight calculation is performed, and the structural perception features and state prediction information are weighted and concatenated based on their respective weights to obtain fused features.

[0168] Specifically, the target fusion model uses a lightweight feature fusion network as an intermediate layer module to dynamically weight and synthesize the structural perception features of the wheel and the state prediction information of the wheel within a second preset time period to obtain fused features.

[0169] More specifically, the target fusion model uses a learnable attention weight mechanism to adaptively calculate the weights of structural perception features and state prediction information, and dynamically determines the adaptive weights of structural perception features and state prediction information.

[0170] Among them, when the weight of structural perception features is greater than the weight of state prediction information during the wheel torque distribution process, structural perception features are dominant over state prediction information in the current control cycle.

[0171] The fused feature vector serves as the sole state input to the target control strategy model, entering the policy Actor network to complete torque allocation decisions. The controller outputs torque allocation coefficients for the four wheels based on the fused features, and calculates the actual target torque value for the motor based on the current power request. This design ensures end-to-end connectivity across the entire model control chain, avoiding information loss and response mismatch between multiple models.

[0172] In addition, the target fusion model includes multiple adjustable interfaces to support the initial setting of fusion weights for the outputs of the target graph neural network model and the target prediction model, the feature priority adjustment logic triggered by various working conditions, and the configuration of attention weight learning rate and stabilization factor. In this way, the target fusion model can meet the needs of various vehicle types, terrains, and adjustment strategies by limiting the weights and effects of the fusion features to obtain fusion features that meet the requirements.

[0173] In this embodiment, an adaptive weight is applied to the structural perception features and state prediction information through a target prediction model and then concatenated. The concatenated features are then input into the target control strategy model to determine the torque distribution coefficient of the wheels. Based on this, combined with the vehicle's power request, the torque distribution coefficient of each wheel can be determined more accurately. Based on this torque distribution coefficient, the target torque can be accurately determined, shortening the time for getting out of trouble when the wheels are abnormal.

[0174] To ensure the torque distribution effect of the target control strategy model, in some embodiments, the wheel torque distribution method further includes: building a training platform in a simulation environment, constructing a reward function based on slip control, escape success rate, energy efficiency target and motion stability, and training the target network structure based on the reward function until convergence, thereby obtaining the target control strategy model.

[0175] In a dedicated simulation environment for off-road / unstructured terrain, a virtual control platform for model training is built, which includes typical road condition simulation modules (gravel, mud, cross-axle, etc.), tire mechanics models, electric drive response models, etc.

[0176] By reproducing real-world operating data, the control strategy can be experimentally optimized in a simulation environment.

[0177] The environment provides a state feedback interface, returning information such as wheel slip rate, current, motor temperature rise, and obstacle clearance determination, which is used for training feedback of the reinforcement learning model.

[0178] Additionally, the target network structure can refer to a policy-value Actor-Critic network architecture.

[0179] The control strategy model employs an Actor-Critic network architecture, constructing separate policy (Actor) and value (Critic) networks. The Actor network responds to the fused features, or GNN-LSTM features, which are the inputs of the concatenated fused features, and outputs the torque distribution ratio of each wheel. The Critic network receives the current wheel state data and the Actor's action output (i.e., the torque distribution ratio of each wheel), and estimates the value of the value function based on this action output in the current state.

[0180] Both the Actor and Critic networks are three-layer fully connected networks. For example, the ReLU activation function can be used as the activation function in these three fully connected layers, and the output layer uses Sigmoid normalization. Sigmoid can map any range of output values ​​to the interval (0, 1).

[0181] Specifically, the target control strategy model can be trained end-to-end and adapted to vehicle deployment platforms.

[0182] Specifically, the Critic network evaluates the torque distribution ratio of each wheel output by the Actor based on the reward function, and adjusts the Actor-Critic network architecture based on the evaluation until convergence, or until the evaluation score of the reward function evaluating the torque distribution ratio of each wheel output by the Actor converges.

[0183] Specifically, a reward function is constructed based on slip control objectives, escape success rate, energy efficiency objectives, and motion smoothness. This reward function guides the Actor network to learn reasonable torque distribution behavior, ensuring that the torque distribution coefficients output by the final target control strategy model effectively reduce wheel slip rate, guide the vehicle to successfully start from a standstill or slipping state, provide positive rewards, reduce unnecessary power output and system losses, and suppress frequent and large torque jumps.

[0184] In addition, the reward function can be further adjusted based on temperature rise suppression, which can effectively control the motor temperature rise and avoid overheating of a single wheel.

[0185] Therefore, in this embodiment, a reward function is constructed based on slip control, escape success rate, energy efficiency target and motion smoothness, and the Actor network is trained based on this reward function in combination with the simulation environment, which can effectively optimize the torque distribution effect of the model.

[0186] It is understandable that the Critic network mentioned above is a value function used to estimate the Actor's output / action in the current state, and is used to guide model training. The actual target control policy model does not include the Critic network part.

[0187] In addition, during the model training phase, the target fusion model and the target control strategy model are usually optimized together, that is, they are trained together to ensure that the fusion features obtained by the target fusion model have a positive guiding effect on the final torque distribution ratio, i.e., the control result.

[0188] For example, the wheel torque distribution process can be as follows: Figure 2As shown, vehicle state data is first collected and preprocessed to obtain wheel state data. Then, based on the wheel state data, an inter-wheel graph structure is constructed. The inter-wheel graph structure, i.e., the graph structure data, is input into the target graph neural network model. Based on the target graph neural network model, the structural perception features of the wheels are extracted. The time series of wheel state data is input into the target prediction model. Based on the target prediction model, the state change trend of the wheels is predicted. Finally, through the target fusion model and the target control strategy model, the structural perception features and the state change trend of the wheels (i.e., the predicted state information) are fused (specifically, spliced ​​fusion) and the inter-wheel torque is allocated to obtain the target torque of each wheel. Then, the torque control command of each wheel is generated. Based on the control command, the inter-wheel torque change is controlled by the vehicle control unit (VCU).

[0189] In some embodiments, the target graph neural network model, target prediction model, and target control strategy model described above can be deployed in a distributed manner in the vehicle controller or the central controller.

[0190] At this point, the target graph neural network model, the target prediction model, and the target control strategy model interact with each other through the intermediate fusion module (Fusion Layer) to exchange data flow and control information. The specific information interaction process can be found in the above scheme description and will not be elaborated upon here.

[0191] For example, the target graph neural network model, the target prediction model, the target control strategy model, and the target fusion model can be deployed on automotive-grade computing platforms such as NXP S32G, NVIDIA DRIVE Orin, Huawei MDC, and Horizon Robotics.

[0192] The deployed target model is compressed using inference frameworks such as TensorRT, ONNX, and TFLite to meet the requirements for vehicle power-on startup and real-time control.

[0193] The torque commands output by all models are sent to the four electric drive controllers via the CAN bus or Ethernet interface to control the motor torque of the wheels.

[0194] Exemplary device like Figure 3 As shown in the figure, this application embodiment also provides a wheel torque distribution device, including a determination module 301, a prediction module 302 and a distribution module 303.

[0195] in, The construction module 301 is used to determine the structural perception features of each wheel based on the current wheel state data, wherein the structural perception features characterize the torque control priority of the wheel. Prediction module 302 is used to predict the state of each wheel in a second preset time period based on wheel state data in a first preset time period, and to obtain state prediction information of each wheel. The first preset time period is a time period lasting a first preset duration up to the current moment, and the second preset time period is a time period lasting a second preset duration after the current moment. The allocation module 303 is used to allocate torque to each wheel based on the structural perception features and the state prediction information, and to determine the target torque of each wheel.

[0196] The wheel torque distribution device provided in this embodiment belongs to the same concept as the wheel torque distribution method provided in the above embodiments of this application. It can execute the method provided in any of the above embodiments of this application and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the specific processing content of the wheel torque distribution method provided in the above embodiments of this application, and will not be repeated here.

[0197] The functions implemented by the determination module 301, prediction module 302 and allocation module 303 can be implemented by the same or different processors calling software, and this application embodiment does not limit this.

[0198] Exemplary electronic devices Another embodiment of this application also provides an electronic device, see [link to relevant documentation] Figure 4 As shown, the electronic device includes a memory 400 and a processor 410.

[0199] The memory 400 is connected to the processor 410 and is used to store programs; The processor 410 is configured to implement the wheel torque distribution method disclosed in any of the above embodiments by running the program stored in the memory 400.

[0200] Specifically, the electronic device may also include: a bus, a communication interface 420, an input device 430, and an output device 440.

[0201] The processor 410, memory 400, communication interface 420, input device 430, and output device 440 are interconnected via a bus. Among them: A bus can include a pathway for transmitting information between various components of a computer system.

[0202] The processor 410 can be a general-purpose processor, such as a general-purpose central processing unit (CPU), a microprocessor, etc., or an application-specific integrated circuit (ASIC), or one or more integrated circuits used to control the execution of the program of the present application. It can also be a digital signal processor (DSP), an application-specific integrated circuit ASIC, a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0203] Processor 410 may include a main processor, as well as a baseband chip, modem, etc.

[0204] The memory 400 stores a program for executing the technical solution of this application, and may also store an operating system and other critical business functions. Specifically, the program may include program code, which includes computer operation instructions. More specifically, the memory 400 may include read-only memory (ROM), other types of static storage devices capable of storing static information and instructions, random access memory (RAM), other types of dynamic storage devices capable of storing information and instructions, disk storage, flash memory, etc.

[0205] Input device 430 may include a device for receiving user input data and information, such as a keyboard, mouse, camera, scanner, light pen, voice input device, touch screen, pedometer, or gravity sensor.

[0206] Output device 440 may include devices that allow information to be output to a user, such as a display screen, printer, speaker, etc.

[0207] The communication interface 420 may include a device that uses any transceiver to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area network (WLAN), etc.

[0208] The processor 410 executes the program stored in the memory 400 and calls other devices, which can be used to implement the various steps of any wheel torque distribution method provided in the above embodiments of this application.

[0209] Those skilled in the art will understand that Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0210] This application also proposes a chip, which includes a processor and a data interface. The processor reads and runs a program stored in a memory through the data interface to execute the wheel torque distribution method described in any of the above embodiments. For details of the processing and its beneficial effects, please refer to the embodiments of the wheel torque distribution method described above.

[0211] This application also provides a vehicle equipped with the aforementioned wheel distribution device, or Figure 4 The electronic device shown.

[0212] In addition to the methods and devices described above, embodiments of this application propose a computer program product comprising computer program instructions that, when executed by a processor, cause the processor to perform the steps of the wheel torque distribution methods according to various embodiments of this application as described in the "Exemplary Methods" section of this specification.

[0213] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0214] Furthermore, embodiments of this application also propose a storage medium storing a computer program, which is executed by a processor in the wheel torque distribution method according to various embodiments of this application described in the "Exemplary Methods" section above.

[0215] The basic principles of the present invention have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in the present invention are merely examples and not limitations, and should not be considered as essential features of each embodiment of the present invention. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the present invention to the necessity of employing the aforementioned specific details.

[0216] The block diagrams of devices, apparatuses, devices, and systems involved in this invention are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0217] It should also be noted that in the apparatus, device, and method of the present invention, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of the present invention.

[0218] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the invention. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of the invention. Therefore, the invention is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0219] It should be understood that the qualifying terms "first", "second", "third", "fourth", "fifth" and "sixth" used in the description of the embodiments of the present invention are only used to more clearly illustrate the technical solutions and are not intended to limit the scope of protection of the present invention.

[0220] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of the invention to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A wheel torque distribution method, characterized in that, The method includes: Based on the current wheel status data, the structural perception features of each wheel are determined, and the structural perception features characterize the torque control priority of the wheel. Based on the wheel state data within a first preset time period, the state of each wheel within a second preset time period is predicted to obtain the state prediction information of each wheel. The first preset time period is a time period lasting a first preset duration up to the current moment, and the second preset time period is a time period lasting a second preset duration after the current moment. Based on the structural perception features and the state prediction information, torque is allocated to each wheel to determine the target torque for each wheel.

2. The wheel torque distribution method according to claim 1, characterized in that, The process of determining the structural sensing features of each wheel based on the current wheel state data includes: Based on the current wheel state data, an inter-wheel diagram structure is constructed, which represents the state of each wheel and the inter-wheel coupling strength. The wheel-to-wheel diagram structure is input into the target graph neural network model for feature extraction to obtain the structural perception features of each wheel.

3. The wheel torque distribution method according to claim 2, characterized in that, The construction of the wheel-to-wheel diagram structure based on the current wheel state data includes: The basic graph structure is constructed by using wheels as nodes in the basic graph structure and establishing edge connections between nodes based on the physical connection relationships between the wheels. Based on the current wheel state data, configure multi-dimensional state features for the nodes of the basic graph structure; Based on the current wheel state data, the wheel state differences are determined, and multi-dimensional weight features are configured for the edges of the basic graph structure based on the wheel state differences to obtain the wheel graph structure. The multi-dimensional state features and the multi-dimensional weight features correspond one-to-one.

4. The wheel torque distribution method according to claim 3, characterized in that, The method further includes: The wheel state data within the first preset time period is input into the target prediction model so that the target prediction model can predict the state of each wheel within the second preset time period, thereby obtaining the state prediction information of each wheel.

5. The wheel torque distribution method according to claim 1, characterized in that, The step of allocating torque to each wheel based on the structural perception features and the state prediction information, and determining the target torque for each wheel, includes: The structure-aware features and the state prediction information are fused to obtain fused features; The fused features are input into the target control strategy model for analysis to determine the torque distribution coefficient of each wheel; Based on the torque distribution coefficient and the vehicle's power request, the target torque for each wheel is determined using the target control strategy model.

6. The wheel torque distribution method according to claim 5, characterized in that, The process of fusing the structure-aware features and the state prediction information to obtain fused features includes: The structure-aware features and the state prediction information are input into the target fusion model, adaptive weight calculation is performed, and the structure-aware features and the state prediction information are weighted and concatenated based on their respective weights to obtain the fusion features.

7. The wheel torque distribution method according to claim 6, characterized in that, The method further includes: Build a training platform in a simulation environment; A reward function is constructed based on slip control, escape success rate, energy efficiency targets, and motion smoothness. The target network structure is trained based on the reward function until convergence, thus obtaining the target control policy model.

8. A wheel torque distribution device, characterized in that, The device includes: The determination module is used to determine the structural perception features of each wheel based on the current wheel state data, wherein the structural perception features characterize the torque control priority of the wheel. The prediction module is used to predict the state of each wheel in a second preset time period based on the wheel state data in a first preset time period, and to obtain the state prediction information of each wheel. The first preset time period is a time period lasting a first preset duration up to the current moment, and the second preset time period is a time period lasting a second preset duration after the current moment. The allocation module is used to allocate torque to each wheel based on the structural perception features and the state prediction information, determine the target torque of each wheel, and generate torque control commands for each wheel. The control module is used to control the torque of the corresponding wheel to its target torque based on the torque control command.

9. An electronic device, characterized in that, Including memory and processor; The memory is connected to the processor and is used to store programs; The processor is used to implement the wheel torque distribution method as described in any one of claims 1 to 7 by running a program in the memory.

10. A vehicle, characterized in that, The vehicle is equipped with the wheel torque distribution device as described in claim 8, or the electronic device as described in claim 9.

Images