Vehicle control method and vehicle

Abstract

The application relates to the technical field of vehicle control, in particular to a vehicle control method and a vehicle, wherein the method comprises the following steps: acquiring vehicle state data of a vehicle, performing feature extraction on the vehicle state data to obtain vehicle state features; inputting the vehicle state features into a wheel-end relationship identification model to obtain wheel-end state features; inputting the vehicle state features into a slip trend prediction model to obtain slip trend prediction features of the vehicle; determining abnormal features of the vehicle based on the wheel-end state features and the slip trend prediction features; determining a control mode of the vehicle based on the abnormal features; and controlling the vehicle according to the control mode. The application realizes more accurate and comprehensive early identification and cooperative control of potential instability states of the vehicle by constructing a wheel-end topology graph for analysis and combining slip trend prediction.

Description

Technical Field

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

[0002] In existing technologies, active safety control of vehicles in complex scenarios often relies on driver experience or independent analysis based on single wheel-end data.

[0003] Driver experience can be biased, and mistakes or delayed reactions can easily endanger the vehicle. Relying on data from a single wheel end is also incomplete and inaccurate, potentially creating safety hazards. Summary of the Invention

[0004] In view of this, the purpose of this application is to propose a vehicle control method and a vehicle to solve the technical problem that current analysis based on driving experience or single wheel-end data is prone to causing safety hazards to the vehicle.

[0005] To achieve the above objectives, this application provides a vehicle control method, which includes:

[0006] Obtain vehicle status data, extract features from the vehicle status data, and obtain vehicle status features. The vehicle state features are input into the wheel-end relationship recognition model. By performing wheel-end topology construction and feature fusion on the vehicle state features, the wheel-end state features are obtained. The vehicle state features are input into the slip trend prediction model, and the slip trend prediction features of the vehicle are obtained by performing slip trend prediction on the vehicle state features. Based on the wheel end state characteristics and the slip trend prediction characteristics, the abnormal characteristics of the vehicle are determined; Based on the aforementioned abnormal characteristics, the vehicle control method is determined; Vehicle control is performed according to the control method described above.

[0007] Based on the same inventive concept, this application also provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.

[0008] Based on the same inventive concept, this application also provides a vehicle including the above-mentioned electronic devices.

[0009] As can be seen from the above, the vehicle control method and vehicle provided in this application, by acquiring vehicle state data and extracting features, effectively represent the real-time operating state of the vehicle, providing a data foundation for subsequent accurate analysis. By inputting the vehicle state features into the wheel-end relationship recognition model for wheel-end topology construction and feature fusion, a deep modeling and understanding of the interaction relationships between each wheel is achieved, enabling the identification of coupling anomalies in the wheel-end cooperative state. By inputting the vehicle state features into the slip trend prediction model for slip trend prediction, the judgment of future instability risks is realized. By determining the abnormal characteristics of the vehicle based on the wheel-end state features and slip trend prediction features, a comprehensive and accurate diagnosis of the vehicle's dynamic state is achieved. By determining and executing the corresponding vehicle control mode based on the abnormal characteristics, proactive and precise intervention in potential risks is achieved, thereby effectively improving the vehicle's passability, safety, and overall control performance under complex working conditions. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic flowchart of the vehicle control method according to an embodiment of this application; Figure 2 This is a schematic diagram of the structure of the vehicle control device according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0013] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0014] CAN: Controller Area Network. The text mentions a communication bus protocol used for data transmission between electronic control units (ECUs) within a vehicle.

[0015] GPS: Global Positioning System. The text mentions a satellite navigation system used to provide reference speed and positioning information for vehicles.

[0016] IMU: Inertial Measurement Unit. The text mentions sensors used to measure the angular velocity and acceleration of a vehicle's three axes.

[0017] GNN: Graph Neural Network. The text mentions the core deep learning model architecture used in the wheel-end relationship recognition model.

[0018] LSTM: Long Short-Term Memory. It is one of the types of temporal neural networks mentioned in the paper that may be used in slip trend prediction models.

[0019] GRU: Gated Recurrent Unit. It is one type of temporal neural network that the text mentions as a possible tool for predicting slip trends.

[0020] MLP: Multilayer Perceptron. The text mentions the neural network infrastructure used to build the first anomaly detection network or control policy mapper.

[0021] Softmax: The Softmax function is a normalized exponential function. The paper mentions its use in the output layer of anomaly classification networks to transform the input into a probability distribution.

[0022] ESP: Electronic Stability Program. The text mentions it as one of the vehicle's underlying actuators, a system used to apply braking force to control vehicle stability.

[0023] ABS: Anti-lock Braking System. Although not mentioned directly in the text, it is often discussed in conjunction with ESP in relevant technical backgrounds or functional scenarios.

[0024] TCS: Traction Control System. Although not directly mentioned in the text, it is often associated with ESP and ABS systems in relevant technical backgrounds or functional scenarios.

[0025] BiLSTM: Bidirectional Long Short-Term Memory. It is one of the specific types of temporal coding networks mentioned in the paper for slip trend prediction models.

[0026] ReLU: Rectified Linear Unit. The text mentions the activation function used in the nonlinear transformation of graph convolutional networks.

[0027] Focal Loss: Focal loss is a classification loss function mentioned in the text, used in training models for predicting slippage trends, specifically for imbalanced samples (such as mutation events being in the minority class).

[0028] Sigmoid: The Sigmoid function. The activation function mentioned in the text when mapping trend strength to probability values.

[0029] Adam: Adaptive Moment Estimation. The paper mentions optimization algorithms used for model parameter optimization.

[0030] AdamW: Adam with Weight Decay. A variant of the Adam optimizer mentioned in the paper for training wheel-end relationship recognition models.

[0031] In related technologies, vehicle stability and traction control mainly rely on the driver's experience or threshold triggering mechanisms based on a single wheel-end sensor (such as a wheel speed sensor) to achieve passive slippage response or torque limitation.

[0032] In low-traction road scenarios, vehicle stability and handling safety face severe challenges. Existing technologies typically rely on the driver's subjective perception of vehicle attitude and power response for control, or monitor the slip rate of individual wheels through wheel speed sensors. When the slip rate exceeds a preset static threshold, independent intervention is initiated through anti-lock braking systems or traction control systems. To improve control capabilities, some solutions introduce model-based vehicle state observers to estimate the vehicle's longitudinal velocity or sideslip angle. However, their control decisions are often based on the whole vehicle kinematic model, without explicit modeling and in-depth analysis of the complex dynamic coupling relationships between the four wheels. Therefore, the system's control basis is local and fragmented: traditional traction control systems focus on whether a single wheel slips, and electronic stability programs focus on the yaw motion of the entire vehicle, lacking a global understanding and forward-looking prediction of the coordinated working state, mutual influence relationships, and instability trends of the multi-wheel system.

[0033] In practical applications, traditional control methods based on thresholds or simple models have limitations when dealing with complex and dynamic wheel-ground interactions. Control based on single-wheel slip ratio thresholds exhibits lag, and fixed thresholds cannot adapt to drastically changing adhesion conditions, such as paved roads or soft sand. Control methods based on whole-vehicle models are highly dependent on model accuracy; under extreme conditions, model mismatch can lead to decreased control performance or even failure. More critically, existing methods fail to fully utilize the interrelationships between wheel states (slip ratio, torque). If the system only processes each wheel slippage event in isolation, ignoring the dynamic process that a single wheel slippage can rapidly trigger torque redistribution in other wheels and lead to a chain reaction of instability, it will result in uncoordinated control actions, potentially causing oscillations or excessive intervention, affecting both throughput and driving safety.

[0034] One intuitive improvement strategy is to employ rule-based torque vectoring, such as allocating more torque to axles or wheels with better traction. However, this strategy is typically based on simple differential principles or predefined terrain patterns, failing to quantify the intensity of interactions between wheels and future risks in real time and with precision. Another approach is to attempt to design a multi-wheel coordination controller using classical or modern control theory. However, due to the high nonlinearity of the vehicle system and the uncertainty of wheel-ground interactions, controller design is complex, parameter tuning is difficult, and robustness in unknown and variable environments is challenged.

[0035] Based on this, the applicant discovered that it is necessary to design a collaborative control scheme based on wheel-end relationship depth recognition and slip trend intelligent prediction in the vehicle stability control system. This is to solve the problem of how to comprehensively utilize the real-time state data of each wheel under complex adhesion conditions and off-road scenarios, model the spatial coupling relationship between wheels and the temporal evolution trend of slip process through graph model and sequence model respectively, and integrate the information of the two to predict the risk of vehicle instability in a forward-looking manner, thereby generating active, accurate and collaborative torque control commands to maximize vehicle stability and traction while improving handling and safety boundaries.

[0036] The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0037] The vehicle control method proposed in the embodiments of this application, such as Figure 1 As shown, it includes: S101, acquire vehicle status data, extract features from the vehicle status data, and obtain vehicle status features.

[0038] In practical implementation, vehicle state data is a multi-dimensional set of information characterizing the dynamic state of each wheel and the entire vehicle, including wheel speed, driving or braking torque, slip ratio, wheel-end acceleration, and parameters such as longitudinal / lateral acceleration and yaw rate of the entire vehicle. Vehicle state data can be collected and generated through wheel speed sensors, torque sensors, inertial measurement units, and the vehicle controller network. For example, wheel speed sensors installed on each wheel measure the instantaneous rotational speed of each wheel at a frequency of 100 frames per second; real-time torque output values ​​of each wheel are obtained through the motor controller or braking system controller; longitudinal acceleration, lateral acceleration, and yaw rate of the vehicle are obtained through the onboard IMU; all data is transmitted and synchronized via the CAN bus in a specific message format. The real-time slip ratio of each wheel is calculated based on the wheel speed and the vehicle reference speed (obtainable through GPS or multi-sensor fusion). Furthermore, vehicle state data can also be estimated or supplemented by a vehicle state observer and formed after data validity verification.

[0039] In practice, after acquiring vehicle state data, feature extraction is performed to obtain vehicle state features. Specifically, feature extraction includes normalizing the vehicle state data, converting the raw data collected by each sensor to a unified dimension and numerical range; extracting time-domain statistical features, including calculating the mean, variance, standard deviation, and range of slip ratio, torque, and wheel speed sequences for each wheel; extracting frequency-domain features, analyzing the spectral energy distribution by performing Fast Fourier Transform on wheel speed or vehicle body vibration signals; and extracting vehicle lateral and longitudinal dynamic coupling features, modeling the dynamic relationship between acceleration, yaw rate, and wheel-end torque based on the Newton-Euler equations to generate a coupling feature vector characterizing the overall vehicle motion coordination. Finally, the extracted features from each dimension are concatenated and dimensionality reduced to form the vehicle state features.

[0040] S102, input the vehicle state features into the wheel-end relationship recognition model, and obtain the wheel-end state features by performing wheel-end topology construction and feature fusion on the vehicle state features.

[0041] In practice, the wheel-end relationship recognition model uses a deep learning model based on graph neural networks. The core function of this graph neural network deep learning model is to learn the state features by constructing a topological relationship graph and aggregating neighborhood information using graph convolution operations.

[0042] Because vehicle state features contain a variety of data, it is necessary to construct a topological relationship graph between the wheels in order to facilitate understanding of the relationships among them. Thus, by using a deep learning model based on graph neural networks as the wheel-end relationship recognition model, it is possible to learn and output wheel-end state features that can characterize the interaction and global collaborative state between the wheel ends based on the topological relationship graph between the wheels.

[0043] In the wheel-end relationship recognition model, each wheel is first mapped to a node in a graph based on vehicle state features, with real-time slip ratio and torque of each wheel serving as the initial features of the corresponding node. Then, the connection strength (edge ​​weights) between nodes is calculated based on the differences in wheel states (e.g., slip ratio difference, torque difference), dynamically constructing a topology graph reflecting the current wheel-end coupling relationship. Next, a graph convolutional layer performs multi-level spatial aggregation and information transfer on this topology graph, enabling each node to integrate the state information of its neighboring nodes, thereby capturing the spatial association and dynamic coupling between wheel ends. Finally, the high-order association features obtained through graph convolution processing are fused with the original node features to output comprehensive wheel-end state features.

[0044] S103, input the vehicle state features into the slip trend prediction model, and obtain the vehicle slip trend prediction features by performing slip trend prediction on the vehicle state features.

[0045] In practice, the slip trend prediction model uses a prediction model based on temporal deep learning. In order to predict the slip trend, it is generally necessary to have feature data on the change of vehicle state characteristics over time, and these feature data have temporal relationships.

[0046] The prediction model based on temporal deep learning is selected as the slip trend prediction model. It can encode the temporal sequence corresponding to the input vehicle state features through the built-in temporal feature extraction module (such as an encoder based on Long Short-Term Memory Network (LSTM) or Gated Recurrent Unit (GRU)) to capture the dynamic patterns and potential laws of slip rate changes.

[0047] The core function of the slip trend prediction model is to analyze the historical evolution sequence of wheel slip states and learn patterns to predict future slip trends. It receives vehicle state characteristics as input, with particular attention to time-series data on the slip ratio and related dynamic parameters of each wheel. Based on this, the trend analysis and prediction module calculates forward-looking indicators such as the rate of change of slip ratio, acceleration, and probability of abrupt changes, thereby generating slip trend prediction features that characterize the degree of slip risk for each wheel in the near future.

[0048] S104. Based on wheel end state characteristics and slip trend prediction characteristics, the abnormal characteristics of the vehicle are determined.

[0049] In practice, the system receives wheel-end state features from the wheel-end relationship recognition model and slip trend prediction features from the slip trend prediction model. These two types of features are aligned and concatenated in both spatial and temporal dimensions to form a fused global state description vector. This global state description vector is then input into a lightweight anomaly classification network (e.g., a multilayer perceptron consisting of fully connected layers and a Softmax function). After training, this network can identify patterns from the fused features that characterize dangerous states such as instability, slippage, and abnormal torque distribution. The output of this anomaly classification network is the anomaly feature, which can be a multi-dimensional vector where each dimension represents the confidence or severity score of a specific anomaly state (e.g., unilateral wheel instability, diagonal wheel slippage), providing precise and quantitative state diagnosis for the final control decision.

[0050] S105, based on the abnormal characteristics, determine the vehicle control method and control the vehicle according to the driving control method.

[0051] In practical implementation, the control method is determined through a control strategy mapping module. This module uses anomalous features as its core input. It incorporates a trained control strategy mapping relationship, which can take the form of a lookup table or a lightweight neural network decision-maker. This mapping relationship defines the optimal vehicle control command set corresponding to different anomalous modes (represented by the dimensions and values ​​of the anomalous features), such as independent braking or torque reduction commands for single-wheel slip anomalies, diagonal wheel torque distribution commands for yaw instability anomalies, or pre-tightening torque limiting commands for predictive slip risks. Based on the input anomalous feature vector, the module matches or calculates the corresponding target control parameters (such as target torque values ​​for each wheel, target braking force, differential lock status, etc.), thereby generating a specific, executable sequence of control commands, providing a clear action basis for the final vehicle execution.

[0052] In practice, vehicle control is executed through a vehicle-level actuator collaborative control module. This module receives the control command sequence and parses it into real-time target setpoints for each specific actuator (such as the motor controller, brake pressure modulation valve, differential controller, suspension controller, etc.). For example, if the control requires torque limiting on the left front wheel and applying slight braking to the right rear wheel simultaneously, a command to reduce the target torque on the left front wheel is sent to the drive motor controller, while a command to establish specific braking pressure on the right rear wheel is sent to the hydraulic control unit of the Electronic Stability Program (ESP). During execution, commands are sent to each controller via the vehicle's CAN bus or a dedicated drive-by-wire channel, and the feedback status of the actuators (such as actual torque, actual wheel speed, and braking pressure) is monitored in real time, forming a high-speed closed-loop control circuit. This ensures that control actions are executed accurately and promptly, ultimately achieving proactive and precise intervention in the vehicle's dynamics.

[0053] The above scheme achieves an effective representation of the vehicle's real-time operating status by acquiring vehicle state data and extracting features, providing a data foundation for subsequent accurate analysis. By inputting vehicle state features into a wheel-end relationship recognition model for wheel-end topology construction and feature fusion, a deep modeling and understanding of the interactions between wheels is achieved, enabling the identification of coupling anomalies in wheel-end cooperative states. By inputting vehicle state features into a slip trend prediction model for slip trend prediction, the risk of future instability is assessed. By determining abnormal vehicle characteristics based on wheel-end state features and slip trend prediction features, a comprehensive and accurate diagnosis of the vehicle's dynamic state is achieved. By determining and executing corresponding vehicle control methods based on abnormal characteristics, proactive and precise intervention in potential risks is realized, thereby effectively improving the vehicle's passability, safety, and overall control performance under complex operating conditions.

[0054] In some embodiments, in S102, the vehicle state features are input into the wheel-end relationship recognition model, and the wheel-end state features are obtained by topological construction and feature fusion of the vehicle state features.

[0055] In specific implementation, the wheel-end relationship recognition model is preferably a dynamic topological relationship modeling model based on graph neural networks. During the model construction phase, massive amounts of historical sensor data on vehicle operation under various conditions (such as different road surface adhesions, different steering and acceleration / deceleration conditions) are collected. After feature extraction processing, corresponding time-series vehicle state feature sequences are obtained and used as training data. Simultaneously, the actual dynamic states of each wheel within the corresponding time period (such as slip ratio and torque distribution relationships) are recorded and labeled. During the training phase, a wheel topology graph is first dynamically constructed based on the input features (with wheels as nodes and edge weights calculated based on state differences). Then, the internal graph convolutional network layer learns the rules for information aggregation and transmission on the graph structure to extract wheel-end association features implied in the topology. The training process iteratively optimizes the model parameters by minimizing the error (such as mean squared error or cross-entropy) between the wheel-end state representation output by the model and the actual dynamic state. This process is repeated, and cross-validation is performed on an independent validation set, ultimately resulting in a wheel-end relationship recognition model with high recognition accuracy and strong generalization ability.

[0056] Thus, after acquiring vehicle state features in real-time applications, the wheel-end relationship recognition model receives this input. The internal processing flow first dynamically constructs or updates the wheel-end topology graph structure based on the features. Next, a graph convolutional network performs multi-layer spatial information aggregation on this topology graph, enabling the features of each wheel node to fuse with the state information of its neighboring nodes, thereby encoding the interaction relationships between the wheel ends. Finally, the high-order features obtained through graph convolution, which contain deep correlation information, are fused with the original input features (e.g., through residual connections or feature concatenation) to generate a comprehensive feature representation, which is the wheel-end state feature.

[0057] S201, Constructing a wheel-end topology map of the vehicle based on vehicle state features, including: S2011 uses each wheel of a vehicle as a node in the topology graph.

[0058] In practice, each node in the topology graph corresponds one-to-one with a physical wheel of the vehicle, forming a set of four nodes: left front wheel (FL), right front wheel (FR), left rear wheel (RL), and right rear wheel (RR). Each node serves as the basic unit for information aggregation and transmission in the graph, representing the dynamic state of the corresponding wheel end.

[0059] S2012, acquire the real-time slip ratio and real-time torque data of each wheel in the vehicle state characteristics.

[0060] In practice, real-time vehicle status characteristics are acquired at the current moment or within the current time window. The real-time slip ratio is calculated based on the rotational speed of each wheel and the vehicle's reference speed, and smoothed using algorithms such as Kalman filtering to reduce noise. Real-time torque data can be directly obtained by parsing high-frequency CAN messages from the motor controller or electric drive axle controller, containing drive torque commands and actual feedback torque values. To ensure data timeliness and accuracy, this embodiment uses a 1kHz sampling frequency to simultaneously acquire the slip ratio and torque data of all four wheels.

[0061] S2013 uses the real-time slip ratio and real-time torque data of each wheel as the node features of the corresponding node for each wheel.

[0062] In practice, for each node in the topology graph, its corresponding real-time slip ratio (s_i) and real-time torque (τ_i) data are combined into a two-dimensional feature vector (s_i, τ_i), which serves as the node feature of that node.

[0063] S2014, determine the slip ratio difference between each wheel based on the real-time slip ratio.

[0064] In practice, the absolute difference in slip ratio between any two nodes in the diagram (corresponding to wheels i and j) is calculated as Δs_ij = |s_i - s_j|. For example, the difference in slip ratio between the left front wheel and the right rear wheel is calculated as Δs_FL - RR.

[0065] S2015 determines the torque difference between each wheel based on real-time torque data.

[0066] In practice, the absolute torque difference Δτ_ij = |τ_i - τ_j| between any two nodes in the diagram (corresponding to wheels i and j) is calculated. When torque vector distribution or torque redistribution is triggered by single-wheel misalignment, the torque difference can effectively reflect the load transfer and changes in drive strategy between the wheel ends.

[0067] S2016, determine the edge weights between nodes based on the slip ratio difference and torque difference between each wheel.

[0068] In practice, the edge weight w_ij is used to quantify the strength of the dynamic association between node i and node j. A weighted function of the difference in slip ratio and torque is used to determine the edge weight, for example: w_ij = exp(-(α×Δs_ij + β×Δτ_ij)), where α and β are trainable or preset positive weight coefficients. This function results in a larger weight for the connecting edge between two wheels whose slip ratio and torque states are more similar (smaller difference), indicating a stronger state association. In a single-wheel decoupling scenario, the edge weight between the decoupling wheel and other wheels will decrease due to the increased state difference, thus exhibiting a relatively isolated trend in the topology graph.

[0069] S2017 constructs the wheel-end topology graph of the vehicle based on nodes, node features, and edge weights.

[0070] In practice, the four nodes defined in the preceding steps, the two-dimensional feature vectors (s_i, τ_i) of each node, and the edge weights w_ij calculated between all node pairs are collectively used to construct a weighted undirected graph G=(V, E, W), where V is the set of nodes, E is the set of edges, and W is the set of edge weights. This graph is the wheel-end topology graph, which is a dynamic structure that is updated every 1ms as the vehicle state changes, capturing changes in the coupling relationships between the wheel ends in real time.

[0071] S202, spatial aggregation and convolution processing are performed on the wheel-end topology map to obtain wheel-end association features, including: S2021, based on the edge weights in the wheel-end topology graph, aggregate the node features of each node to obtain aggregated features.

[0072] In practice, for each node in the graph (e.g., the left front wheel FL), the feature information of all its neighboring nodes (FR, RL, RR) is aggregated. The aggregation operation uses a weighted summation method: H_i^(agg) = Σ_{j∈N(i)}w_ij × H_j, where H_j is the feature vector of neighboring node j, w_ij is the edge weight, and N(i) is the set of neighbors of node i (usually including all other nodes). The aggregated feature H_i^(agg) contains state information from other wheels. After edge weight modulation, the neighbor whose state differs more from the target node's own contributes less.

[0073] S2022, the aggregated features are subjected to nonlinear transformation and feature mapping to obtain the wheel-end associated features.

[0074] In practice, the obtained aggregated features H_i^(agg) are input into a learnable neural network layer for nonlinear transformation and feature mapping, for example: H_i'=σ(H_i^(agg)×θ+b), where θ is a trainable weight matrix, b is a bias vector, and σ is a nonlinear activation function (such as ReLU). This operation corresponds to the feature transformation step in graph convolution, which can extract higher-order and more abstract inter-wheel association patterns. After one or more such graph convolution operations, the feature H_i' output by each node is the wheel-end association feature.

[0075] S203, fuse wheel-end correlation features with vehicle state features to obtain wheel-end state features.

[0076] In practice, the output wheel-end association features (high-order, contextual features) are fused with the original node features. The fusion is performed using feature concatenation: for each wheel i, its final wheel-end state feature F_i = [H_i', (s_i, τ_i)]. The resulting F_i is a comprehensive feature vector that combines precise low-level state with high-level association semantics. As a deep representation of the current wheel-end collaborative working state, it will be fed into the subsequent fusion and decision-making modules for anomaly detection.

[0077] As a preferred embodiment, the training process of this wheel-end relationship recognition model is as follows: (1) Construct an initial wheel-end relationship recognition model. Based on the graph neural network architecture, construct a dynamic topology relationship modeling model that includes a dynamic topology graph construction module, a multi-layer graph convolutional network layer, and a feature fusion module. The graph convolutional layer can be implemented using graph attention mechanism or spectral graph convolution, etc., to aggregate spatial information on the constructed wheel-end topology graph.

[0078] (2) An initial training dataset was constructed based on actual operating data collected from a large number of test vehicles of automakers under various typical working conditions (such as constant speed driving on high-adhesion paved roads, acceleration on low-adhesion ice surfaces, braking on split-road surfaces, and single-wheel encounter with low adhesion on off-road surfaces). Multi-dimensional sensor data, including wheel speed, torque, slip ratio, and the vehicle's lateral and longitudinal acceleration and yaw rate, along with corresponding timestamps, were extracted synchronously from each wheel via the vehicle sensor network and CAN bus. These data samples were processed through a feature extraction process to obtain the corresponding time-series vehicle state feature sequences. Simultaneously, the actual dynamic states of each wheel within the corresponding time period (such as "good adhesion," "slight slip," and "severe slip and loss of adhesion") and the inter-wheel torque distribution relationship were labeled from the original data using a high-precision dynamic model and expert rules as supervisory labels. The processed samples constituted the basic training sample set.

[0079] (3) Perform preprocessing for each training sample. The vehicle state feature sequence is truncated using a sliding window to form multiple fixed-length input sequence segments. The corresponding real dynamic state labels are simultaneously truncated and aligned. To unify the input scale, all features are normalized. The preprocessed input sequence segments are used as the model input.

[0080] (4) Input the constructed input sequence fragment into the initial wheel-end relationship recognition model. The model first dynamically constructs a wheel-end topology map according to the features at the current time step S201 to S2017. Then, the graph convolutional network layer (S202) performs information transfer and aggregation on the topology map to learn and represent the spatial association patterns between wheels. The feature fusion module (S203) fuses the high-order association features output by the graph convolution with the original input features to generate the final wheel-end state features as the model output.

[0081] (5) A multi-task objective function combining state representation error loss function and classification loss function is used for optimization. State representation error loss (e.g., mean squared error) supervises the degree to which the wheel end state features output by the model fit the real dynamic state in the latent space; classification loss function (e.g., cross-entropy loss) supervises the accuracy of the model in classifying the attachment state of each wheel based on the wheel end state features. The total loss value is calculated, and all learnable parameters of the model are updated through backpropagation algorithm combined with an adaptive optimizer (e.g., AdamW), including the weight coefficients α and β in the topological graph edge weight calculation and the parameters of the graph convolutional layer.

[0082] (6) Repeat steps (4) and (5) to train the model iteratively multiple times using the basic training sample set. After each training round, evaluate the model performance using an independent validation set that covers multiple unseen road surfaces and driving modes, and monitor changes in indicators such as the similarity index of wheel end state representation, the accuracy, precision, and recall of attachment state classification. When the model's performance on the validation set tends to stabilize and the risk of overfitting is low, end this stage of training to obtain the pre-trained model.

[0083] (7) To improve the model's recognition accuracy and generalization robustness under complex, extreme, or transient conditions (such as a single wheel instantly crossing an ice surface during high-speed driving, emergency obstacle avoidance on a low-adhesion road surface, and sudden changes in wheel-end coupling under active intervention of the torque vector system), an additional challenging condition sample set is constructed. This challenging condition sample set focuses on the above-mentioned scenarios with intense dynamics, blurred state boundaries, or multiple superimposed disturbances, and performs more refined data cleaning and state labeling.

[0084] (8) Fine-tune the pre-trained model using a challenging set of working conditions. At this stage, a lower learning rate is used for training, and some of the underlying feature extraction layers may be frozen, so that the model can be further refined to improve its ability to model and recognize extreme or rare coupling patterns based on the learned general wheel-end relationship patterns.

[0085] (9) The fine-tuned model was finally evaluated on a completely independent comprehensive test set covering the entire operating condition spectrum. Evaluation metrics included the F1 score for wheel-end attachment state classification, the detection latency for single wheel deattachment events (required to be less than 5ms), the false alarm rate, and the utility improvement of the wheel-end state features output by the model in the downstream control strategy learning task. Finally, the model whose performance met the preset accuracy, real-time performance, and generalization thresholds was determined as the deployable wheel-end relationship recognition model.

[0086] Through the aforementioned technical solution, dynamic topology modeling based on graph neural networks directly constructs and learns the dynamic relationship structure between wheel ends from the original vehicle state features. Graph convolution operations aggregate neighbor information in the spatial dimension, thereby encoding deep-level wheel-to-wheel interaction relationships. Finally, by fusing with the original features, a comprehensive wheel-end state representation is formed, possessing both accurate low-level state and high-level semantic relationships. This process effectively overcomes the limitations of traditional methods that rely on isolated analysis of single-wheel states or fixed physical models, making it difficult to capture complex, time-varying coupling relationships. It can also more sensitively capture global dynamic correlation variations caused by single-wheel deattachment. The resulting wheel-end state features provide accurate and structurally rich input for subsequent anomaly comprehensive judgment and intelligent torque collaborative control, significantly improving the vehicle's stability control and passability optimization potential and reliability under extreme conditions.

[0087] In some embodiments, in S103, vehicle state features are input into the slip trend prediction model, and slip trend prediction features of the vehicle are obtained by performing slip trend prediction on the vehicle state features.

[0088] In practice, the slippage trend prediction model is preferably a sequence prediction model based on a temporal deep neural network. It consists of a parameterized temporal coding network that can learn the dynamic evolution of wheel slippage state.

[0089] During the model building phase, massive amounts of historical sensor data on vehicle operation under various conditions (such as acceleration, braking, and steering under different surface adhesions) are collected. After feature extraction, corresponding time-series vehicle state feature sequences are obtained and used as training data. Simultaneously, the actual evolution trajectory of each wheel's slip ratio within the corresponding time period is recorded and labeled. During the training phase, the vehicle state feature sequences (especially slip ratio and related dynamic parameter sequences) are input into the model. The model learns the mapping relationship from historical feature sequences to future slip ratio trends through its internal temporal coding network (such as a bidirectional long short-term memory network BiLSTM or a Transformer encoder). By learning temporal dependency patterns and abrupt change features in the historical sequences, the model predicts the trajectory of each wheel's slip ratio change and the probability of abrupt change risk over a future period. The training process iteratively optimizes the model parameters by minimizing the error between the predicted trend and the actual slip evolution. This process is repeated and cross-validated on an independent validation set, ultimately resulting in a slip trend prediction model with high prediction accuracy and strong generalization ability.

[0090] Thus, after acquiring vehicle state characteristics in real-time applications, the slip trend prediction model receives this input. Its internal temporal coding network extracts and encodes the temporal pattern of slip state evolution based on the feature sequences of the current and historical moments, calculating the predicted trajectory of slip ratio changes and the corresponding probability of abrupt changes over a future period. Subsequently, features directly related to vehicle stability (e.g., predicted slip ratios of each wheel within the next 50 milliseconds, slip ratio change rate, and abrupt change probability intensity) are extracted from this predicted trajectory. These features constitute the slip trend prediction features. These slip trend prediction features provide crucial trend information input for subsequent anomaly comprehensive judgment and forward-looking control.

[0091] S301, obtain the time series sequence of slip ratio of each wheel from the vehicle state characteristics.

[0092] In practice, the slip ratio values ​​of each wheel (left front FL, right front FR, left rear RL, right rear RR) within a continuous time window are separated from the vehicle state characteristics. This slip ratio time series is usually constructed based on a preset sampling frequency (e.g., 1kHz) and window length (e.g., 500ms), forming a slip ratio data sequence {s_i(t), s_i(t-Δt), s_i(t-2Δt),...} containing multiple consecutive time points, where i represents the wheel number.

[0093] S302, perform first-order difference processing on the slip rate time series to obtain the corresponding slip rate difference series.

[0094] In practice, for the slip ratio time series of each wheel, the difference in slip ratio between adjacent time points is calculated, i.e., a first-order difference operation is performed: Δs_i(t) = s_i(t) - s_i(t - Δt). This operation is performed on all time points to obtain a difference sequence {Δs_i(t), Δs_i(t - Δt), ...} that reflects the instantaneous rate and direction of change of slip ratio. This difference sequence can effectively amplify the abrupt change signal of slip ratio and filter out some low-frequency trend or steady-state noise.

[0095] S303 performs bidirectional time series modeling on the slip ratio time series to obtain slip ratio time series features.

[0096] In practice, the original slip ratio time series sequence {s_i(t), s_i(t-Δt),...} for each wheel is input into a bidirectional temporal neural network (such as BiLSTM) for encoding. The bidirectional temporal neural network processes the sequence simultaneously in both forward and backward directions, effectively capturing the causal influence of past states on future states, as well as the potential prediction of the current state by future context. The final hidden state or the output of a specific layer of the bidirectional temporal neural network is extracted as a high-dimensional feature vector. This feature vector is the slip ratio temporal feature, which deeply encodes the evolution pattern, periodicity, and potential trend information of the wheel's slip ratio over time.

[0097] S304, the slip rate time series features are fused with the corresponding slip rate difference series to obtain slip trend features.

[0098] In practice, the obtained slip rate temporal features (high-dimensional, semantically rich encoded features) are fused with the obtained slip rate difference sequence (a numerical sequence reflecting recent instantaneous changes). The fusion method can be to map the difference sequence to the same dimension as the temporal features through a fully connected layer and then add them element-wise (feature addition), or to directly concatenate the two along the feature dimension (feature concatenation). The resulting feature vector is the slip trend feature, which simultaneously contains both the long-term dependency pattern of slip rate evolution and short-term mutation dynamics, providing a more comprehensive input for the final mutation probability prediction.

[0099] S305, based on slip trend characteristics, determines the slip abrupt change probability of each wheel, and uses the slip abrupt change probability as a slip trend prediction feature, including: S3051, perform time-series abrupt change analysis on slip trend characteristics to obtain the slip rate change trend intensity of each wheel within a preset time window.

[0100] In practice, the fused slip trend features are input into a specially designed temporal abrupt change analysis module. This module can be a shallow neural network or a learnable attention mechanism. It analyzes the temporal patterns encoded in the features, focusing on assessing the likelihood of a rapid, non-stationary change in the slip ratio within a predetermined time window (e.g., the next 20ms). The temporal abrupt change analysis module outputs a scalar value I_i, representing the strength of the slip ratio change trend; a higher value indicates a greater risk of a sharp deterioration in slip in the near future.

[0101] S3052, based on the intensity of the slip ratio change trend and the preset slip mutation threshold, determines the slip mutation probability of each wheel.

[0102] In practice, the calculated trend strength I_i is converted into a probability value P_i between 0 and 1 through a probability mapping function, i.e., the slippage probability. A common implementation is to use the sigmoid function: P_i = 1 / (1 + exp(-k × (I_i - θ))), where θ is a preset slippage threshold (an adjustable parameter), and k is a slope factor that controls the steepness of the probability curve. When the trend strength I_i significantly exceeds the threshold θ, the slippage probability P_i will rapidly approach 1.

[0103] S3053 uses the slip mutation probability as a slip trend prediction feature.

[0104] In practice, the calculated slip change probabilities P_i (i=FL,FR,RL,RR) of each wheel are organized into a vector [P_FL,P_FR,P_RL,P_RR]. This vector is the final slip trend prediction feature. This feature directly and quantitatively characterizes the predicted risk of each wheel experiencing instability and slippage in the near future, providing a clear and crucial input signal for subsequent comprehensive vehicle anomaly assessment and forward-looking control decisions.

[0105] As a preferred embodiment, the training process of this slip trend prediction model is as follows: (1) Constructing an initial slip trend prediction model. Based on the temporal deep neural network architecture, a sequence prediction model is constructed, which includes a bidirectional temporal coding network, a feature fusion module, a temporal mutation analysis module, and a probability mapping module. Among them, the bidirectional temporal coding network can be implemented using a BiLSTM or Transformer encoder to perform deep feature extraction on the slip rate temporal sequence.

[0106] (2) Based on the actual operating data collected from a large number of test vehicles of car manufacturers under various typical working conditions (such as rapid acceleration on high-adhesion roads, braking on low-adhesion ice surfaces, steady-state driving on split roads, and continuous bumpy off-road roads), an initial training dataset is constructed. Through the vehicle sensor network and CAN bus, sensor data such as the slip ratio, wheel speed, driving torque, and longitudinal acceleration of the whole vehicle, as well as the corresponding timestamps, are extracted synchronously. These data samples are processed through a feature extraction process to obtain the corresponding time-series vehicle state feature sequence. At the same time, based on high-precision reference signals and differential calculations, the true change trend of the slip ratio of each wheel within a short-term time window (such as the next 50ms) (such as "stable", "gradual increase", "sharp increase") and whether a sudden change occurs are labeled as supervision signals. The processed samples constitute the basic training sample set.

[0107] (3) For each training sample, perform preprocessing. Extract the slip ratio time series subsequence of each wheel from the vehicle state feature sequence (S301). Perform first-order differencing on the slip ratio time series subsequence to obtain the difference sequence (S302). Standardize or normalize the slip ratio time series. Use the preprocessed slip ratio time series, the difference sequence, and the corresponding future trend labels as the input and supervision target of the model.

[0108] (4) Input the constructed input sequence into the initial slip rate prediction model. The model first encodes the slip rate time series through a bidirectional temporal coding network (S303) to obtain the slip rate time series features. Then, the feature fusion module (S304) fuses the slip rate time series features with the difference sequence to form slip rate trend features. Next, the temporal mutation analysis module (S3051) analyzes the slip rate trend features and outputs the slip rate change trend intensity. Finally, the probability mapping module (S3052) maps the trend intensity to slip mutation probability as the model output (S3053).

[0109] (5) A multi-task objective function combining trend prediction error loss function and mutation classification loss function is used for optimization. Trend prediction error loss (such as smoothing L1 loss) supervises the model's prediction accuracy of the future short-term trajectory of slip rate; mutation classification loss function (such as focal loss) focuses on supervising the model's classification accuracy of whether slip mutation has occurred, especially the ability to identify minority classes (mutation events). The total loss value is calculated, and all learnable parameters of the model are updated through backpropagation algorithm combined with an adaptive optimizer (such as Adam), including the parameters in the temporal coding network, fusion layer, mutation analysis module, and probability mapping function.

[0110] (6) Repeat steps (4) and (5) to train the model iteratively multiple times using the basic training sample set. After each training round, evaluate the model performance using an independent validation set that covers multiple unseen driving scenarios and road conditions, and monitor changes in metrics such as mean absolute error of trend prediction, precision and recall of mutation detection, and F1 score. When the model's performance on the validation set tends to stabilize and the risk of overfitting is low, end this stage of training to obtain the pre-trained model.

[0111] (7) To improve the prediction accuracy and robustness of the model under complex, extreme or transient conditions (such as drastic oscillations in slip ratio when ABS / TCS intervenes frequently, rapid changes in slip ratio caused by oversteering / understeering on roads with extremely low adhesion, and slip impact at the moment of vehicle landing after a jump), an additional challenging condition sample set is constructed. This challenging condition sample set focuses on the above-mentioned scenarios with severe dynamics, non-stationary signals or low signal-to-noise ratios, and performs more refined data alignment and label correction.

[0112] (8) Fine-tune the pre-trained model using a challenging set of working conditions. At this stage, a lower learning rate is used for training, and gradient freezing may be applied to the front-end feature extraction part of the model, so that the model can be further refined to capture and predict extreme or rare mutation patterns based on the learned general slip evolution pattern.

[0113] (9) The fine-tuned model was finally evaluated on a completely independent comprehensive test set covering the entire operating condition spectrum. Evaluation indicators included the root mean square error of the slip rate prediction in the next 20ms, the detection delay of abrupt events (required to be less than 10ms), the false alarm rate, and the contribution of the slip trend prediction features output by the model to the downstream anomaly detection model. Finally, the model whose performance met the preset accuracy, timeliness, and generalization thresholds was determined as the deployable slip trend prediction model.

[0114] The above technical solution, based on a hybrid architecture of deep temporal networks and feature fusion, directly learns and predicts future evolutionary trends and mutation risks from historical slip rate sequences and their differential signals. Bidirectional encoding fully mines temporal contextual information, and the differential signal enhances sensitivity to mutations. Finally, an intuitive mutation probability is output through a learnable analytical framework. This process effectively overcomes the limitations of traditional slip warning methods based on fixed thresholds or simple moving averages, which suffer from strong lag and poor adaptability, and can capture early and reliable precursor signals of slip deterioration. The resulting slip trend prediction features provide crucial, quantified risk warning inputs for subsequent comprehensive diagnosis of vehicle abnormal states and proactive stability control, significantly improving the vehicle's stability control boundaries and active safety performance under extreme adhesion conditions.

[0115] In some embodiments, in S104, based on wheel-end state characteristics and slip trend prediction characteristics, abnormal characteristics of the vehicle are generated, including: S401, based on the wheel end state characteristics, determine the first abnormal feature of each wheel.

[0116] In specific implementation, the wheel end state features F_i of each wheel output by the wheel end relationship recognition model are respectively input into a first anomaly detection network (e.g., a lightweight multilayer perceptron MLP). After training, the first anomaly detection network can identify whether there is an anomaly based on the current topological relationship perspective from the comprehensive information of "the wheel end's own state and its relationship with other wheels" encoded by F_i. The output of the first anomaly detection network is the first anomaly feature A_i^(1), which can be in the form of a scalar (anomaly score), a multidimensional vector (representing the confidence of multiple anomaly types), or a category label (such as "normal", "coupling instability", "torque unevenness").

[0117] S402, based on slip trend prediction features, determines the second abnormal feature of each wheel.

[0118] In practice, the slip mutation probability P_i of each wheel output by the slip trend prediction model is used as the core input. P_i itself is a quantified risk prediction. In order to form an expression that matches the dimension or semantics of the first anomalous feature, P_i can be input into a second anomalous mapping module. This second anomalous mapping module can be a simple function (such as linear scaling) or a small neural network. Its purpose is to transform and expand the predicted probability P_i into a second anomalous feature A_i^(2), so that it can not only reflect the "level of future slip risk", but also characterize the anomalous type or severity inferred from the trend prediction in a format compatible with A_i^(1).

[0119] S403, the first and second abnormal features of the same wheel are fused to obtain the fused abnormal features of each wheel.

[0120] In practice, for each wheel i, its first abnormal feature A_i^(1) and second abnormal feature A_i^(2) are fused. The fusion method depends on the specific form of the two features. Common methods include: feature concatenation (C_i=Concat(A_i^(1),A_i^(2))), weighted summation (C_i=α×A_i^(1)+β×A_i^(2), where α and β are learnable or preset weights), or attention-based fusion (dynamically calculating the importance weights of the two features and then weighting them together). The feature vector C_i obtained after fusion is the fused abnormal feature of the wheel. It integrates two types of information: "abnormal wheel end correlation state at the current moment" and "abnormal future slip trend prediction", providing a more comprehensive and reliable basis for judging the abnormal condition of the wheel.

[0121] S404 aggregates the fusion anomaly features of each wheel to obtain the vehicle's anomaly features.

[0122] In practice, the fused anomalous features C_FL, C_FR, C_RL, and C_RR of each of the four wheels are aggregated to form vehicle anomalous features representing the overall vehicle-level anomalous state. The aggregation operation needs to capture the spatial correlation and overall impact between the anomalous states of the four wheels. Specific methods can include: global pooling (e.g., max pooling to highlight the most severe anomaly, or average pooling to reflect the overall anomalous level); encoding aggregation through a fully connected network to learn the relationship between the four feature vectors and output a comprehensive vector; or constructing a lightweight graph network again, using wheels as nodes and fused anomalous features as node features, and aggregating the entire vehicle information through a single graph convolution layer. The final output vehicle-level feature vector is the vehicle's anomalous feature, a high-order, compact encoded representation of the vehicle's current and recent potential overall dynamic anomalous states.

[0123] By independently fusing the current associated anomalies and future trend risks for each wheel, the local accuracy and information complementarity of anomaly detection are ensured. Subsequently, through intelligent aggregation of the fused features of all wheels, a higher-dimensional and generalized understanding of the vehicle's anomaly state is achieved, moving from local anomalies to the overall vehicle anomaly state. This effectively integrates the spatial structure analysis capabilities of the wheel-end relationship recognition model with the temporal forecasting capabilities of the slippage trend prediction model, overcoming the limitations of a single model's limited perspective. It generates more comprehensive and discriminative vehicle anomaly features, providing crucial decision input for ultimately generating accurate and coordinated vehicle control strategies.

[0124] In some embodiments, the process of determining the second abnormal feature in S402 includes: S4021, Based on the slip trend prediction features, determine the slip change probability of each wheel.

[0125] S4022, determine the average slip change probability based on the slip change probability of each wheel.

[0126] In practice, since the slip trend prediction features include the slip mutation probability of each wheel, the average slip mutation probability is obtained by summing the slip mutation probabilities of all wheels and dividing by the number of wheels. This average slip mutation probability can characterize the overall trend of slip mutation of each wheel of the vehicle.

[0127] S4023, For each wheel, determine the deviation value between the slip change probability of the wheel and the average slip change probability, and use the deviation value as the second abnormal feature of the wheel.

[0128] In practice, to accurately determine the individual deviations of each wheel's slip change, the deviation value between the wheel's slip change probability and the average slip change probability is calculated. This deviation value can be used as a second anomalous feature of that wheel. This deviation value characterizes the individual slip deviation of that wheel relative to other wheels; the larger the deviation value, the greater the slip deviation of that wheel from other wheels, and the higher the corresponding level of vehicle driving danger.

[0129] The above method can determine the deviation value of the average slip change probability of each wheel relative to the wheel as a whole. Based on this deviation value, the individual slip deviation of the wheel relative to other wheels can be accurately determined, so that the deviation value as the second abnormal feature can better reflect the slip change between wheels.

[0130] In some embodiments, S105, determining the vehicle control method based on abnormal characteristics includes: S501, based on abnormal features, determines the current adhesion state and slip change intensity of each wheel.

[0131] In practice, vehicle anomaly features characterizing the abnormal state of the vehicle's dynamics are input into a state decoder. This state decoder is typically a pre-trained lightweight neural network that decodes and maps high-dimensional, abstract anomaly feature vectors back to specific, physically interpretable wheel-end state parameters. The decoder's output for each wheel contains two key pieces of information: first, a classification or quantification score of the current adhesion state (e.g., a continuous value from 0 to 1 representing the degree of "fully attached" to "fully unattached," or a discrete classification of "good," "marginal," or "unattached"), and second, a specific numerical value of the slip mutation intensity (this value is directly derived from or related to the mutation probability in the slip trend prediction features, but is integrated and calibrated to a unified decision scale here).

[0132] S502 inputs the current adhesion state and slip change intensity of each wheel into the torque control mapping matrix to obtain the corresponding torque control command.

[0133] In practice, the torque control mapping matrix is ​​a predefined or online-learnable strategy lookup table or function mapping relationship. Its core logic defines the baseline torque adjustment actions to be taken for different combinations of "adhesion state-slip mutation intensity." The input is the two-dimensional state of each wheel (adhesion state value, mutation intensity value), and the output is the torque adjustment command for that wheel, typically including the torque adjustment direction (increase, maintain, decrease), adjustment magnitude (percentage or absolute value), and the urgency of the adjustment (response rate). For example, when a wheel is determined to be on the "adhesion edge" and has "high slip mutation intensity," the mapping matrix might output a command to "rapidly reduce the torque of this wheel by 50%."

[0134] S503 obtains the current operating parameters of the vehicle, corrects the torque control command based on the current operating parameters, and generates the corrected torque control command.

[0135] In practice, the system acquires real-time vehicle operating parameters from the vehicle bus. These parameters include, but are not limited to: vehicle speed, steering wheel angle, yaw rate, longitudinal / lateral acceleration, current gear, driver power request (e.g., accelerator pedal opening), and braking status. These parameters reflect the vehicle's overall motion and the driver's intent. Subsequently, an operating condition adaptation module uses these parameters to coordinate and correct the reference torque control commands for each wheel. The goal of this correction is to ensure that the torque adjustment of a single wheel aligns with the vehicle's dynamic objectives (e.g., maintaining driving stability, following the driver's steering intent, and optimizing overall traction), avoiding vehicle instability caused by localized adjustments. For example, during high-speed cornering, even if the inner drive wheel is at risk of slippage, excessive torque reduction could lead to insufficient driving force and understeer. Therefore, the correction module will appropriately reduce the reduction magnitude and may compensate by fine-tuning the torque of the outer wheels.

[0136] S504 determines the vehicle control method based on the modified torque control command.

[0137] In practice, the final torque control commands for each wheel, after being adjusted for operating conditions, are packaged and formatted with other necessary chassis control commands (such as differential lock locking requests, braking system intervention requests, suspension stiffness adjustment requests, etc.) to form a complete "control mode" instruction set that can be understood and executed by the vehicle's underlying actuators. This instruction set specifies the target values ​​or action sequences of each actuator in the next control cycle. For example, the control mode might ultimately be expressed as: {Left front motor: torque output set to -200Nm; Right front motor: torque output maintained; Rear axle electronic differential lock: lock requested; ESP: prepare for slight braking intervention on the right rear wheel}.

[0138] The above technical solution concretizes comprehensive anomaly information through state decoding, achieves rapid response using preset torque control mapping, and then incorporates vehicle operating parameters for global coordination and correction, ultimately generating a control strategy that balances local risk mitigation with overall vehicle dynamic performance optimization. This process ensures that the control system can react to sudden anomalies within milliseconds and intelligently integrate this local intervention into overall vehicle dynamic management, thereby effectively improving vehicle passability and safety while guaranteeing driving smoothness and controllability.

[0139] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides a vehicle control device.

[0140] refer to Figure 2 Vehicle control device, including: The data acquisition and feature extraction module 210 is used to acquire vehicle status data and extract features from the vehicle status data to obtain vehicle status features. The wheel-end state feature generation module 220 is used to input vehicle state features into the wheel-end relationship recognition model, and obtain wheel-end state features by performing wheel-end topology construction and feature fusion on the vehicle state features; The slip trend prediction feature generation module 230 is used to input vehicle state features into the slip trend prediction model, and obtain the vehicle slip trend prediction features by performing slip trend prediction on the vehicle state features. The abnormal feature determination module 240 is used to determine the abnormal features of the vehicle based on wheel end state features and slip trend prediction features. The control mode determination module 250 is used to determine the vehicle's control mode based on abnormal characteristics and to control the vehicle according to the control mode.

[0141] In some embodiments, the wheel end state feature generation module 220 is configured to: Construct a wheel-end topology map of the vehicle based on vehicle state characteristics; Spatial aggregation and convolution processing are performed on the wheel end topology graph to obtain wheel end association features; The wheel-end associated features are fused with the vehicle state features to obtain the wheel-end state features.

[0142] In some embodiments, the wheel end state feature generation module 220 is further configured to: Each wheel of the vehicle is used as a node in the topology graph; Obtain real-time slip ratio and real-time torque data of each wheel in the vehicle state characteristics; The real-time slip ratio and real-time torque data of each wheel are used as the node features of the corresponding node of each wheel. The difference in slip ratio between each wheel is determined based on the real-time slip ratio. The torque difference between each wheel is determined based on real-time torque data; The edge weights between nodes are determined based on the differences in slip ratio and torque between each wheel; Based on nodes, node features, and edge weights, a wheel-end topology graph of the vehicle is constructed.

[0143] In some embodiments, the wheel end state feature generation module 220 is further configured to: Based on the edge weights in the wheel-end topology graph, the node features of each node are aggregated to obtain aggregated features; The aggregated features are subjected to nonlinear transformation and feature mapping to obtain the wheel-end associated features.

[0144] In some embodiments, the slip trend prediction feature generation module 230 is configured to: Obtain the time series sequence of slip ratio of each wheel from the vehicle state characteristics; The slip ratio time series is subjected to first-order differencing to obtain the corresponding slip ratio difference series; Bidirectional time series modeling is performed on the slip ratio time series to obtain slip ratio time series features; By fusing the time-series features of slip ratio with the corresponding slip ratio difference sequence, slip trend features are obtained; Based on the slip trend characteristics, the slip change probability of each wheel is determined, and the slip change probability is used as the slip trend prediction feature.

[0145] In some embodiments, the slip trend prediction feature generation module 230 is further configured to: A time-series abrupt change analysis was performed on the slip trend characteristics to obtain the intensity of the slip ratio change trend of each wheel within a preset time window; Based on the intensity of the slip ratio change trend and the preset slip abrupt change threshold, the slip abrupt change probability of each wheel is determined; The probability of slip change is used as a feature for predicting slip trend.

[0146] In some embodiments, the anomaly feature determination module 240 is configured to: Based on the wheel end state characteristics, the first abnormal characteristic of each wheel is determined; Based on the slip trend prediction characteristics, the second anomaly characteristics of each wheel are determined; The first and second abnormal features of the same wheel are fused to obtain the fused abnormal features of each wheel. By aggregating the fusion anomaly features of each wheel, the anomaly features of the vehicle are obtained.

[0147] In some embodiments, the anomaly feature determination module 240 is further configured to: Based on the slip trend prediction characteristics, the slip change probability of each wheel is determined; The average slip change probability is determined based on the slip change probability of each wheel; For each wheel, the deviation between the slip mutation probability and the average slip mutation probability of that wheel is determined, and the deviation value is used as the second anomalous feature of that wheel.

[0148] In some embodiments, the control mode determination module 250 is configured to: Based on the abnormal characteristics, the current adhesion status and slip change intensity of each wheel are determined; The current adhesion state and slip change intensity of each wheel are input into the torque control mapping matrix to obtain the corresponding torque control command; Obtain the vehicle's current operating parameters, correct the torque control command based on the current operating parameters, and generate the corrected torque control command; The vehicle control method is determined based on the revised torque control command.

[0149] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware.

[0150] The apparatus of the above embodiments is used to implement the corresponding method in any of the foregoing embodiments and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0151] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides an electronic device, the electronic device including a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the methods described in any of the above embodiments.

[0152] Figure 3 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0153] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0154] The memory 1020 can be implemented in the form of ROM (Read-Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0155] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0156] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0157] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0158] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0159] The electronic devices described above are used to implement the corresponding methods in any of the foregoing embodiments and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0160] Based on the same inventive concept, this application also provides a vehicle including the device or electronic device described in the above embodiments. The beneficial effects of embodiments having corresponding devices or electronic devices will not be elaborated further here.

[0161] It is understood that before using the technical solutions of the various embodiments in this disclosure, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.

[0162] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose, based on the prompt message, whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media performing the operations of this disclosed technical solution.

[0163] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0164] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0165] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application (including the claims) is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in the details for the sake of brevity.

[0166] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0167] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0168] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. A vehicle control method characterized by, include: Obtain vehicle status data, extract features from the vehicle status data, and obtain vehicle status features; The vehicle state features are input into the wheel-end relationship recognition model. By performing wheel-end topology construction and feature fusion on the vehicle state features, the wheel-end state features are obtained. The vehicle state features are input into the slip trend prediction model, and the slip trend prediction features of the vehicle are obtained by performing slip trend prediction on the vehicle state features. Based on the wheel end state characteristics and the slip trend prediction characteristics, the abnormal characteristics of the vehicle are determined; Based on the aforementioned abnormal characteristics, the vehicle control method is determined, and the vehicle is controlled according to the stated control method.

2. The method of claim 1, wherein, The process involves inputting the vehicle state features into the wheel-end relationship recognition model, and obtaining wheel-end state features through topology construction and feature fusion of the vehicle state features, including: A wheel-end topology map of the vehicle is constructed based on the vehicle state characteristics; Spatial aggregation and convolution processing are performed on the wheel end topology map to obtain wheel end association features; The wheel-end associated features are fused with the vehicle state features to obtain the wheel-end state features.

3. The method of claim 2, wherein, The construction of the wheel-end topology map of the vehicle based on the vehicle state features includes: Each wheel of the vehicle is used as a node in the topology graph; Obtain the real-time slip ratio and real-time torque data of each wheel in the vehicle state characteristics; The real-time slip ratio and real-time torque data of each wheel are used as the node features of the corresponding node of each wheel. The slip ratio difference between each wheel is determined based on the real-time slip ratio. The torque difference between each wheel is determined based on the real-time torque data. The edge weights between nodes are determined based on the differences in slip ratio and torque between each wheel; Based on the nodes, node features, and edge weights, a wheel-end topology graph of the vehicle is constructed.

4. The method of claim 3, wherein, The spatial aggregation and convolution processing of the wheel end topology map to obtain wheel end association features includes: Based on the edge weights in the wheel-end topology graph, the node features of each node are aggregated to obtain aggregated features; The aggregated features are subjected to nonlinear transformation and feature mapping to obtain wheel-end associated features.

5. The method of claim 1, wherein, The step of inputting the vehicle state features into the slip trend prediction model and obtaining the vehicle's slip trend prediction features by performing slip trend prediction on the vehicle state features includes: Obtain the time sequence of slip ratio of each wheel from the vehicle state characteristics; The slip ratio time series is subjected to first-order difference processing to obtain the corresponding slip ratio difference series; Bidirectional time series modeling is performed on the slip ratio time series to obtain slip ratio time series features; The slip ratio time series features are fused with the corresponding slip ratio difference sequence to obtain slip trend features; Based on the slip trend characteristics, the slip change probability of each wheel is determined, and the slip change probability is used as the slip trend prediction feature.

6. The method of claim 5, wherein, The step of determining the slip change probability of each wheel based on the slip trend characteristics, and using the slip change probability as the slip trend prediction feature, includes: A time-series abrupt change analysis was performed on the slip trend characteristics to obtain the slip ratio change trend intensity of each wheel within a preset time window; Based on the intensity of the slip ratio change trend and the preset slip abrupt change threshold, the slip abrupt change probability of each wheel is determined; The slippage mutation probability is used as the slippage trend prediction feature.

7. The method of claim 1, wherein, The generation of vehicle anomaly features based on the wheel end state features and the slip trend prediction features includes: Based on the wheel end state characteristics, the first abnormal characteristic of each wheel is determined; Based on the slip trend prediction features, the second abnormal features of each wheel are determined; The first and second abnormal features of the same wheel are fused to obtain the fused abnormal features of each wheel. By aggregating the fusion anomaly features of each wheel, the anomaly features of the vehicle are obtained.

8. The method of claim 7, wherein, The determination of the second abnormal feature of each wheel based on the slip trend prediction feature includes: Based on the slip trend prediction features, the slip change probability of each wheel is determined; The average slip change probability is determined based on the slip change probability of each wheel; For each wheel, the deviation value between the slip mutation probability of that wheel and the average slip mutation probability is determined, and the deviation value is used as the second abnormal feature of that wheel.

9. The method of claim 1, wherein, The process of determining the vehicle control method based on the abnormal characteristics includes: Based on the aforementioned abnormal characteristics, the current adhesion state and slip change intensity of each wheel are determined. The current adhesion state and slip change intensity of each wheel are input into the torque control mapping matrix to obtain the corresponding torque control command; Obtain the current operating parameters of the vehicle, and modify the torque control command based on the current operating parameters to generate a modified torque control command; The vehicle control method is determined based on the modified torque control command.

10. A vehicle, including electronic equipment, characterized in that, The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method as described in any one of claims 1 to 9.

Images