Vehicle control method and vehicle

Abstract

This application provides a vehicle control method and a vehicle, relating to the field of vehicle chassis control technology. The method includes acquiring multi-source perception data during vehicle operation; preprocessing the multi-source perception data to obtain temporal perception data; generating an obstacle crossing structure map using a map generation model based on the temporal perception data; generating a rear suspension state prediction result and prediction error data using a predictive coding model based on the obstacle crossing structure map; generating rear suspension impact pulse data using an impact recognition model based on the rear suspension state prediction result and prediction error data; and generating rear suspension control commands based on the obstacle structure map, rear suspension state prediction result, prediction error data, and rear suspension impact pulse data. These commands are then executed by the suspension control system, achieving accurate prediction and pre-hardening control of rear suspension obstacle crossing impacts, effectively avoiding chassis impact risks in unstructured terrain.

Description

Technical Field

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

[0002] During obstacle crossing, the vehicle chassis control system often uses a fixed threshold judgment method or simplified model to trigger dynamic adjustments of the vehicle body based on the obstacle crossing status of the front suspension. This can lead to problems such as chassis impact and sudden changes in posture when the rear wheels pass over obstacles, affecting the overall vehicle passability and comfort. Summary of the Invention

[0003] In view of this, the purpose of this application is to propose a vehicle control method and a vehicle to solve problems such as chassis impact and sudden attitude changes when the rear wheels of a vehicle pass through an obstacle.

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

[0005] Multi-source perception data is acquired during vehicle operation, and the multi-source perception data is preprocessed to obtain temporal perception data characterizing the vehicle's obstacle-crossing process. Based on the time-series sensing data, an obstacle-crossing structure map is generated using a pre-trained map generation model. Based on the obstacle crossing structure map, the vehicle's rear suspension state prediction results and prediction error data are generated through a pre-trained prediction coding model. Based on the rear suspension state prediction results and the prediction error data, rear suspension impact pulse data are generated using a pre-trained impact recognition model. Based on the obstacle structure diagram, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data, a rear suspension control command is generated. The rear suspension control command is sent to the vehicle's suspension control system to execute the rear suspension control command.

[0006] Optionally, the step of generating an obstacle-crossing structure map based on the time-series-aware data using a pre-trained map generation model includes: The temporal sensing data is encoded into feature vectors of each node in the obstacle crossing structure map by the temporal encoder in the map generation model; wherein each node represents a component of the vehicle. Based on the feature vectors of each node and prior knowledge of vehicle physics, connection edges are established between each node to generate an initial structural graph. The initial structure map is input into the graph neural network in the map generation model for iterative propagation to generate the obstacle-crossing structure map.

[0007] Optionally, the step of generating the vehicle's rear suspension state prediction result and prediction error data based on the obstacle-crossing structure map and a pre-trained prediction coding model includes: Extract the temporal data of the target nodes from the obstacle crossing structure map; The encoder in the predictive coding model extracts features from the temporal data of the target node to obtain the hidden state vector. Based on the hidden state vector, multi-step forward state prediction is performed through the prediction network in the prediction coding model to generate the rear-hanging state prediction result. Based on the rear suspension state prediction results and real-time sensor data, the error is calculated through the error network in the prediction coding model to generate the prediction error data.

[0008] Optionally, the step of generating rear suspension impact pulse data based on the rear suspension state prediction result and the prediction error data using a pre-trained impact recognition model includes: The pulse coding layer in the impact recognition model converts the rear suspension state prediction result and the prediction error data into pulse sequence data. The pulse sequence data is weighted by the synaptic weight modulation layer in the impact recognition model to form weighted pulse sequence data. Based on the weighted pulse sequence data, the decision output layer in the impact identification model outputs the rear suspension impact pulse data using a membrane potential integration mechanism.

[0009] Optionally, generating rear suspension control commands based on the obstacle-crossing structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data includes: A consistency check is performed based on the obstacle crossing structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data. In response to passing the consistency check, a rear suspension control command is generated based on the rear suspension impact pulse data, so that the vehicle's suspension control system executes the rear suspension control command.

[0010] Optionally, the method further includes; In response to failing the consistency check, the rear suspension impact pulse data is corrected using the obstacle crossing structure map, the rear suspension state prediction result, and the prediction error data, and the rear suspension control command is generated.

[0011] Optionally, the consistency verification based on the obstacle-crossing structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data includes: Based on the obstacle crossing structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data, a time consistency check is performed. Based on the obstacle crossing structure map, the rear suspension state prediction results, the prediction error data, and the rear suspension impact pulse data, a consistency verification of the results is performed. In response to passing both the time consistency check and the result consistency check, it is determined that the consistency check has been passed.

[0012] Optionally, before sending the rear suspension control command to the vehicle's suspension control system, the method further includes: Obtain real-time operating status data of the vehicle; The rear suspension control command is adjusted based on the real-time operating status data, and the adjusted rear suspension control command is sent to the suspension control system.

[0013] Optionally, the training process of the impact recognition model includes: Construct the initial impact recognition model and training dataset; The data in the training dataset is input into the initial impact recognition model, and the initial impact recognition model outputs the predicted rear overhang impact pulse data. With the goal of minimizing the deviation between the predicted impact trigger time and the actual impact trigger time in the predicted rear overhang impact pulse data, a timing accuracy loss function is constructed; Based on the predicted rear overhang impact pulse data, construct the false trigger penalty loss function and the missed detection penalty loss function; Based on the time-series precision loss function, the false triggering penalty loss function, and the missed detection penalty loss function, an event-driven joint loss function is constructed by weighted summation; The initial impact recognition model is trained by minimizing the event-driven joint loss function to obtain the impact recognition model.

[0014] Based on the same inventive concept, a second aspect of this application also provides a vehicle, comprising: a memory for storing executable program code; and a processor for calling and running the executable program code from the memory, causing the vehicle to perform the method as described in the first aspect.

[0015] As can be seen from the above, the vehicle control method and vehicle provided in this application include: acquiring multi-source perception data during vehicle operation; preprocessing the multi-source perception data to obtain temporal perception data characterizing the vehicle's obstacle-crossing process. The temporal perception data is data generated during the vehicle's obstacle-crossing process, specifically data from the start of obstacle crossing by the front wheels to the point when the rear wheels are about to cross the obstacle. Based on the temporal perception data, an obstacle-crossing structure map is generated using a pre-trained map generation model. The obstacle-crossing structure map abstracts the high-dimensional, chaotic multi-source perception data into a structure map that conforms to vehicle dynamics and has physical interpretability, providing structured prior knowledge for subsequent models. Based on the obstacle-crossing structure map, a pre-trained prediction coding model generates a prediction result and prediction error data for the vehicle's rear suspension state. The rear suspension state prediction result can predict the rear suspension state at several future time steps, and combined with the prediction error data, accurate identification before rear suspension impact is achieved, realizing the goal of forward-looking prediction and solving the problem of lag in existing chassis control systems. Based on the predicted rear suspension state and the prediction error data, rear suspension impact pulse data is generated using a pre-trained impact recognition model. The rear suspension impact pulse data is then used to convert the predicted rear suspension state into a clear rear suspension hardening trigger signal, including the precise trigger time and hardening intensity. Based on the obstacle structure map, the predicted rear suspension state, the prediction error data, and the rear suspension impact pulse data, a collaborative decision is made to generate rear suspension control commands. These commands integrate the output information from the three aforementioned models, improving their reliability and accuracy. The rear suspension control commands are then sent to the vehicle's suspension control system, which executes them, overcoming the response lag bottleneck of existing suspension control systems. This achieves accurate prediction and early hardening control of rear suspension obstacle crossing impacts, effectively mitigating the risk of chassis impacts in unstructured terrain. Simultaneously, it improves the vehicle's passability and comfort on terrains such as cliffs and steps. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic flowchart of a 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 3This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0019] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0020] Current vehicle chassis control systems, when dealing with unstructured and complex terrains such as off-road vehicles, construction sites, and unpaved roads, which include features like cliffs, steep slopes, steps, and potholes, mostly rely on the front suspension's obstacle-crossing status (such as front wheel compression, vehicle pitch changes, and sudden changes in wheel acceleration) as the core trigger condition. This drives dynamic response strategies such as vehicle attitude adjustment, suspension stiffness adjustment, or damping control. However, chassis control systems generally lack a precise identification and advanced prediction mechanism for the critical timing difference between "front suspension obstacle crossing completion and rear suspension impending impact with the ground." This prevents the establishment of a dynamic causal relationship between front and rear axle obstacle-crossing events, leading to safety and comfort issues such as chassis scraping, suspension bottoming out, severe vehicle pitch, and excessive impact vibration even when the rear wheels pass over an obstacle. The front suspension refers to the vehicle's front suspension system, and the rear suspension refers to the vehicle's rear suspension system.

[0021] Existing vehicle chassis control architectures often rely on fixed threshold judgments or simplified dynamic models, making it difficult to analyze the complex coupling relationships between vehicle structure, terrain features, attitude changes, and load transfer during obstacle crossing in real time. This is especially true under conditions such as high-speed obstacle crossing, terrain with large elevation differences, and strong nonlinear impacts, where there are generally problems with perception lag, decision lag, and execution lag. Because the system cannot identify the precursors of rear suspension impacts in advance, it is even more difficult to actively activate the rear suspension hardening control within the critical time window before the impact occurs. It can only passively adjust after the impact, resulting in a significant reduction in control effectiveness.

[0022] This is because traditional vehicle chassis control systems lack the structural perception and dynamic modeling capabilities to handle the entire obstacle-crossing process. They cannot analyze the transmission pattern of front-wheel obstacle crossing to rear-wheel impact at the structural level, making it difficult to anticipate and respond quickly to rear suspension impacts. This prevents chassis control strategies from achieving synchronized matching and coordinated control of front and rear axle obstacle-crossing events, ultimately resulting in vehicle passability, driving stability, and ride comfort failing to meet the high-end, intelligent requirements of complex unstructured terrain.

[0023] In view of this, this application proposes a vehicle control method that, through the fusion of multiple models for collaborative decision-making, overcomes the response lag bottleneck of existing suspension control systems, achieving accurate prediction and early hardening control of rear suspension impacts during obstacle crossings, effectively avoiding chassis collision risks in unstructured terrain. Simultaneously, it improves the vehicle's passability and comfort on terrains such as cliffs and steps.

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

[0025] This application provides a vehicle control method, referring to... Figure 1 This includes the following steps: Step 102: Acquire multi-source perception data during vehicle operation, and preprocess the multi-source perception data to obtain temporal perception data characterizing the vehicle's obstacle-crossing process.

[0026] Specifically, multi-source perception data includes raw data collected in real time by various vehicle sensors, encompassing three dimensions: vehicle motion state, spatial environment, and vehicle control. This data comprehensively captures the dynamic changes of the vehicle during obstacle crossing. Multi-source perception data includes at least wheel acceleration sensor data, suspension displacement sensor data, inertial measurement unit (IMU) data, camera data, millimeter-wave radar data, and local area network (LAN) bus signals. Wheel acceleration reflects the impact acceleration changes when the wheels contact obstacles, while suspension displacement records the changes in suspension travel of the front and rear suspensions, reflecting the suspension compression / extension state. The IMU captures core motion parameters such as the vehicle's attitude, angular velocity, and acceleration. Cameras can perceive environmental information such as road obstacles and terrain contours from a visual perspective. Millimeter-wave radar acquires spatial information such as the distance and relative speed between the vehicle and obstacles. LAN bus signals include core data from the vehicle control layer, including vehicle operating parameters such as wheel speed, motor torque, and steering angle.

[0027] Multi-source sensing data suffers from issues such as inconsistent acquisition frequencies, inconsistent coordinate systems, noise interference, and systematic errors. This step addresses these issues through preprocessing to achieve spatiotemporal consistency and improve data accuracy. Because different sensors have varying acquisition frequencies (e.g., high-frequency acquisition by the inertial measurement unit and low-frequency acquisition by cameras / radar), direct use would lead to timing discrepancies. Therefore, a unified timestamp mechanism is needed for data synchronization. Specifically, the acquisition clock of the inertial measurement unit is used as a high-frequency reference clock. Data from all sensors at different frequencies is aligned to this clock reference, constructing a master timeline centered on the inertial measurement unit. This ensures that all data remains consistent in the time dimension, laying the foundation for subsequent spatiotemporal joint analysis. The raw data from cameras and millimeter-wave radar are based on their own device coordinate systems, which are incompatible with the coordinate systems of vehicle body and suspension sensor data, making direct structured analysis impossible. During the preprocessing process, spatial calibration is completed through extrinsic parameter matrix transformation: the environmental perception data from cameras and millimeter-wave radar are uniformly transformed into the vehicle coordinate system, allowing data from different modalities such as vision, radar, and vehicle sensors to be correlated and analyzed under the same spatial reference system, eliminating the information fragmentation problem caused by differences in spatial coordinate systems.

[0028] Raw data from the inertial measurement unit (IMU) and suspension displacement sensors are susceptible to high-frequency noise from non-road obstacle-crossing factors such as vehicle vibration and minor road bumps. Retaining this noise would interfere with subsequent model learning of core obstacle-crossing features. The preprocessing stage employs a combination of Kalman filtering and median filtering to specifically denoise the IMU and suspension data, filtering out invalid high-frequency noise and retaining core data that accurately reflects the vehicle's obstacle-crossing dynamics. Furthermore, vehicle sensors, under long-term use or different operating conditions, may exhibit systematic errors such as bias and drift (e.g., non-zero output values ​​when the sensor is stationary, or data slowly shifting over time), leading to decreased data accuracy. The preprocessing stage uses known static conditions (e.g., the vehicle's attitude and wheel load when parked on a flat road) as a benchmark to perform bias correction and drift compensation on the outputs of all sensors, eliminating systematic errors at the source and ensuring absolute data accuracy.

[0029] Through the above preprocessing, the chaotic, heterogeneous, and noisy multi-source sensing data is transformed into standardized sensing data that is spatiotemporally unified, high-precision, and reproducible, providing high-quality data for subsequent models. Sensing data from various moments within a given time period are merged into temporal sensing data. When the vehicle determines that the temporal sensing data falls within the period from the front wheels contacting the obstacle to the rear overhang about to touch the ground, the resulting temporal sensing data is input into the subsequent map generation model. If it is determined that the temporal sensing data does not fall within the period from the front wheels contacting the obstacle to the rear overhang about to touch the ground, the vehicle control method of this application is not executed, avoiding the consumption of vehicle-side computing resources.

[0030] When training subsequent models, a time-segment dataset is constructed based on the time-series-aware data to serve as training data. Using the key moments of the obstacle-crossing process—the first obstacle clearing by the front wheel and the moment the rear wheel is about to clear an obstacle—as time anchors, the continuous time-series-aware data is segmented into independent obstacle-crossing event segments, ultimately constructing an event segment dataset. This dataset provides accurate sample annotations for the supervised training of subsequent models, enabling them to specifically learn the temporal correlation features of "front overhang obstacle clearing - rear overhang impending impact." In addition to preprocessed data, the event segment dataset includes annotations of key obstacle-crossing time nodes to clearly define the start and end range and key time-series nodes of each event segment. The key time points for obstacle crossing include the moment the front wheels first touch the obstacle (T0, the instant the front tires make rigid contact with the obstacle), the moment the vehicle begins to pitch (T1, the instant the impact force from the front suspension is transferred to the vehicle body, and the vehicle pitch angular velocity increases significantly from the baseline value), the moment the front suspension takes off (T2, the instant the front suspension transitions from maximum compression to extension, and the front wheels begin to detach from the obstacle), and the moment the rear suspension touches the obstacle (T3, the instant the rear tires make contact with the ground / obstacle). For example, starting at T0 and ending 50ms after T3 (when impact energy absorption is complete), each segment ensures that it fully covers the entire core process of "front wheel obstacle contact → front suspension compression → front suspension take-off → rear suspension awaiting impact → rear suspension impact." The event segment dataset also includes dedicated supervision labels for model training, used for supervised model training.

[0031] Step 104: Based on the time-series sensing data, generate an obstacle-crossing structure map using a pre-trained map generation model.

[0032] Specifically, based on time-series perceived data, an initial structure graph is generated by constructing feature vectors for nodes and connecting edges between nodes. Each node represents a component of the vehicle, and the time-series perceived data is mapped to a high-dimensional feature vector for each node according to each component. The dimension of the high-dimensional feature vector of each node is fixed, and can be 256 dimensions. Learnable connecting edges are established for each node according to the vehicle's physical coupling rules and dynamic temporal dependencies, and edge weights are initialized. The final output is an initial structure graph containing nodes, edges, and initial edge weights. The initial structure graph is iteratively propagated through a graph neural network to update the feature vectors of nodes and edge weights, generating the final obstacle-crossing structure graph.

[0033] Step 106: Based on the obstacle crossing structure map, generate the vehicle's rear suspension state prediction result and prediction error data through the pre-trained prediction coding model.

[0034] Specifically, the predictive coding model includes an encoder, a prediction network, and an error network. The encoder is responsible for feature encoding and compression of the obstacle crossing structure map, transforming it into a dynamic semantic hidden state vector. Based on the hidden state vector, the prediction network predicts the rear overhang state at the next T time steps, enabling the predictive network to make advance predictions. The error network calculates the error between the candidate state prediction results and the real-time sensor observations of the vehicle, generating prediction error data. During the training process of the predictive coding model, the prediction error data can be backpropagated within the model. Through the backpropagation mechanism, the network weights of the encoder and prediction network are dynamically adjusted, allowing the predictive coding model to continuously refine its internal representation of the obstacle crossing dynamic process.

[0035] Step 108: Based on the rear suspension state prediction results and the prediction error data, generate rear suspension impact pulse data using the pre-trained impact recognition model.

[0036] Specifically, the impact recognition model in this embodiment is a spiking neural network. The spiking neural network monitors impact risk in real time through a membrane potential integration mechanism. Once the accumulated risk reaches a threshold, a hard trigger signal is immediately generated. First, the rear suspension state prediction results and prediction error data are converted into pulse signals that the spiking neural network can process. The spiking neural network adopts an event-driven mode; only when the change in input data exceeds a preset threshold will the neurons be triggered to calculate and generate a hard trigger signal, saving computational power. The inference process of the spiking neural network is a biomimetic process of "continuous signal → pulse encoding → spatiotemporal integration → threshold triggering." During critical obstacle crossing periods, this process can operate at a microsecond-level granularity, greatly improving the vehicle's response speed. The rear suspension impact pulse data includes the rear suspension hardening trigger signal, hardening degree, and timestamp.

[0037] Step 110: Generate a rear suspension control command based on the obstacle structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data; send the rear suspension control command to the vehicle's suspension control system so that the rear suspension control command can be executed by the suspension control system.

[0038] Specifically, in this step, the outputs of the map generation model, predictive coding model, and impact recognition model are fused to achieve collaborative decision-making and rear suspension control command generation. By fusing the output data of the Sanger model, false triggering caused by sensor interference and feature bias in a single model can be avoided, improving the accuracy and reliability of the rear suspension control commands. Afterward, the rear suspension control commands are sent to the suspension control system for execution.

[0039] Based on steps 102 to 110 above, this embodiment provides a vehicle control method, including acquiring multi-source perception data during vehicle operation, and preprocessing the multi-source perception data to obtain temporal perception data characterizing the vehicle's obstacle-crossing process. The temporal perception data is generated during the vehicle's obstacle-crossing process, specifically the data from the start of obstacle crossing by the front wheels to the point when the rear wheels are about to cross the obstacle. Based on the temporal perception data, an obstacle-crossing structure map is generated using a pre-trained map generation model. The obstacle-crossing structure map abstracts the high-dimensional, chaotic multi-source perception data into a structure map that conforms to vehicle dynamics and has physical interpretability, providing structured prior knowledge for subsequent models. Based on the obstacle-crossing structure map, a pre-trained predictive coding model generates a prediction result and prediction error data for the vehicle's rear suspension state. The rear suspension state prediction result can predict the rear suspension state at several future time steps, and combined with the prediction error data, accurate identification before rear suspension impact is achieved, realizing the goal of forward-looking prediction and solving the problem of lag in existing chassis control systems. Based on the predicted rear suspension state and the prediction error data, rear suspension impact pulse data is generated using a pre-trained impact recognition model. The rear suspension impact pulse data is then used to convert the predicted rear suspension state into a clear rear suspension hardening trigger signal, including the precise trigger time and hardening intensity. Based on the obstacle structure map, the predicted rear suspension state, the prediction error data, and the rear suspension impact pulse data, a collaborative decision is made to generate rear suspension control commands. These commands integrate the output information from the three aforementioned models, improving their reliability and accuracy. The rear suspension control commands are then sent to the vehicle's suspension control system, which executes them, overcoming the response lag bottleneck of existing suspension control systems. This achieves accurate prediction and early hardening control of rear suspension obstacle crossing impacts, effectively mitigating the risk of chassis impacts in unstructured terrain. Simultaneously, it improves the vehicle's passability and comfort on terrains such as cliffs and steps.

[0040] In some embodiments, generating an obstacle-crossing structure map based on the time-series-aware data using a pre-trained map generation model includes: The temporal sensing data is encoded into feature vectors of each node in the obstacle crossing structure graph by the temporal encoder in the graph generation model; wherein each node represents a component of the vehicle; based on the feature vectors of each node and prior knowledge of vehicle physics, connection edges are established between each node to generate an initial structure graph; the initial structure graph is input into the graph neural network in the graph generation model for iterative propagation to generate the obstacle crossing structure graph.

[0041] Specifically, a timing encoder can convert continuous time-series sensing data into high-dimensional feature vectors for each vehicle component node in a graph structure. Based on the vehicle's physical structure, the input time-series sensing data is split into independent component data channels. For example, front wheel acceleration, front suspension travel, and front suspension speed are assigned as inputs to the front suspension node, while pitch angle, yaw angle, and longitudinal acceleration from the vehicle body inertial measurement unit are assigned as inputs to the body node. The timing encoder processes the data from the most recent N time steps (e.g., 50ms for each component), and the data corresponding to each vehicle component is ultimately compressed into a fixed-dimensional node feature vector. For example, M node feature vectors are generated, where M represents the preset number of vehicle components, such as front wheels, front suspension, body, rear suspension, rear wheels, and road environment.

[0042] Based on the feature vectors of each node and prior knowledge of vehicle physics, connection edges are established between nodes to form an initial structural graph. According to the vehicle's mechanical structure, force transmission paths are formed, generating connection edges between nodes, such as front wheel → front suspension → body → rear suspension → rear wheel. Connection edges are established between nodes by calculating the real-time correlation between the feature vectors of each node. For example, when the impact feature vector of the "front wheel node" shows a strong correlation with the height difference feature of the "road environment node," a connection edge is established between the "front wheel node" and the "road environment node." When the vehicle body pitches violently, the connection strength between the "body node" and the "rear suspension node" is enhanced. Finally, an M×M binary adjacency matrix is ​​generated, where 1 indicates the existence of a connection edge between two nodes, and 0 indicates no connection. After determining the feature vectors of each node and the connection edges between them, the initial structural graph is formed.

[0043] The initial structure map is input into the graph neural network in the map generation model for iterative propagation, updating node feature vectors and edge weights so that the generated obstacle-crossing structure map can reflect the dynamic coupling system of various vehicle components. The graph neural network can be a graph convolutional network or a graph attention network. First, the edge weights are initialized. Based on the adjacency matrix, each connecting edge is assigned an initial weight value (between 0 and 1) using the similarity of the feature vectors of each node. The weight represents the coupling strength between the two components at that moment. For example, when the front suspension just touches the obstacle, the edge weight of "front suspension-body" is higher (force is being strongly transmitted), while the edge weight of "body-rear suspension" is relatively lower (the rear suspension has not yet been affected). Each node collects the features of its neighboring nodes, multiplies them by the corresponding edge weights, and then aggregates them to update its own feature vector. For example, the features of the body node will incorporate the impact information from the front suspension, and the features of the rear suspension node will "sense" in advance the pitch attitude change that the body will undergo. During the propagation process, vehicle dynamics regularization terms are embedded. For example, the "suspension travel" feature is forced to not exceed physical limits, and the "force transmission" direction is forced to conform to the laws of mechanics to prevent the graph neural network from outputting results that violate physical common sense. After multiple rounds of iterative propagation, the calculation stops, the node feature vectors and edge weights at the current moment are solidified, and an obstacle-crossing structure graph is generated.

[0044] Furthermore, since obstacle crossing is a continuous process, the current obstacle crossing structure map needs to be correlated with historical obstacle crossing structure maps. The current obstacle crossing structure map is added to a sliding time window buffer (e.g., buffering the map of the most recent 10 frames) to provide a data foundation for the vehicle state evolution of subsequent models.

[0045] During the training of the graph generation model, the event fragment dataset constructed in the aforementioned embodiments is used as the training dataset. The event fragment dataset includes time-series-aware data, obstacle crossing stage labels, ground truth graph structure values, and physical coupling strength annotations. Simultaneously, physical constraint reference values ​​(such as suspension travel limits and force transmission efficiency thresholds) output by the vehicle dynamics simulation model are introduced as auxiliary data sources during training. The training dataset covers obstacle crossing data with different combinations of obstacle height, impact angle, and vehicle speed. The joint loss function during training can include a graph construction loss function, a stage determination loss function, and a physical constraint regularization term. The graph construction loss function can use a weighted mean square error loss to calculate the deviation between the generated graph structure and the ground truth graph structure. The stage determination loss function can use a cross-entropy loss function to constrain the consistency between the predicted probability distribution of obstacle crossing stages and the obstacle crossing stage labels. The physical constraint regularization term uses a penalty loss; if node features (such as suspension travel) or edge weights (such as force transmission strength) in the graph exceed physical limits, a tiered penalty is applied based on the magnitude of the exceedance (the greater the exceedance, the heavier the penalty). Once the model converges, the trained map generation model is obtained.

[0046] By constructing the obstacle crossing structure map in this embodiment, the "black box problem" of traditional time series models is solved. By structuring the time series sensing data, the downstream predictive coding model can use the dynamic force relationship map that conforms to physical laws when making predictions, so that the prediction of rear suspension impact has interpretability and high robustness.

[0047] In some embodiments, generating vehicle rear suspension state prediction results and prediction error data based on the obstacle-crossing structure map and a pre-trained prediction coding model includes: Temporal data of the target node is extracted from the obstacle crossing structure map; features of the temporal data of the target node are extracted by the encoder in the predictive coding model to obtain a hidden state vector; based on the hidden state vector, multi-step forward state prediction is performed by the prediction network in the predictive coding model to generate the rear overhang state prediction result; based on the rear overhang state prediction result and real-time sensor data, error calculation is performed by the error network in the predictive coding model to generate the prediction error data.

[0048] Specifically, target nodes refer to key nodes, such as the front suspension, vehicle body, rear suspension, front tires, and road surface environment. The temporal data of target nodes is extracted from the obstacle crossing structure map sequence, including the feature vectors of the nodes at various time points and the dynamic edge weight matrix between nodes. This temporal data characterizes the state evolution trajectory of the target nodes. The data input to the predictive coding model can also include physical constraint information, such as suspension travel limits and vehicle pitch angle ranges, providing physical boundaries for the predictive coding model and avoiding invalid predictions. The final target node temporal data is a sequence of node state evolutions directly related to the rear suspension impact prediction, and undergoes lightweight processing such as normalization and dimensionless elimination to form a standard input feature stream that the predictive coding model can directly receive.

[0049] The predictive coding model comprises an encoder, a prediction network, and an error network. The encoder extracts features from the temporal data of the target nodes, compressing high-dimensional, continuous temporal features into low-dimensional, semantically unified latent state vectors. These latent state vectors represent the dynamic features of the vehicle at its current obstacle-crossing stage, preserving both the core dynamics of front suspension obstacle crossing and the transmission trend of vehicle posture. The encoder structure can be a one-dimensional convolutional neural network plus a lightweight gated recurrent unit. The one-dimensional convolutional neural network extracts local features, such as sudden increases and decreases in front suspension travel, peak values ​​of vehicle pitch angle acceleration, and instantaneous changes in tire impact signals. These features are core local representations of the impending impact during obstacle crossing. The lightweight gated recurrent unit uses gating mechanisms (update gate, reset gate) to filter and model long-term temporal dependencies, such as the continuous evolution trend of vehicle pitch angle after front suspension compression and the temporal correlation between front suspension take-off and vehicle head-up. Invalid historical features are eliminated, while valuable temporal patterns for rear suspension prediction are retained. Finally, a hidden state vector with a unified dimension is obtained. This vector condenses all the key obstacle-crossing dynamic features from the current time step forward N time steps and is the core input data of the prediction network.

[0050] Based on the hidden state vector, a multi-step forward prediction of the state is performed through the prediction network in the predictive coding model to generate the rear suspension state prediction result. The network structure of the prediction network may include a fully connected prediction layer with residual connections + a lightweight temporal decoder. Through a fully connected layer with residual connections, the low-dimensional hidden state vector of the encoder is mapped to the feature space adapted to the prediction network, supplementing the feature dimension while avoiding gradient vanishing and ensuring prediction accuracy. Using a lightweight temporal decoder, based on the mapped features, the multi-dimensional core states of the rear suspension are predicted simultaneously for the next T time steps, generating the rear suspension state prediction result (dimension: future time step T × rear suspension feature dimension K). The rear suspension state prediction result may include the rear suspension displacement and displacement change rate, rear suspension compression / extension acceleration, future evolution values ​​of vehicle pitch angle / angular velocity, impact probability of the rear suspension tires and the ground (between 0 and 1), and predicted value of the rear suspension travel. The rear suspension state prediction result provides trend features of the future state of the rear suspension for the impact recognition model and is the "basic judgment basis" for impact recognition. The prediction process incorporates hard constraints on vehicle dynamics, and real-time corrections are made to predicted values ​​that exceed physical limits (e.g., if the predicted value of the rear suspension displacement exceeds the maximum travel, it is forcibly corrected to the upper limit of the threshold), ensuring that the prediction results conform to the laws of vehicle mechanical structure and avoiding invalid predictions.

[0051] Based on the rear suspension state prediction results and real-time sensor data, error calculation is performed through the error network in the predictive coding model to generate prediction error data. In practice, real-time sensor observations of the vehicle's rear suspension and body are collected synchronously with the prediction inference, including rear suspension displacement, acceleration, and body pitch angle, ensuring spatiotemporal consistency with the rear suspension state prediction results. For each rear suspension feature at each future time step, a weighted mean square error is used to calculate the single-time-step error vector. The error vectors of the next T time steps are arranged chronologically to form the original error sequence. A lightweight first-order low-pass filter is applied to the original error sequence to remove invalid error abrupt changes caused by instantaneous sensor glitches. Simultaneously, gradient enhancement processing is applied to error abrupt change features (such as error values ​​suddenly increasing above a threshold in a short time) to highlight the error signal changes corresponding to the impending impact, ultimately generating prediction error data. Based on the mean and variance of the overall error, a credibility score (between 0 and 1) is assigned to the multi-step prediction results; the smaller the error, the higher the score. This score is incorporated into the output as an auxiliary feature to provide a reference for subsequent impact recognition models and fusion decisions. Prediction error data provides the impact identification model with dynamic deviation characteristics of the rear overhang state from the normal trend, and sudden error changes are "strong signals" of impending impact. The credibility score provides a credibility reference for the prediction results of the impact identification model, which assigns higher weight to predictions with high credibility.

[0052] During the training of the predictive coding model, the training data includes obstacle-crossing structure maps output by the map generation model, as well as sensor observation data of the rear suspension and body during actual vehicle obstacle crossing. The actual observed values ​​of the rear suspension state at the next T time steps corresponding to the time-series data of the labeled target nodes, including suspension displacement / rate of change, compression acceleration, body pitch angle, and rear suspension impact probability, are used as prediction labels for the rear suspension state. The deviation between the initial prediction values ​​of the predictive coding model and the actual observation values ​​of the actual vehicle is calculated and used as the true label for the error signal. Humans rate the data based on its validity and completeness, labeling the prediction credibility score as an auxiliary label to supervise the output credibility score of the predictive coding model. During training, the primary objective is to minimize the multi-step prediction errors of rear suspension displacement, body pitch angle, and impact intensity, ensuring a high degree of consistency between the predicted rear suspension state and the measured values. Simultaneously, several constraint objectives are defined, including: suppressing abrupt changes in prediction between adjacent time steps to maintain temporal continuity; introducing edge weights of the obstacle-crossing structure map as regularization terms to maintain physical coupling stability; and limiting high-frequency oscillations in the prediction error data to improve the robustness of impact precursor identification. A joint loss function is constructed based on the main objective and multiple constraint objectives, in the following form: (1), (2), in, This represents the multi-step state prediction loss function. Represents the time consistency loss function. Represents the structural loss function, representing Smoothing loss function This represents the weight of each loss function. T represents the total number of time steps, represents the predicted hang-over state at time step t, and represents the predicted hang-over state label at time step t. Through Consistency between predicted and measured values ​​of rear suspension displacement, vehicle pitch angle, and impact strength. Suppress abrupt changes in prediction between adjacent time points to ensure a smooth and continuous predicted trajectory, avoiding jumps. Through... By introducing edge weights from the obstacle-crossing structure graph as regularization terms, the physical coupling relationship is kept stable, ensuring that the prediction conforms to the laws of vehicle dynamics. To limit high-frequency oscillations in prediction error data and improve the robustness of shock precursor identification, the predictive coding model is trained by minimizing the joint loss function. After convergence, the trained predictive coding model is obtained.

[0053] In this embodiment, the predictive coding model acts as a bridge between the map generation model and the impact recognition model, playing a crucial role in the conversion of "structured features → temporal predictive features" throughout the entire rear overhang obstacle crossing impact prediction and active hardening control chain. By performing multiple forward predictions and anticipating future rear overhang states in advance, it provides rear overhang hardening control with more time to respond, solving the problem of lag in traditional systems. It also provides the impact recognition model with rear overhang state prediction results and prediction error data, enabling the impact recognition model to not only identify the predicted rear overhang state but also capture dynamic abrupt changes in state deviation, significantly improving the robustness and accuracy of impact precursor identification.

[0054] In some embodiments, generating rear suspension impact pulse data based on the rear suspension state prediction result and the prediction error data using a pre-trained impact recognition model includes: The pulse coding layer in the impact recognition model converts the rear suspension state prediction result and the prediction error data into pulse sequence data; the synaptic weight modulation layer in the impact recognition model performs weight modulation on the pulse sequence data to form weighted pulse sequence data; based on the weighted pulse sequence data, the decision output layer in the impact recognition model outputs the rear suspension impact pulse data using a membrane potential integration mechanism.

[0055] Specifically, the impact recognition model in this embodiment is a spiking neural network. The spiking neural network includes a perceptual input layer, a pulse coding layer, a salient weight modulation layer, and a decision output layer. The data input to the spiking neural network includes the overhang state prediction result, prediction error data, and prediction confidence score. The perceptual input layer is responsible for receiving, normalizing, and distributing the continuous overhang state prediction results and prediction error data output by the prediction coding model, converting them into analog input currents that neurons can understand. In practice, the overhang state prediction result, prediction error data, and prediction confidence score are mapped one-to-one to the corresponding input neurons, and the values ​​are normalized to the 0-1 range to ensure that the intensity of the input current is within the dynamic range of the neuron. The normalized feature values ​​are directly used as presynaptic currents and passed to the next layer, the pulse coding layer.

[0056] The pulse coding layer transforms the continuous analog current from the sensing input layer into a discrete pulse sequence using specific coding rules. In practice, a hybrid coding strategy is employed to address different characteristics of the obstacle crossing scenario. For the rear-hanging state prediction result, frequency coding is used: the larger the input current (the higher the impact probability represented by the rear-hanging state prediction result), the faster the neurons fire pulses, thus transforming a sustained high impact risk into a dense burst of pulses. For the prediction error data, time coding is used: the steeper the slope of the input current change (a sudden increase in error), the earlier the neurons fire pulses. The output of the pulse coding layer is no longer a continuous numerical value, but an asynchronous pulse event. The pulse coding layer can convert the rear-hanging state prediction result into "neural language." If the rear-hanging impact probability is predicted to spike from 0.2 to 0.8 based on the rear-hanging state prediction result, the pulse coding layer will immediately fire a strong pulse within milliseconds.

[0057] The prominent weight modulation layer, acting as a gating and weighting layer in the spiking neural network, can adjust the weights of synaptic connections in real time, determining which pulse signals should be amplified and which should be suppressed, thereby improving the robustness of the decision. The prominent weight modulation layer first loads the base weight matrix fixed during training, which represents the inherent importance of different features to impact prediction (e.g., the weight of "impact probability" is higher than that of "vehicle pitch angle"). It then scales the base weights in real time based on the received confidence score. If the confidence score is high, the input is trusted and the full weights are used; if the confidence score is low, the weights are reduced to prevent false triggers. Effective weights are generated through this adjustment of the base weights. The pulse sequence data from the pulse coding layer is multiplied by the corresponding effective weights, transforming it into weighted postsynaptic potentials (i.e., weighted pulse sequence data), which are then passed to the decision output layer.

[0058] The decision output layer generates the final rear suspension impact pulse data through membrane potential integration. This data includes the rear suspension hardening trigger signal, the degree of hardening, and a timestamp. The decision output layer accumulates membrane potentials using a leakage current integration ignition neuron model and ultimately determines whether to generate a rear suspension hardening trigger signal based on a threshold. Each neuron maintains a membrane potential; it rises upon receiving weighted pulse data and decays with a time constant leakage current if no pulse input is received. When the membrane potential exceeds a preset trigger threshold, an impending impact is determined. The decision output layer immediately outputs a high-level pulse, the rear suspension hardening trigger command. Afterward, the membrane potential is instantly reset to its resting potential and enters a very short refractory period. Neurons include trigger neurons and intensity neurons. The trigger neuron outputs whether a trigger has occurred, while the intensity neuron outputs the degree of hardening (e.g., stiffness / damping) based on the number of pulses fired per unit time. Simultaneously, the decision output layer can also output timing information, specifically the precise timestamp of the pulse firing, synchronized with the inertial measurement unit's master clock for time consistency verification with the map generation model and predictive coding model. The impact recognition model ensures that, in unstructured terrain, the vehicle chassis can accurately and promptly harden the rear suspension with a biological-like response speed, ultimately avoiding rear suspension impacts and maintaining vehicle balance.

[0059] In some embodiments, the training process of the impact recognition model includes: An initial impact recognition model and a training dataset are constructed. Data from the training dataset is input into the initial impact recognition model, which outputs predicted rear-end impact pulse data. A temporal accuracy loss function is constructed with the objective of minimizing the deviation between the predicted impact trigger time and the actual impact trigger time in the predicted rear-end impact pulse data. A false trigger penalty loss function and a missed detection penalty loss function are constructed based on the predicted rear-end impact pulse data. An event-driven joint loss function is constructed by weighted summation based on the temporal accuracy loss function, the false trigger penalty loss function, and the missed detection penalty loss function. The initial impact recognition model is trained by minimizing the event-driven joint loss function to obtain the impact recognition model.

[0060] Specifically, the initial impact recognition model and training dataset are constructed first. The model structure of the initial impact recognition model is the same as that of the impact recognition model in the aforementioned embodiment. The training dataset includes the predicted rear-hanging state prediction results and prediction error data output by the predictive coding model, as well as prediction confidence scores. The label data in the training dataset includes the actual rear-hanging impact time and binary classification labels for impact / non-impact. A value of 1 indicates a true impact; the binary label value is... The time frame represents non-impact events, used for monitoring false triggers / missed detections. An event-driven joint loss function is used during training. The specific format is as follows: (3), in, This represents the time precision loss function, used to constrain the prediction of the impact trigger time. Compared to the actual impact trigger time To minimize deviations and ensure precise triggering timing. This represents the false trigger penalty loss function, which is used to penalize false trigger pulses in non-impact scenarios and reduce false triggers. This represents the missed detection penalty loss function, which is used to penalize missed pulses in real impact scenarios to avoid no response in critical scenarios. , These are the weighting coefficients. This represents the probability of a predicted impact in a non-impact scenario. This indicates that in a scenario where there should be no impact, the impact recognition model incorrectly generated a trigger signal, i.e., a false trigger occurred. The purpose is to penalize only false triggers, without incurring loss for correct, impact-free predictions. The initial impact recognition model is trained by minimizing the event-driven joint loss function, resulting in a fully trained impact recognition model. Through event-driven supervised learning in the training process of this embodiment, the impact recognition model can accurately identify rear suspension impact precursors and generate millisecond-level trigger signals, while strictly controlling false triggers and missed detection rates, providing a reliable core criterion for rear suspension hardening control.

[0061] In some embodiments, generating rear suspension control commands based on the obstacle structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data includes: Based on the obstacle crossing structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data, a consistency check is performed; in response to passing the consistency check, a rear suspension control command is generated based on the rear suspension impact pulse data, so that the vehicle's suspension control system executes the rear suspension control command.

[0062] Specifically, in this embodiment, the outputs of the obstacle crossing structure map, the predictive coding model, and the impact recognition model are fused to achieve collaborative decision-making and rear suspension control command generation. The input data for the fusion process consists of the output data from the three models, including obstacle crossing structure maps, rear suspension state prediction results, prediction error data, and rear suspension impact pulse data. The obstacle crossing structure map includes obstacle crossing stage judgment values; the rear suspension state prediction results include confidence scores and predicted risk levels; and the rear suspension impact pulse data includes rear suspension hardening trigger signals, hardening degree values, and timestamps. All input data are strictly aligned with the vehicle's inertial measurement unit's main timeline to ensure spatiotemporal consistency, laying the data foundation for collaborative decision-making.

[0063] The core design of the fusion mechanism follows the principles of "division of labor and cooperation, mutual verification and constraint, and fault tolerance and correction." Based on the functional characteristics of the three models, it constructs three independent reasoning paths: a structural semantic flow (graph generation model), a dynamic prediction flow (predictive coding model), and an event-driven flow (impact recognition model). Each path makes independent decisions based on its own input, and the fusion mechanism then achieves information exchange and consistency verification. The structural semantic flow defines the "stage" (which stage the obstacle crossing is in and whether hardening can be performed), the dynamic prediction flow defines the "risk" (how great is the rear suspension impact risk and how much hardening is needed), and the event-driven flow defines the "timing" (when to perform hardening, the core trigger point). These three aspects respectively cover the three core dimensions of decision-making: spatial boundaries, risk level, and temporal sequence, with no functional overlap. The highest priority rear suspension hardening operation is triggered only when the decision results of the three flows reach consistency, avoiding false triggering caused by sensor interference or feature bias in a single model.

[0064] Furthermore, the consistency verification based on the obstacle-crossing structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data includes: Based on the obstacle crossing structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data, a time consistency check is performed; based on the obstacle crossing structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data, a result consistency check is performed; in response to passing both the time consistency check and the result consistency check, it is determined that the consistency check has been passed.

[0065] Specifically, for the structural semantic flow, "rear suspension hardening can be triggered" only when the obstacle-crossing phase enters the core window period of "front suspension jumping to rear suspension about to touch the ground," and the relative height between the rear suspension and the ground is lower than a preset threshold, and the vehicle body-rear suspension coupling weight exceeds the activation threshold; otherwise, it is determined as "not triggerable." Simultaneously, the spatiotemporal boundaries of hardening execution (e.g., earliest triggerable time, latest trigger cutoff time) are output to define the execution range of subsequent decisions. For the dynamic prediction flow, based on the impact probability in the rear suspension state prediction results, four levels of risk (low / medium / high / extremely high) are classified, matching preset four-level hardening degree intervals (e.g., low risk: stiffness / damping increased by 20%, extremely high risk: stiffness / damping increased by 100%). Simultaneously, the hardening degree is weighted and adjusted based on the prediction credibility score (higher credibility means a harder hardening degree more closely matches the risk level; lower credibility means a lower hardening degree is appropriately reduced to avoid over-control). Finally, the suggested trigger time and suggested hardening degree are output. For event-driven flows, if the hardening trigger signal is 1 (triggered), the core trigger time is the pulse release timestamp, and the hardening degree value is the final hardening degree; if the hardening trigger signal is 0, it is determined that "not triggering for now".

[0066] The consistency verification includes time consistency verification and result consistency verification. For time consistency verification, it checks whether the trigger times of the three stream outputs fall within a preset time window (e.g., 5ms). If the time difference exceeds the window, the time is considered inconsistent and requires subsequent correction; if it is within the window, the time is considered consistent, and the time consistency verification passes. For result consistency verification, it includes two layers of verification: ① Hardening execution feasibility: The graph generation model determines "triggerable," and the prediction coding model determines a risk level ≥ medium, while the impact identification model determines "triggerable," indicating consistent feasibility; ② Hardening degree matching: The deviation between the hardening degree of the impact identification model and the suggested hardening degree of the prediction coding model is ≤ a preset threshold (e.g., 20%), indicating consistent degree. If both layers of verification are satisfied, the result consistency verification is confirmed to pass. If both time consistency verification and result consistency verification pass, the consistency verification is confirmed to pass. If the consistency verification passes, a rear suspension control command is generated based on the rear suspension impact pulse data. The rear suspension control command includes the precise trigger time of rear suspension hardening, the rear suspension hardening degree, and the hardening holding time (indicating the duration of maintaining the hardened state after triggering, automatically returning to the normal state after the impact ends).

[0067] Through the consistency verification in this embodiment, the false trigger rate of rear suspension hardening can be significantly reduced, filtering out false triggers caused by noise, interference, and anomalies in a single model, ensuring that erroneous hardening does not occur in non-impact scenarios. It effectively avoids missed triggers, improving the reliability of impact recognition; triggering is only executed when all three models jointly confirm "impending impact," ensuring that real dangers are not missed. It enhances decision robustness, adapting to complex unstructured terrain, maintaining stable decision-making without jitter or abrupt changes even under highly nonlinear conditions such as cliffs, steps, and high-speed obstacle crossings. It ensures the physical rationality and temporal legality of control commands, guaranteeing that hardening commands are only executed during permitted obstacle crossing phases (e.g., front suspension jump → rear suspension awaiting impact), without being premature, delayed, or violating vehicle dynamics constraints. It improves overall vehicle control safety and ride comfort, avoiding body stiffness and abnormal suspension stiffening due to false triggers; and avoiding chassis impacts and sudden attitude changes due to missed triggers.

[0068] In some embodiments, the method further includes; In response to failing the consistency check, the rear suspension impact pulse data is corrected using the obstacle crossing structure map, the rear suspension state prediction result, and the prediction error data, and the rear suspension control command is generated.

[0069] Specifically, cases failing the consistency check include: passing the time consistency check but failing the result consistency check; failing the time consistency check but passing the result consistency check; and failing both the time consistency check and the result consistency check. Following this, the conflict adaptive correction phase begins, initiating a consistency gating correction mechanism. Using the output data of the impact recognition model as the core, and combining the structural constraints of the graph generation model and the dynamic trends of the predictive coding model, the triggering time and hardening degree of the impact recognition model are attenuated, delayed, or corrected to generate corrected rear suspension control commands. This avoids false triggering while ensuring control accuracy. The core basis for correction is a weighted confidence score, constructing a weighted confidence score for the three streams. The confidence score for the graph generation model is based on the physical interpretability of the obstacle-crossing structure graph (e.g., whether the edge weights conform to vehicle dynamics) and the stability of the obstacle-crossing stage judgment; a higher score indicates more reliable structural constraints. The confidence score for the predictive coding model directly uses prediction credibility scoring; a higher score indicates more accurate trend prediction. The confidence score of the impact identification model is based on the stability of the pulse delivery (e.g., whether it is a single isolated pulse, whether the pulse frequency matches the impact risk). A higher score indicates a more reliable event triggering. The weights of the three models are fixed during the training phase, such as impact identification model: predictive coding model: graph generation model = 0.5:0.3:0.2, to ensure the dominance of the impact identification model.

[0070] If the consistency check fails because time consistency is successful but result consistency fails, the hardening level of the impact recognition model is reduced based on the confidence scores of the predictive coding model and the map generation model (e.g., if the confidence score of the predictive coding model is 0.2, the hardening level is reduced by 50%). If the map generation model determines that it is "untriggerable," the trigger signal of the impact recognition model is directly blocked, and the rear suspension control command is not generated temporarily. If the consistency check fails because time consistency fails but result consistency is successful, the earliest triggerable time defined by the map generation model is used as the corrected trigger time, and the trigger command of the impact recognition model is delayed to ensure that the hardening operation conforms to the vehicle's physical structure constraints and avoids abnormal vehicle posture caused by premature hardening. If the consistency check fails because both time consistency and result consistency fail, and the confidence score of the impact recognition model is much higher than the other two (e.g., ≥0.8), the trigger is delayed (e.g., delayed by 5ms) and the hardening level is reduced, and the output results of the subsequent three models are observed. If the confidence score of the impact recognition model is low, the trigger signal is blocked, and the rear suspension control command is not generated temporarily. After the correction is completed, a unique fusion decision result (the final rear suspension control command) is output, including the final trigger time, the final hardening degree, and the hardening hold time, to ensure the uniqueness and reliability of the decision.

[0071] It should be noted that the fusion mechanism adopts an end-to-end collaborative learning mechanism, jointly training the three models as a whole. Through closed-loop feedback of the execution effect of the final control output, the weight distribution, input feature sensitivity, and parameters such as the confidence score weight, consistency time window, and correction rules of the fusion framework are adjusted in reverse. This achieves collaborative adaptation of the three models: matching the structural modeling of the map generation model, the trend prediction of the predictive coding model, and the event triggering of the impact recognition model, avoiding conflicts in real-time decisions caused by model biases during the training phase. It optimizes the engineering adaptability of the fusion rules: adjusting parameters such as confidence score weights and correction thresholds based on real-vehicle training data, making the fusion rules more closely reflect the actual obstacle-crossing dynamics of the vehicle. It ensures the feasibility of control commands: introducing execution constraints of the suspension electronic control system (such as the adjustment rate of stiffness / damping and response latency) during the training phase ensures that the generated commands are executable in engineering, without overshoot or execution failure.

[0072] In this embodiment, by using the structural constraints of the graph generation model and the trend prediction of the predictive coding model, isolated trigger pulses caused by instantaneous sensor glitches and high-frequency interference in the impact recognition model can be filtered out. With the impact recognition model as the core, its trigger signals are not easily rejected; instead, adjustments are made to adapt to the structure and trend, ensuring accurate capture of real impacts. The decision results from the three streams are mutually constrained and verified, avoiding frequent changes in control commands due to fluctuations in the decision-making of a single model, thus improving the smoothness of suspension control.

[0073] In some embodiments, before sending the rear suspension control command to the vehicle's suspension control system, the method further includes: Acquire real-time operating status data of the vehicle; adjust the rear suspension control command based on the real-time operating status data, and send the adjusted rear suspension control command to the suspension control system.

[0074] Specifically, the purpose of this implementation is to optimize the rear suspension control commands based on the vehicle's real-time operating status data. Real-time operating status data includes at least longitudinal vehicle speed, vehicle load, and tire pressure. The rear suspension control commands are fine-tuned using this data. For example, the faster the vehicle speed, the shorter the hardening holding time is appropriately shortened to avoid over-hardening. The greater the vehicle load, the greater the hardening degree is appropriately increased to resist compression with higher stiffness / damping. The lower the tire pressure, the more appropriately the triggering time is delayed. The adjusted rear suspension control commands are sent to the suspension control system to trigger the rear suspension hardening execution. This method ensures the executability and adaptability of the rear suspension control commands.

[0075] It should be noted that this application also provides a closed-loop verification and model parameter tuning mechanism based on real-vehicle obstacle crossing data. The goal is to deploy the three models under the fusion mechanism in a real-vehicle testing environment, perform closed-loop verification of the model's outputs through real obstacle crossing scenarios, and continuously optimize parameters based on actual vehicle response feedback to ensure the reliability and accuracy of the rear suspension pre-hardening strategy in dynamic environments. The inputs to the closed-loop verification are the rear suspension control commands output by the fusion mechanism and the vehicle's measured sensor data, including the rear suspension hardening trigger time, suspension response intensity, vehicle impact acceleration, and vehicle attitude changes. The outputs are parameter update suggestions for each model, rear suspension control correction commands, and execution effect evaluation indicators. The closed-loop verification consists of three parts: model output verification, control execution verification, and feedback-driven optimization. Model output verification mainly evaluates the consistency and stability of each model in real-vehicle operation, especially the adaptability of the obstacle crossing structure map under different terrains, the tracking error of the predictive coding model for state trends, and the time accuracy of the impact recognition model in impact precursor recognition. The control execution verification focuses on the effectiveness of the rear suspension hardening strategy. By comparing key indicators such as peak vehicle impact, rear wheel bounce amplitude, and chassis ground impact frequency under no-control and controlled conditions, it determines whether the rear suspension control commands effectively reduce impact at the appropriate time. The feedback-driven optimization mechanism records the model's performance under different terrains, speeds, and loads, fine-tuning key parameters such as input feature weights, decision thresholds, and time window lengths to ensure the model's robustness and adaptability in varying environments. This mechanism employs a phased parameter feedback mode; after each round of testing, the system automatically calculates the deviation between the preset target and the measured results, using this deviation information to guide local updates of model parameters, gradually converging to the optimal control behavior. During closed-loop verification, manually annotated keyframes are also introduced as a verification benchmark to accurately evaluate the model's accuracy in tasks such as impact precursor detection and obstacle recognition. This ensures a seamless transition from simulation training to real-vehicle deployment, enabling the rear suspension hardening control to execute reliably in real-world road scenarios and further improving the vehicle's passability and comfort on unstructured terrain.

[0076] It should be noted that the method in this embodiment can be executed by a single device, such as a computer or server. The method can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method in this embodiment, and the multiple devices will interact with each other to complete the method described.

[0077] It should be noted that some embodiments of this application have been described above. In some cases, the actions or steps described in the above embodiments can be performed in a different order than that shown in the above embodiments and the desired result can still be achieved. In addition, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0078] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides a vehicle control device.

[0079] refer to Figure 2 The vehicle control device includes: The acquisition module 202 is configured to acquire multi-source perception data during vehicle operation, and to obtain temporal perception data characterizing the vehicle obstacle crossing process by preprocessing the multi-source perception data. The first generation module 204 is configured to generate an obstacle crossing structure map based on the time-series sensing data and through a pre-trained map generation model. The second generation module 206 is configured to generate the vehicle's rear suspension state prediction result and prediction error data based on the obstacle crossing structure map and a pre-trained prediction coding model. The third generation module 208 is configured to generate rear suspension impact pulse data based on the rear suspension state prediction result and the prediction error data, using a pre-trained impact recognition model. The fourth generation module 210 is configured to generate a rear suspension control command based on the obstacle structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data; and send the rear suspension control command to the vehicle's suspension control system so that the rear suspension control command can be executed by the suspension control system.

[0080] In some embodiments, the first generation module 204 is configured to encode the time-series sensing data into feature vectors of each node in the obstacle crossing structure map using a time-series encoder in the map generation model; wherein each node represents a component of the vehicle; based on the feature vectors of each node and prior knowledge of vehicle physics, establish connection edges between each node to generate an initial structure map; and input the initial structure map into a graph neural network in the map generation model for iterative propagation to generate the obstacle crossing structure map.

[0081] In some embodiments, the second generation module 206 is configured to extract temporal data of the target node from the obstacle crossing structure map; extract features from the temporal data of the target node through the encoder in the predictive coding model to obtain a hidden state vector; and perform multi-step forward state prediction through the prediction network in the predictive coding model based on the hidden state vector to generate the rear-hanging state prediction result. Based on the rear suspension state prediction results and real-time sensor data, the error is calculated through the error network in the prediction coding model to generate the prediction error data.

[0082] In some embodiments, the third generation module 206 is configured to convert the rear suspension state prediction result and the prediction error data into pulse sequence data through the pulse coding layer in the impact recognition model; to perform weighted modulation on the pulse sequence data through the synaptic weight modulation layer in the impact recognition model to form weighted pulse sequence data; and to output the rear suspension impact pulse data based on the weighted pulse sequence data by the decision output layer in the impact recognition model using a membrane potential integration mechanism.

[0083] In some embodiments, the fourth generation module 210 is configured to perform a consistency check based on the obstacle crossing structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data; in response to passing the consistency check, to generate a rear suspension control command based on the rear suspension impact pulse data, so that the vehicle's suspension control system executes the rear suspension control command.

[0084] In some embodiments, the fourth generation module 210 is configured to, in response to failing the consistency check, correct the rear suspension impact pulse data using the obstacle crossing structure map, the rear suspension state prediction result, and the prediction error data, and generate the rear suspension control command.

[0085] In some embodiments, the fourth generation module 210 is configured to perform time consistency verification based on the obstacle crossing structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data; perform result consistency verification based on the obstacle crossing structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data; and determine that the consistency verification has been passed in response to passing both the time consistency verification and the result consistency verification.

[0086] In some embodiments, before sending the rear suspension control command to the vehicle's suspension control system, a correction module is further included, configured to acquire real-time operating status data of the vehicle; adjust the rear suspension control command according to the real-time operating status data; and send the adjusted rear suspension control command to the suspension control system.

[0087] In some embodiments, a training module is further included, configured to: construct an initial impact recognition model and a training dataset; input data from the training dataset into the initial impact recognition model, and output predicted rear-end impact pulse data through the initial impact recognition model; construct a temporal accuracy loss function with the objective of minimizing the deviation between the predicted impact trigger time and the actual impact trigger time in the predicted rear-end impact pulse data; construct a false trigger penalty loss function and a missed detection penalty loss function based on the predicted rear-end impact pulse data; construct an event-driven joint loss function by weighted summation based on the temporal accuracy loss function, the false trigger penalty loss function, and the missed detection penalty loss function; and train the initial impact recognition model by minimizing the event-driven joint loss function to obtain the impact recognition model.

[0088] 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.

[0089] The apparatus of the above embodiments is used to implement the corresponding vehicle control method in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0090] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the vehicle control method described in any of the above embodiments.

[0091] Figure 3 This embodiment illustrates a more specific hardware structure of an electronic device. The device may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0092] 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.

[0093] 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.

[0094] 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.

[0095] 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.).

[0096] 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.

[0097] 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.

[0098] The electronic devices described above are used to implement the corresponding vehicle control methods in any of the foregoing embodiments and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0099] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides a vehicle, including: a memory for storing executable program code; and a processor for calling and running the executable program code from the memory, so that the vehicle performs the vehicle control method as described in the above embodiments.

[0100] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium that stores computer instructions for causing the computer to execute the vehicle control method as described in any of the above embodiments.

[0101] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0102] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the vehicle control method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0103] Based on the same concept, corresponding to any of the above embodiments, this application also provides a computer program product, including computer program instructions, which, when run on a computer, cause the computer to perform the method described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0104] 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.

[0105] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to choose, based on the prompt message, whether to provide personal information to the software or hardware such as the electronic device, application, server, or storage medium performing the operations of this disclosed technical solution.

[0106] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" regarding the provision of personal information by the electronic device.

[0107] 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.

[0108] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0109] 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.

[0110] 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.

[0111] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. A vehicle control method, characterized in that, include: Multi-source perception data is acquired during vehicle operation, and the multi-source perception data is preprocessed to obtain temporal perception data characterizing the vehicle's obstacle-crossing process. Based on the time-series sensing data, an obstacle-crossing structure map is generated using a pre-trained map generation model. Based on the obstacle crossing structure map, the vehicle's rear suspension state prediction results and prediction error data are generated through a pre-trained prediction coding model. Based on the rear suspension state prediction results and the prediction error data, rear suspension impact pulse data are generated using a pre-trained impact recognition model. Based on the obstacle structure diagram, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data, a rear suspension control command is generated; the rear suspension control command is sent to the vehicle's suspension control system to execute the rear suspension control command through the suspension control system.

2. The method according to claim 1, characterized in that, The process of generating an obstacle-crossing structure map based on the time-series-aware data using a pre-trained map generation model includes: The temporal sensing data is encoded into feature vectors of each node in the obstacle crossing structure map by the temporal encoder in the map generation model; wherein each node represents a component of the vehicle. Based on the feature vectors of each node and prior knowledge of vehicle physics, connection edges are established between each node to generate an initial structural graph. The initial structure map is input into the graph neural network in the map generation model for iterative propagation to generate the obstacle-crossing structure map.

3. The method according to claim 1, characterized in that, Based on the obstacle-crossing structure map, the pre-trained prediction coding model generates prediction results and prediction error data for the vehicle's rear suspension state, including: Extract the temporal data of the target nodes from the obstacle crossing structure map; The encoder in the predictive coding model extracts features from the temporal data of the target node to obtain the hidden state vector. Based on the hidden state vector, multi-step forward state prediction is performed through the prediction network in the prediction coding model to generate the rear-hanging state prediction result. Based on the rear suspension state prediction results and real-time sensor data, the error is calculated through the error network in the prediction coding model to generate the prediction error data.

4. The method according to claim 1, characterized in that, The step of generating rear suspension impact pulse data based on the predicted rear suspension state and the predicted error data, using a pre-trained impact recognition model, includes: The pulse coding layer in the impact recognition model converts the rear suspension state prediction result and the prediction error data into pulse sequence data. The pulse sequence data is weighted by the synaptic weight modulation layer in the impact recognition model to form weighted pulse sequence data. Based on the weighted pulse sequence data, the decision output layer in the impact identification model outputs the rear suspension impact pulse data using a membrane potential integration mechanism.

5. The method according to claim 1, characterized in that, The step of generating rear suspension control commands based on the obstacle-crossing structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data includes: A consistency check is performed based on the obstacle crossing structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data. In response to passing the consistency check, a rear suspension control command is generated based on the rear suspension impact pulse data, so that the vehicle's suspension control system executes the rear suspension control command.

6. The method according to claim 5, characterized in that, The method further includes; In response to failing the consistency check, the rear suspension impact pulse data is corrected using the obstacle crossing structure map, the rear suspension state prediction result, and the prediction error data, and the rear suspension control command is generated.

7. The method according to claim 5, characterized in that, The consistency verification based on the obstacle-crossing structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data includes: Based on the obstacle crossing structure map, the rear suspension state prediction result, the prediction error data, and the rear suspension impact pulse data, a time consistency check is performed. Based on the obstacle crossing structure map, the rear suspension state prediction results, the prediction error data, and the rear suspension impact pulse data, a consistency verification of the results is performed. In response to passing both the time consistency check and the result consistency check, it is determined that the consistency check has been passed.

8. The method according to claim 1, characterized in that, Before sending the rear suspension control command to the vehicle's suspension control system, the method further includes: Obtain real-time operating status data of the vehicle; The rear suspension control command is adjusted based on the real-time operating status data, and the adjusted rear suspension control command is sent to the suspension control system.

9. The method according to claim 1, characterized in that, The training process of the impact recognition model includes: Construct the initial impact recognition model and training dataset; The data in the training dataset is input into the initial impact recognition model, and the initial impact recognition model outputs the predicted rear overhang impact pulse data. With the goal of minimizing the deviation between the predicted impact trigger time and the actual impact trigger time in the predicted rear overhang impact pulse data, a timing accuracy loss function is constructed; Based on the predicted rear overhang impact pulse data, construct the false trigger penalty loss function and the missed detection penalty loss function; Based on the time-series precision loss function, the false triggering penalty loss function, and the missed detection penalty loss function, an event-driven joint loss function is constructed by weighted summation; The initial impact recognition model is trained by minimizing the event-driven joint loss function to obtain the impact recognition model.

10. A vehicle, characterized in that, The vehicles include: Memory, used to store executable program code; A processor for calling and running the executable program code from the memory, causing the vehicle to perform the method as described in any one of claims 1 to 9.

Images