A control method, device and vehicle

Abstract

This application discloses a control method, device, and vehicle. It acquires multi-source data during the switching process and encodes it into state segments. Based on these state segments, a candidate damping strategy sequence is generated using a generator in a dual-stream generative adversarial network (GAN), while an encoder outputs low-dimensional latent variables characterizing the vehicle's operating conditions. Using the low-dimensional latent variables and the candidate strategy sequence as input, an optimal damping control path is obtained through iterative optimization using a policy gradient reinforcement learning model. A discriminant subnetwork scores the candidate strategy sequence and the optimal damping control path, and the final damping adjustment strategy sequence is determined based on the scoring results. This sequence is then converted into physical commands that meet the execution conditions of the suspension controller and output. This application achieves smooth transition and real-time adaptive adjustment of the damping strategy through the coordinated operation of three models: generation, understanding, and decision-making, significantly improving the comfort and stability of the vehicle under dynamic driving conditions.

Description

Technical Field

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

[0002] With the continuous development of intelligent chassis technology, the damping control of vehicle suspension systems is gradually evolving from passive adjustment to active adaptive control. Especially in driving scenarios where off-road and on-road conditions frequently switch, road surface characteristics, vehicle posture, and driving demands all undergo drastic changes. Achieving a smooth transition in damping parameters directly affects vehicle stability and ride comfort. Currently, intelligent chassis systems integrate various sensors, such as accelerometers, gyroscopes, suspension height sensors, cameras, and radar, enabling them to perceive vehicle status and road environment, providing a data foundation for adaptive damping adjustment. However, how to utilize this multi-source information to generate reasonable damping control strategies during complex condition transitions remains a technological challenge for the industry.

[0003] Traditional chassis control systems typically adjust damping based on fixed rules or simple mapping functions. This fails to accurately identify abnormal damping parameter maintenance caused by sudden changes in operating conditions, especially when merging from off-road terrain to highway at high speeds. High-damping states are often incorrectly maintained, leading to body roll and ride discomfort. Specifically, existing systems rely on pre-calibrated lookup tables or linear interpolation methods to map finite state variables such as current speed and acceleration to damping setpoints, lacking a deep understanding of scene transitions. When a vehicle rapidly transitions from off-road conditions to a flat highway, the system often fails to promptly perceive the change in road surface type, continuing to use the high-damping strategy of off-road mode. This results in an overly stiff suspension response and excessive vertical body roll, severely impacting the driving experience. Furthermore, fixed-rule methods lack learning and generalization capabilities, making it difficult to adapt to different driving styles, load conditions, and road surface combinations, highlighting a rigid strategy. Summary of the Invention

[0004] In view of the above problems, this application provides a control method, device and vehicle.

[0005] The embodiments of this application disclose the following technical solutions: The first aspect of this application provides a control method, including: Acquire multi-source data during the process of the vehicle switching from the first state to the second state, and encode the multi-source data to obtain state segments; Based on the state segments, a sequence of candidate damping strategies is generated using a generator from a two-stream generative adversarial network; based on the encoder, feature extraction and reparameterization sampling are performed on the state segments to output low-dimensional latent variables that characterize the vehicle's operating conditions. Using the low-dimensional latent variables and the candidate damping policy sequence as inputs, the optimal damping control path is obtained by iterative optimization through a policy gradient reinforcement learning model. The candidate damping strategy sequence is scored against the optimal damping control path by a discriminant subnetwork, and the final damping adjustment strategy sequence is determined based on the scoring results. The final damping adjustment strategy sequence is converted into physical commands that meet the execution conditions of the suspension controller and output to the suspension controller.

[0006] This application provides a control system in embodiment two, including: The acquisition unit is used to acquire multi-source data during the process of the vehicle switching from the first state to the second state, and to encode the multi-source data to obtain a state segment. The generation unit is used to generate a sequence of candidate damping strategies based on the state segments using a generator from a two-stream generative adversarial network; and to perform feature extraction and reparameterization sampling on the state segments based on the encoder, outputting low-dimensional latent variables to characterize the vehicle's operating conditions. An iterative unit is used to take the low-dimensional latent variables and the candidate damping policy sequence as inputs, and perform iterative optimization through a policy gradient reinforcement learning model to obtain the optimal damping control path. The determining unit is used to score the candidate damping strategy sequence and the optimal damping control path through a discriminant subnetwork, and determine the final damping adjustment strategy sequence based on the scoring results; The output unit is used to convert the final damping adjustment strategy sequence into physical commands that meet the execution conditions of the suspension controller and output them to the suspension controller.

[0007] A third aspect of this application provides an electronic device, including: a processor, a memory, and a system bus; the processor and the memory are connected via the system bus; the memory is used to store one or more programs, the one or more programs including instructions, which, when executed by the processor, cause the processor to perform the control method of the first aspect described above.

[0008] A fourth aspect of this application provides a vehicle including an electronic device as described in the third aspect above.

[0009] Compared with the prior art, this application has the following beneficial effects: Multi-source data during vehicle state transitions is acquired and encoded into structured state fragments, enabling the vehicle to comprehensively perceive scene information such as vehicle speed, road surface type, and suspension status, overcoming the limitations of traditional fixed-rule-based systems that rely on finite state quantities. Next, a generator using a dual-stream generative adversarial network (GAN) generates a sequence of candidate damping strategies with smoothness and responsiveness. Simultaneously, an encoder outputs low-dimensional latent variables representing the evolution trend of operating conditions, achieving a deep understanding and compressed representation of scene transitions. Then, using the latent variables and candidate strategies as input, an iterative optimization model using a policy gradient reinforcement learning model yields the optimal damping control path, enabling the control strategy to possess dynamic learning and adaptive capabilities, overcoming the rigidity and inability to generalize of fixed-rule strategies. A discriminant sub-network scores and dynamically merges the candidate strategies and optimized paths, ensuring the final strategy's advantages in control rationality, operating condition adaptability, and smoothness. Finally, the strategy is converted into physical commands and output to the suspension controller. Thus, the vehicle can perceive changes in operating conditions in real time during off-road to on-road transitions, dynamically generate and optimize damping strategies, avoid maintaining high-damping errors, achieve a smooth transition, and improve ride comfort and vehicle stability. Attached Figure Description

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

[0011] Figure 1 A flowchart of a control method provided in an embodiment of this application; Figure 2 A flowchart illustrating another control method provided in this application embodiment; Figure 3 This is a structural diagram of a control system provided in an embodiment of this application. Detailed Implementation

[0012] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0013] To facilitate understanding of the technical solutions provided in the embodiments of this application, the technical terms involved in the embodiments of this application will be explained below.

[0014] Dropout is a regularization technique used in neural networks, also known as a random deactivation layer. Its core idea is to randomly set the output of some neurons to zero with a certain probability during model training, preventing them from participating in the current forward and backward propagation. This avoids excessively strong co-fitting relationships between neurons, effectively preventing overfitting and improving the model's generalization ability. In the technical solution of this patent, the introduction of a Dropout layer into the discriminator is precisely to enhance its generalization performance and avoid overfitting to noise or specific patterns in the training data.

[0015] The Transformer Variational Autoencoder (VAE) is a deep learning model that combines the Transformer architecture with the variational autoencoder framework. The Transformer part utilizes multi-head self-attention and positional encoding to capture long-range dependencies between different time steps and modal features in the input sequence (such as multi-source state segments). The variational autoencoder part maps the input to the mean and variance of the latent variable distribution through an encoder, then obtains low-dimensional latent variables through reparameterization sampling, and finally reconstructs the original input through a decoder. This model balances the global expressive power of sequence modeling with the continuous smoothness of the probabilistic latent space, making it suitable for efficient compression and scene feature extraction of high-dimensional, temporal, and multimodal state segments in this application.

[0016] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0017] As mentioned earlier, intelligent chassis systems generally suffer from problems such as delayed damping control response, rigid control strategies, and scene judgment errors during the transition from off-road to on-road conditions, making it difficult to meet the demands for ride comfort and vehicle stability under dynamic driving conditions. Traditional chassis damping control systems often rely on fixed rule bases or simple mapping functions for control decisions, failing to accurately identify abnormal damping parameter holding phenomena caused by sudden changes in operating conditions. A typical scenario is that when a vehicle merges into a highway at high speed after completing off-road driving, the suspension system incorrectly maintains a high damping setting, resulting in continuous body sway and significantly increased ride discomfort, affecting the driving experience and driving safety.

[0018] Analysis revealed that the root cause of the aforementioned problems lies in the existing system's lack of deep understanding of the scene transition process, making it unable to dynamically generate and update damping control strategies based on real-time changes in road surface type, vehicle dynamics, and environmental information. Specifically, traditional solutions fail to establish an effective collaborative mechanism between generative, interpretive, and decision-making models, resulting in delays in system recognition of operating condition changes, lags in strategy adjustments, and even control command failures. Furthermore, existing methods lack the ability to smoothly transition damping control paths when dealing with continuously changing transitional scenarios, making it difficult to balance response speed and execution stability.

[0019] To address the aforementioned issues, this application implements a unified encoding method for multi-source data during the vehicle's transition from off-road to on-road driving, resulting in state segments. Next, a generator using a dual-stream generative adversarial network (GAN) generates candidate damping strategy sequences, while an encoder outputs low-dimensional latent variables representing the driving conditions. Then, using the latent variables and candidate sequences as input, iterative optimization through policy gradient reinforcement learning yields the optimal damping control path. A discriminant subnetwork scores the two strategy paths, dynamically weighting or selecting the best to determine the final damping adjustment strategy sequence. Finally, the damping smooth transition mapping function converts the signals into physical commands, which are then output to the suspension controller, and closed-loop parameter tuning is performed based on real-vehicle data. This application, through the synergistic interaction of generation, understanding, and decision-making models, achieves a smooth transition and real-time adaptive adjustment of damping strategies during the off-road to on-road scenario transition, effectively solving problems of response lag, policy rigidity, and scenario misjudgment, significantly improving the comfort and stability of the vehicle under dynamic driving conditions.

[0020] It should be noted that the control methods, systems, products, devices, and media provided in this application can be applied to the field of automotive electronics technology. The above are merely examples and do not limit the application areas of the control methods, systems, products, devices, and media provided in this application. Furthermore, the embodiments of this application may not limit the executing entity of the control. For example, the control method in the embodiments of this application can be executed by a controller in a vehicle. The controller can be a chassis domain controller, a central computing platform, a high-performance processor, an edge computing node, or a control module integrated into the suspension electronic control unit. The controller may include a processor and a memory, the memory storing executable computer program instructions, and the processor executing the instructions to implement the aforementioned steps. The controller can also interact with the vehicle's perception system, positioning system, power system, and execution system through a unified data interface module to obtain multi-source data and issue damping control commands. In addition, the controller can be deployed on a single hardware node or distributed across multiple nodes, with the nodes coordinating through an in-vehicle communication network. This application does not limit the specific implementation form and deployment method of the controller.

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0022] The following embodiment illustrates a control method provided in this application. See also: Figure 1 ,Should Figure 1 A flowchart of a control method provided in this application embodiment, the method comprising: S101. Acquire multi-source data during the process of the vehicle switching from the first state to the second state, and encode the multi-source data to obtain a state segment.

[0023] The system acquires multi-source data during the vehicle's transition from a first state to a second state. The data collection scope covers multi-dimensional input information from vehicle sensors, environmental perception modules, chassis control systems, and navigation and positioning units. The first state can be an off-road state, and the second state can be a road state.

[0024] Specifically, the vehicle sensor layer mainly includes accelerometers, gyroscopes, suspension height sensors, wheel speed sensors, and the current damping setpoint and response status provided by the chassis domain controller. Environmental perception data comes from forward-facing cameras, surround-view systems, and millimeter-wave radar, which are responsible for capturing information such as changes in road structure, terrain undulations, curbs, or potholes. Simultaneously, the navigation system, combined with high-precision map matching, extracts classification labels for the vehicle's current road segment, such as gravel road, slope, asphalt, dry, and slippery. To enhance the accuracy of scene labels, vehicle-to-everything (V2X) information is also introduced to determine the road topology and traffic conditions around the vehicle. All of the above data together constitute multi-source input.

[0025] After data acquisition, unified timestamp alignment and interpolation are performed across different data sources to ensure data structure consistency and temporal continuity. Data preprocessing also includes sensor noise filtering, outlier removal, and standardization. All preprocessed data is then fed into the feature encoding module for initial encoding: continuous variables, such as velocity, acceleration, and damping setpoints, are standardized and represented as dense vectors; image data undergoes size unification and normalization; and geographic and road labels are represented as discrete feature vectors using one-hot encoding.

[0026] Finally, by constructing a unified data structure, data from different sources are aligned and packaged into unified state segments. Each segment covers a 5-10 second time window during the vehicle's transition from off-road to on-road conditions, serving as the basic unit for model training. The output is a structured state segment dataset, containing the original data, preprocessed feature vectors, and stage labels for the vehicle's current operating condition. This state segment will serve as the common input for the subsequent Dual-Stream GAN, Transformer-VAE, and Policy Gradient RL models, providing a stable data foundation for multi-model collaborative learning and ensuring support for real-time data stream input in practical deployments. The overall goal is to achieve multi-source information fusion and standardized processing, providing accurate input for dynamic damping adjustment during complex scene transitions.

[0027] S102. Based on the state segment, a candidate damping strategy sequence is generated by the generator of a two-stream generative adversarial network; based on the encoder, feature extraction and reparameterization sampling of the state segment are performed to output low-dimensional latent variables to characterize the vehicle's operating conditions.

[0028] The Dual-Stream Generative Adversarial Network (DAN) described in this application refers to a generator and a discriminator each forming an independent adversarial learning stream. These two streams collaborate through game-like training to form a cooperative optimization mechanism: the generator uses a multi-layer residual convolutional network architecture, receiving encoded state fragments and generating candidate damping policy sequences; the discriminator employs a fully connected neural network structure, receiving real damping control trajectories and the generator's output, and judging realism based on dimensions such as physical plausibility, policy continuity, and scene matching. The two adversarial streams iterate alternately and mutually promote each other, ultimately enabling the generator to output damping policy sequences with high realism and scene adaptability. This dual-stream structure achieves controllable generation of damping control paths in a high-dimensional state space without explicit modeling of dynamic equations, providing a priori policy foundation for the subsequent multi-model fusion control in this application.

[0029] On the one hand, the state fragments are input into the generator of the dual-stream generative adversarial network. The generator uses a multi-layer residual convolutional network and embeds a scene label conditional embedding mechanism. It extracts features from the state fragments to obtain high-dimensional feature representations, and then upsamples and maps them into a continuous damping adjustment parameter sequence to form a candidate damping strategy sequence covering each control cycle.

[0030] On the other hand, the encoder includes a Transformer variational autoencoder. The same state segment is input into the Transformer variational autoencoder, and the global correlation features between multi-source data at different times are extracted using a multi-head self-attention mechanism. After obtaining a high-dimensional working condition feature vector, it is fed into the mean network and the variance network respectively, and the mean parameter and variance parameter of the latent variable distribution are output. Then, after reparameterization sampling, a low-dimensional latent variable that follows a preset distribution is generated. This low-dimensional latent variable is used to characterize the current working condition of the vehicle.

[0031] In one possible implementation, the step of generating a candidate damping strategy sequence based on the state fragment using a generator of a two-stream generative adversarial network includes: inputting the state fragment output from S101 into the generator of the two-stream generative adversarial network. This generator employs a multi-layer residual convolutional network as its backbone architecture. Its front-end network is responsible for extracting high-dimensional latent representations from encoded features. The middle layer integrates scene labels, such as road surface type and slope, into the latent space through a conditional embedding mechanism, enabling the generation process to have context-aware capabilities. The back-end layer outputs a continuous damping adjustment value sequence through a progressive upsampling method, corresponding to the damping setpoint change trajectory for each control cycle within the next few seconds. This sequence is the candidate damping strategy sequence, possessing temporal consistency and boundary constraint capabilities.

[0032] In one possible implementation, the dual-stream generative adversarial network also includes a discriminator. This discriminator employs a fully connected neural network structure, containing multilayer perceptrons, using LeakyReLU activation functions, and incorporating dropout layers to prevent overfitting. During training, the discriminator receives real damping control trajectories and candidate damping policy sequences output by the generator, judging the latter's level of realism in terms of physical plausibility, policy continuity, and scene matching. The generator and discriminator alternately optimize through adversarial training: the generator strives to output policy sequences indistinguishable from real damping control trajectories, while the discriminator works to improve discrimination accuracy; together, they enhance the plausibility and diversity of generated policies. A phased strategy is employed during training: policy pre-training is first performed in a single scene, followed by joint adversarial learning under mixed scene conditions. Finally, the trained generator can quickly generate candidate damping policy sequences based on real-time input states in actual deployment, serving as one of the inputs for subsequent fusion modules.

[0033] The true damping control trajectory refers to a series of damping adjustment commands generated by an empirically calibrated or verified suspension control system during actual vehicle operation. It records the actual damping force or damping coefficient changes performed by the damper in each control cycle as the vehicle switches from one state to another. This trajectory serves as a supervisory label for training the dual-stream generative adversarial network, reflecting the optimal or suboptimal control behavior that conforms to physical laws, vehicle dynamics response, and ride comfort requirements. It provides a benchmark for the discriminator to judge authenticity, guiding the generator to learn and produce damping strategy sequences that closely resemble the calibration results achieved by human engineers.

[0034] In one possible implementation, the step of extracting features and reparameterizing the state segment based on the encoder to output low-dimensional latent variables characterizing the vehicle's operating condition includes: The same state segments are input into a Transformer variational autoencoder. This model combines the global dependency capability of Transformer in long sequence modeling with the advantages of VAE in unsupervised latent variable learning. The encoder part consists of a multi-layer Transformer structure with positional encoding, which uses a multi-head self-attention mechanism to capture the long-term dependencies and multi-dimensional interactions between different time points and modal data in the state sequence, such as vehicle speed, acceleration, image features, and road labels, and extracts a high-dimensional condition feature vector. Subsequently, this high-dimensional feature vector is input into the mean network and variance network of the variational autoencoder, respectively, and outputs the mean and variance parameters of the latent variable distribution. Based on this, a reparameterization technique is used to sample from this distribution to generate a low-dimensional latent variable z.

[0035] This latent variable encapsulates the complex evolution trajectory, dynamic behavior trends, and road environment changes of the vehicle during the transition from off-road to on-road conditions. It serves as the state encoding input for the subsequent policy gradient reinforcement learning model and is also fused with the intermediate layer features of the dual-stream GAN generator to semantically enhance and constrain the generated damped policy. The training objective of Transformer-VAE is to simultaneously minimize the reconstruction loss (enabling the decoder to reconstruct the original state from the latent variable) and the KL divergence loss (constraining the latent variable distribution to approximate a standard normal distribution), thereby maintaining the reconstructability of information and the smooth continuity of the latent space while reducing dimensionality. Through this step, high-dimensional redundant state fragments are transformed into low-dimensional controllable latent variable representations, providing an efficient and interpretable information foundation for the overall system's model fusion and control decisions.

[0036] S103. Using the low-dimensional latent variables and the candidate damping policy sequence as input, the optimal damping control path is obtained through iterative optimization using a policy gradient reinforcement learning model.

[0037] This step constructs and trains a policy gradient-based reinforcement learning model to learn the optimal damping control path from historical damping adjustment behaviors and real-time scenario latent variables. The model's goal is not simply to predict damping values, but rather, under the premise of ensuring vehicle attitude stability and comfort constraints, to select the optimal dynamic damping response sequence in different states through comparison and optimization with the generated policy, thereby achieving a smooth transition from off-road to on-road scenarios. The model input consists of two main parts: first, a low-dimensional latent variable z output by the Transformer-VAE, which serves as a compressed representation of the current overall vehicle operating condition, condensing the vehicle's evolution trajectory, dynamic behavior trends, and road environment changes during state transitions; second, a sequence of candidate damping policies generated by the Dual-Stream GAN generator, which serves as a reference policy and initial trajectory for learning, providing prior guidance for reinforcement learning.

[0038] The reinforcement learning network adopts an Actor-Critic structure based on policy gradients. The policy network (Actor) takes the low-dimensional latent variables of the current state as input and outputs continuous actions, i.e., a series of future damping adjustment magnitudes or target values. The action space is modeled using continuous actions to support fine-tuning of the damping setpoint. The value network (Critic) estimates the expected cumulative reward of the current policy in this state, which is used to guide the update of the policy network parameters.

[0039] Considering multiple control objectives, the reward function can include at least one of the following four rewards: smoothness reward, used to quantify the fluctuation of the vehicle's vertical acceleration (smaller fluctuations result in higher rewards); responsiveness reward, used to evaluate the contribution of damping adjustments to vehicle attitude control under the current road type, i.e., whether the adjustment can quickly and effectively suppress dynamic responses such as roll and pitch; policy smoothness reward, used to penalize excessively rapid or abrupt damping jumps to improve ride comfort; and matching reward, used to guide the policy output to approach the target policy generated by the Dual-Stream GAN, thereby achieving synergy between the generative model and the reinforcement learning model. Model training employs Proximal Policy Optimization (PPO) to improve convergence stability, and combines an experience replay mechanism to batch update historical state-action-reward sequences under different road conditions, improving sample utilization efficiency. In each training iteration, the policy network outputs the damping adjustment trajectory for the current latent variable state, while the value network simultaneously estimates its long-term performance, thereby dynamically optimizing the policy function.

[0040] Through the aforementioned iterative optimization, the finally trained policy network can receive real-time encoded low-dimensional latent variables during actual vehicle operation and directly output damping adjustment commands for each control cycle, achieving online inference and control output of the optimal damping control path. This optimal damping control path satisfies vehicle attitude stability and comfort constraints, serving as both the final control input of the actual controller and a fusion comparison with the generator output to achieve cross-validation and adaptive adjustment between models. The core innovation of this reinforcement learning model lies in introducing the generator policy as a guiding path, while itself acting as a dynamic correction mechanism. This compensates for the response lag problem of the generative model in rapidly changing scenarios, achieving complementary control through both generation and policy mechanisms. Consequently, it provides a more continuous and predictive dynamic response in the nonlinear control space, effectively supporting the autonomous control capability of the chassis system during complex road condition transitions.

[0041] S104. The candidate damping strategy sequence and the optimal damping control path are scored by the discriminant subnetwork, and the final damping adjustment strategy sequence is determined based on the scoring results.

[0042] The candidate damping policy sequence generated by the two-stream generative adversarial network generator in the preceding steps is compared and evaluated with the optimal damping control path optimized by the policy gradient reinforcement learning model. Based on the evaluation results, the final damping adjustment policy issued to the execution layer is dynamically determined. The process relies on a lightweight discriminant subnetwork, whose structure is derived from a simplified version of the two-stream generative adversarial network discriminator, used to evaluate the merits of the two policy paths under the current vehicle state.

[0043] For example, the candidate damping strategy sequence and the optimal damping control path are first synchronously input into the discriminant subnetwork. The discriminant subnetwork extracts the temporal features of the two strategy sequences respectively. The extracted feature parameters include the damping adjustment amplitude, adjustment frequency, and temporal continuity features within each control cycle. The adjustment frequency is the number of adjustments per unit time, and the temporal continuity features include the gradient of change between adjacent cycles and the overall fluctuation trend of the sequence. Based on the extracted feature parameters, the discriminant subnetwork compares and analyzes the two strategy sequences from three dimensions: Control rationality, assessing whether the damping adjustment exceeds the physical boundaries of the suspension actuator, such as maximum and minimum damping limits, response delay constraints, and whether the adjustment direction aligns with vehicle attitude control requirements (suppressing roll and pitch); Operating condition adaptability, associating and matching the strategy sequence with the currently input low-dimensional latent variables, which represent road surface type, vehicle speed, vertical acceleration, etc., to determine whether the damping setting meets the typical requirements of the current road conditions, such as high damping for off-road surfaces and low damping for highway surfaces; and Control smoothness, calculating the damping change rate between adjacent control cycles in the strategy sequence, assessing whether there are sudden jumps or severe oscillations, and counting the frequency of changes exceeding a preset threshold. The discriminant subnetwork, based on preset scoring criteria, weights and synthesizes the analysis results from the above three dimensions, outputting the scores corresponding to the candidate damping strategy sequence and the optimal damping control path, respectively. This score quantifies the degree to which two strategy sequences are adapted to the vehicle's suspension control requirements. A higher score indicates that the strategy is better in terms of physical rationality, scenario matching, and execution smoothness.

[0044] After obtaining the scores of the two policy sequences, the final damping adjustment policy sequence is determined based on the scores. For example, specific implementation methods include: dynamically weighting the candidate damping policy sequence and the optimal damping control path according to the score value; that is, weighting the corresponding control period damping values ​​of the two policy sequences according to the score ratio to generate a compromise policy; or directly selecting the policy sequence with the higher score as the final output. Furthermore, confidence assessments can be performed on the stability of the low-dimensional latent variables output by the Transformer-VAE and the consistency of outputs between models. If the confidence level is lower than a preset threshold, the output priority of the optimal damping control path will be automatically triggered to improve the system's control robustness under abnormal conditions. Simultaneously, the system also establishes a control intent feedback mechanism to track and model the deviation between the final output policy and the historical real policy, feeding back the deviation characteristics to the intermediate layer between the generator and the policy network to achieve continuous adaptive adjustment of control performance. Through the aforementioned scoring and fusion mechanism, the system can fully leverage the complementary advantages of the generative model and the reinforcement learning model, ensuring that the final output damping adjustment strategy sequence combines the prior rationality of the generative strategy with the dynamic correction capability of the reinforcement learning strategy, thereby achieving smooth, stable, and comfortable damping control during the transition from off-road to highway scenarios.

[0045] S105. The final damping adjustment strategy sequence is converted into physical commands that meet the execution conditions of the suspension controller and output to the suspension controller.

[0046] The final damping adjustment strategy sequence output after multi-model fusion is transformed into physical control commands that can be executed in real time by the vehicle's electronic control unit (ECU). A smooth damping transition mapping function is constructed between the strategy layer and the execution layer to ensure that the strategy output strictly adapts to the response characteristics and control cycle of the chassis control hardware while meeting dynamic performance requirements. The final damping adjustment strategy sequence output by the fused model is a set of damping adjustment values ​​with time steps, representing the target damping settings that the four wheels should be adjusted to within several future control cycles. However, this sequence is still in the logic-level control domain and needs to be converted into hardware-executable command parameters through a specially designed mapping function.

[0047] The mapping function is constructed based on the actual control interface of the chassis control system. First, according to the vehicle platform and suspension system type, the maximum damping setpoint, minimum damping setpoint, minimum adjustable resolution, and response delay threshold of each channel are set.

[0048] These parameters define the physical boundaries of damping adjustment. The maximum damping setting is the upper limit of the maximum damping force or damping coefficient that the suspension actuator can provide. This value is determined by the vehicle platform and suspension system type. Damping commands exceeding this upper limit will not be executed or will be forcibly truncated to prevent hardware overload or damage. For example, higher damping may be required under aggressive off-road conditions, but this upper limit must still be adhered to. The minimum damping setting is the lower limit of the minimum damping force or damping coefficient that the suspension actuator can provide. Damping commands below this lower limit are also invalid or must be forcibly increased. For example, low damping is required for comfort when driving on a flat road, but it cannot be lower than the lowest point of the hardware's adjustable range. The minimum adjustable resolution is the smallest step size between adjacent damping values ​​that the suspension actuator can recognize and respond to. This value is determined by the actuator's control precision, such as the fineness of current adjustment. In the numerical quantization step, floating-point damping values ​​need to be rounded or truncated according to this resolution to generate integer commands, avoiding the issuance of indistinguishable minute changes. The response delay threshold is the shortest time required for the suspension actuator to complete the damping value adjustment from receiving the command, i.e., the actuator's response lag time. This threshold is used by the dynamic filter to limit the time gradient of the damping sequence: if the damping change between adjacent control cycles divided by the time interval exceeds the maximum rate of change corresponding to this threshold, the single-step change is forcibly clipped to the range that the actuator can complete within the delay time, thereby preventing the command issuance speed from exceeding the hardware response capability and avoiding system oscillation or command backlog.

[0049] Secondly, the floating-point damping values ​​in the final damping adjustment strategy sequence are sampled and aligned on the time axis, and then quantized according to the set maximum damping setpoint, minimum damping setpoint, and minimum adjustable resolution to obtain a quantized damping sequence synchronized with the electronic control unit's control cycle, such as 10 milliseconds or 20 milliseconds. Sampling alignment uses linear interpolation or nearest-neighbor interpolation methods to map irregular or high-resolution time points in the original sequence to the discrete control time points of the ECU; numerical quantization converts the floating-point damping values ​​in each control cycle into the closest integer multiple of the resolution value and truncates them within the maximum and minimum damping limits, thereby avoiding system oscillations caused by excessively frequent setpoint fluctuations.

[0050] Then, based on the response delay threshold, the time gradient of the quantized damping sequence is smoothed using a dynamic filter. The dynamic filter incorporates a moving average and edge suppression mechanism: the moving average algorithm performs local averaging on the damping sequence, and the edge suppression mechanism limits the maximum amplitude of a single-step change, ensuring that each setpoint change conforms to the actual response capability of the suspension system in both amplitude and direction, preventing execution lag or oscillation caused by command changes exceeding hardware response speed. The resulting smoothed damping sequence is then obtained.

[0051] Next, dynamic weighting factors are constructed based on vehicle speed, lateral acceleration, and steering angle information to perform scene-adaptive compensation adjustment on the smoothed damping sequence. The mapping function calls upon the vehicle's current speed, lateral acceleration, and steering angle information to construct dynamic weighting factors reflecting different operating conditions such as high-speed straight driving and low-speed turning. These factors are then used to compensate and adjust the damping setpoint, forming a damping control decision jointly driven by the vehicle's dynamic state and the damping strategy output. The compensated result must again be constrained within the range of the maximum and minimum damping setpoints to ensure that the final value does not exceed physical limits, thus obtaining the compensated damping setpoint.

[0052] Finally, the compensated damping setpoints are packaged into structured parameter packets according to four channels: front left, front right, rear left, and rear right. These packets are then sent to the electronic control unit (ECU) of the suspension controller via a CAN bus or high-speed Ethernet interface. Simultaneously, the ECU's internal strategy receiving module loads the drive parsing protocol that interfaces with the AI ​​model output. It performs channel-by-channel unpacking, timestamp verification, and pre-execution evaluation of the damping setpoints in the parameter packets to ensure that the instructions do not conflict with the currently executing control commands and do not cause physical anomalies. The mapping function also supports parameterized configuration structures, allowing for soft decoupling deployment based on the response models of different automakers and chassis system platforms. This ensures seamless migration of the AI ​​model across various vehicle models and controller interfaces. Through the above conversion and adaptation, a physically executable damping control command stream is ultimately generated. This stream is strictly synchronized with the actual hardware control cycle, filtered and dynamically adjusted, and has been interface-encapsulated before being output to the suspension actuator, thus achieving a closed-loop connection from intelligent recognition to intelligent control.

[0053] In one possible implementation, the model group trained offline can be deployed to the vehicle-side control architecture, and combined with real vehicle operation data for dynamic evaluation, parameter fine-tuning, and adaptive policy updates, thus constructing a closed-loop deployment system with online learning capabilities and self-evolving control strategies.

[0054] The deployment phase begins with quantization compression and structural pruning of the three sub-models: the dual-stream generative adversarial network, the Transformer variational autoencoder, and the policy gradient reinforcement learning. This ensures the models can run efficiently on the chassis controller or central computing platform. The compression process employs techniques such as model distillation, parameter sharing, and low-rank approximation to significantly reduce the model's computational resource consumption while preserving its original control performance to the maximum extent possible. Furthermore, to meet the cyclic requirements of vehicle chassis control, the model inference speed is stably controlled below 5 milliseconds, satisfying the core requirements of real-time vehicle control.

[0055] The deployment environment can be flexibly configured according to the actual situation of the automaker's electronic and electrical architecture. The system supports deploying the three sub-models on the chassis domain controller, high-performance central processing unit, or edge computing nodes respectively, adapting to the electronic and electrical architecture design requirements of different automakers. At the same time, a unified data interface module enables seamless integration with the vehicle's perception system, positioning system, power system, and execution system, ensuring real-time synchronization of data flow between the systems and reliable delivery of model strategy outputs to the actuators, guaranteeing timely response to control commands.

[0056] In the initial stage of model deployment, the system will enter the real-vehicle parameter tuning phase. This phase requires collecting vehicle operational data in various off-road to on-road transition scenarios, focusing on the response of the control strategy under different combinations of slope, grip, and road surface undulations, comprehensively capturing the system's operational status under complex real-world conditions. The system constructs a complete playback sample set by online recording of damping output sequences, vehicle dynamic response indicators, and driver subjective evaluation data. This sample set will be used for subsequent regression analysis of model performance, providing data support for parameter adjustment.

[0057] Building upon this foundation, the parameter tuning process utilizes a lightweight simulation module to conduct batch strategy playback evaluations. By setting a target weight function, multi-dimensional scoring is performed on three core dimensions: strategy response latency, transition smoothness, and execution error. This, combined with real-vehicle operating indicators, forms a comprehensive evaluation result. Based on the evaluation results, the system automatically adjusts the model fusion parameters, strategy weight allocation, and mapping function response thresholds, completing the first round of local optimization iterations to initially adapt the model parameters to the characteristics of real-vehicle operation.

[0058] After parameter tuning, the system officially enters the closed-loop optimization phase. A low-frequency strategy deviation evaluation module is built on the vehicle side. This module periodically compares the control strategy output by the model with the actual vehicle response data after execution, continuously monitoring the strategy's effectiveness. If the vehicle response continuously deviates from the expected behavior and exceeds a set threshold, an incremental learning mechanism is automatically triggered. Current operating condition data is collected as new training samples, and the policy network weights of the policy gradient reinforcement learning model are updated through online fine-tuning or edge learning modes. This allows the model to gradually adapt to changes in actual vehicle performance and complex road conditions, achieving adaptive optimization of the strategy.

[0059] Furthermore, this application introduces a version management mechanism, clearly marking each strategy optimization version and establishing a comprehensive rollback control function. When strategy degradation or model training instability occurs, version rollback can quickly restore to the previous stable version, effectively avoiding control failures caused by strategy anomalies, significantly improving the overall robustness and maintainability of the system, and ensuring long-term stable operation of the system.

[0060] This application possesses dynamic learning capabilities, strategy evolution capabilities, and self-closed-loop parameter tuning capabilities in real-world vehicle scenarios. It can continuously adjust itself based on vehicle operating status, hardware characteristics, and environmental changes, ensuring optimal damping control performance throughout long-term operation. This deployment and parameter tuning mechanism not only bridges the key gap between laboratory research and development and mass production vehicle deployment of artificial intelligence models, but also constructs a truly closed-loop self-optimizing intelligent chassis control system through real-world vehicle feedback. It represents one of the core technological paths for automotive-grade artificial intelligence control in the era of intelligent driving, possessing extremely high engineering transformation value and promising industrial application prospects.

[0061] In one possible implementation, the above-mentioned technical features are integrated to obtain the overall process, such as... Figure 2 As shown, Figure 2 A flowchart of another control method provided in the embodiments of this application includes: The first step is the multi-source data acquisition and unified encoding of key features for off-road to on-road state transition. The system collects vehicle sensor data (speed, acceleration, damping value, etc.), environmental perception data (camera, radar), map and V2X information. Through multi-source heterogeneous data acquisition, timestamp alignment, anomaly removal, unified encoding and feature packaging, it outputs the encoded unified state fragment, i.e., structured time window data.

[0062] The second step involves constructing a dual-stream generative adversarial network (GAN) and training it with damped policy generation. Using state fragment feature vectors and scene labels as input, the system trains a conditional generator, optimizes the discriminator's discriminative ability, designs an adversarial learning mechanism, and pre-trains multiple scene generation policies. The output is a damped policy generation sequence, i.e., continuously set values. The third step involves Transformer variational autoencoder (VAE) feature compression and scene latent variable modeling. Received from the state fragment sequence and scene context labels from the first step, the Transformer encoder extracts temporal-cross-modal features, performs latent space modeling, optimizes reconstruction error and KL divergence, and finally outputs a latent variable vector z as the scene encoding. It should be noted that steps two and three can be performed simultaneously, or their execution order can be set according to the actual situation.

[0063] The fourth step is policy gradient reinforcement learning (RL) for control path optimization. The inputs are latent variables z from the VAE and the generator policy from the GAN. An Actor-Critic structure based on policy gradients is used to design a reward function. Generation-policy contrastive learning and adaptive action output are then performed, resulting in an optimized damping control policy sequence.

[0064] The fifth step involves designing a three-model fusion strategy and a damping-coordinated control mechanism. This involves aligning the GAN output, VAE latent variables, and RL output strategies using features, comparing and judging strategies, dynamically weighting or selecting strategies, and combining this with control confidence assessment and priority switching mechanisms to output the fused damping control strategy.

[0065] The sixth step is the construction and ECU adaptation of the damping transition mapping function. Based on the fusion strategy output and the vehicle's current dynamic information, action quantization, frequency filtering, dynamic compensation, command packaging, and CAN interface encapsulation are performed, and aligned with the control cycle, ultimately outputting a sequence of control commands executable by the ECU. The seventh step is closed-loop optimization and online parameter tuning based on real vehicle data. After deployment on the real vehicle, operational data is collected, and model parameters are updated through deviation evaluation and incremental learning to achieve adaptive evolution of the control strategy.

[0066] The above are some specific implementations of the control method provided in the embodiments of this application. Based on this, this application also provides a corresponding control system. The system provided in the embodiments of this application will be described below from the perspective of functional modularity. Figure 3 This is a structural diagram of a control system provided in an embodiment of this application.

[0067] The system includes: Acquisition unit 110 is used to acquire multi-source data during the process of the vehicle switching from the first state to the second state, and to encode the multi-source data to obtain a state segment. The generation unit 111 is used to generate a sequence of candidate damping strategies based on the state segment using a generator of a two-stream generative adversarial network; and to perform feature extraction and reparameterization sampling on the state segment based on the encoder, and output a low-dimensional latent variable to characterize the vehicle's operating condition. The iteration unit 112 is used to take the low-dimensional latent variables and the candidate damping policy sequence as inputs, and perform iterative optimization through the policy gradient reinforcement learning model to obtain the optimal damping control path. The determining unit 113 is used to score the candidate damping strategy sequence and the optimal damping control path through the discriminant subnetwork, and determine the final damping adjustment strategy sequence based on the scoring results. Output unit 114 is used to convert the final damping adjustment strategy sequence into physical commands that meet the execution conditions of the suspension controller and output them to the suspension controller.

[0068] This application also provides corresponding devices and computer storage media for implementing the control scheme provided in this application.

[0069] The device includes a memory and a processor. The memory is used to store instructions or code, and the processor is used to execute the instructions or code to cause the device to perform the control method described in any embodiment of this application.

[0070] The computer storage medium stores code, and when the code is executed, the device running the code implements the control method described in any embodiment of this application.

[0071] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

[0072] It should be understood that in this application, "at least one" refers to one or more items, and "more" refers to two or more items. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one" or similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, "at least one" of a, b, or c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0073] It should be understood that the terms center, longitudinal, transverse, up, down, front, back, left, right, vertical, horizontal, top, bottom, inside, outside, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0074] It should be noted that, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0075] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the statement "including a…" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0076] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0077] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A control method characterized by, include: Acquire multi-source data during the process of the vehicle switching from the first state to the second state, and encode the multi-source data to obtain state segments; Based on the state segments, a sequence of candidate damping strategies is generated using a generator from a two-stream generative adversarial network; based on the encoder, feature extraction and reparameterization sampling are performed on the state segments to output low-dimensional latent variables that characterize the vehicle's operating conditions. Using the low-dimensional latent variables and the candidate damping policy sequence as inputs, the optimal damping control path is obtained by iterative optimization through a policy gradient reinforcement learning model. The candidate damping strategy sequence is scored against the optimal damping control path by a discriminant subnetwork, and the final damping adjustment strategy sequence is determined based on the scoring results. The final damping adjustment strategy sequence is converted into physical commands that meet the execution conditions of the suspension controller and output to the suspension controller.

2. The method of claim 1, wherein, Based on the state fragment, the generator of the dual-stream generative adversarial network generates a sequence of candidate damping strategies, including: The state fragment is input into the generator of a two-stream generative adversarial network. The generator is based on a multi-layer residual convolutional network and incorporates a scene label conditional embedding mechanism to extract features from the state fragment and obtain a high-dimensional feature representation. The high-dimensional feature representation is mapped to a continuous sequence of damping adjustment parameters through upsampling processing, forming a candidate damping strategy sequence that covers each control cycle during the vehicle state switching process.

3. The method of claim 2, wherein, The training methods for the dual-stream generative adversarial network include: The discriminator of a two-stream generative adversarial network is used to verify the authenticity of the candidate damping strategy sequence. The discriminator includes a fully connected neural network structure and a dropout layer. By comparing the consistency between the generated candidate damping strategy sequence and the real damping control trajectory, the parameters of the generator are adjusted in feedback until the deviation between the candidate damping strategy sequence output by the generator and the real damping control trajectory is within a preset threshold range.

4. The method of claim 1, wherein, The encoder includes a Transformer variational autoencoder. The encoder is used to extract features and reparameterize samples from the state segments, outputting low-dimensional latent variables characterizing the vehicle's operating conditions, including: The state segment is input into the Transformer variational autoencoder, and the global correlation features between multi-source data at different times are extracted using the multi-head self-attention mechanism to obtain a high-dimensional working condition feature vector. The high-dimensional working condition feature vector is input into the mean network and variance network of the variational autoencoder, respectively, and the mean parameter and variance parameter of the latent variable distribution are output. Reparameterized sampling is performed based on the mean and variance parameters to generate low-dimensional latent variables that follow a preset distribution.

5. The method of claim 1, wherein, The process of taking the low-dimensional latent variables and the candidate damping policy sequence as inputs, and iteratively optimizing them through a policy gradient reinforcement learning model to obtain the optimal damping control path includes: The low-dimensional latent variables are used as the state representation of the current vehicle operating condition, and the candidate damping strategy sequence is used as the reference trajectory. Both are input into the policy gradient reinforcement learning model. The policy gradient reinforcement learning model adopts a policy network-value network structure. The policy network is used to output the continuous damping adjustment action sequence, and the value network is used to estimate the expected cumulative reward of the current policy and guide the update of the policy network parameters. The policy gradient reinforcement learning model is trained based on the proximal policy optimization algorithm. The policy network parameters are iteratively updated until the model training converges, and the optimal damping control path that satisfies the vehicle attitude stability and comfort constraints is output.

6. The method of claim 1, wherein, The step of scoring the candidate damping strategy sequence against the optimal damping control path using a discriminant subnetwork includes: The candidate damping strategy sequence and the optimal damping control path are input into the discriminant subnetwork, which extracts feature parameters from them. Based on these feature parameters, the discriminant subnetwork compares and analyzes the differences between the candidate damping strategy sequence and the optimal damping control path in terms of control rationality, operating condition adaptability, and control smoothness. The feature parameters include damping adjustment amplitude, adjustment frequency, and time sequence continuity. Based on a preset scoring standard, the discriminant subnetwork outputs a score value corresponding to each of the candidate damping strategy sequence and the optimal damping control path. The score value is used to quantitatively characterize the degree of adaptability of the candidate damping strategy sequence and the optimal damping control path to the vehicle suspension control requirements.

7. The method of claim 1, wherein, The determination of the final damping adjustment strategy sequence based on the scoring results includes: Based on the scoring results, the candidate damping strategy sequence and the optimal damping control path are dynamically weighted and combined to obtain the final damping adjustment strategy sequence, or the one with the higher score in the scoring results is selected as the final damping adjustment strategy sequence.

8. The method of claim 1, wherein, The step of converting the final damping adjustment strategy sequence into physical commands that meet the execution conditions of the suspension controller and outputting them to the suspension controller includes: Set the maximum damping setting, minimum damping setting, minimum adjustable resolution, and response delay threshold for each channel according to the vehicle platform and suspension system type. The floating-point damping values ​​in the final damping adjustment strategy sequence are sampled and aligned on the time axis, and numerical quantization is performed based on the maximum damping setpoint, minimum damping setpoint, and minimum adjustable resolution to obtain a quantized damping sequence synchronized with the control cycle of the electronic control unit. Based on the response delay threshold, the time gradient of the quantized damping sequence is smoothed by a dynamic filter to obtain a smoothed damping sequence; a dynamic weighting factor is constructed based on vehicle speed, lateral acceleration and steering angle information to perform scene adaptive compensation adjustment on the smoothed damping sequence, and the compensation result is constrained within the range of the maximum damping set value and the minimum damping set value to obtain the compensated damping set value. The compensated damping setpoint is packaged into a structured parameter package and sent to the electronic control unit of the suspension controller via the controller area network bus or high-speed Ethernet interface.

9. An electronic device, characterized in that, include: Processor, memory, system bus; The processor and the memory are connected via the system bus; The memory is used to store one or more programs, the one or more programs including instructions that, when executed by the processor, cause the processor to perform the control method according to any one of claims 1-8.

10. A vehicle, characterized in that, Including an electronic device as described in claim 9.

Images