Dual-mode parallel cooperative control system and control method of intelligent chassis

Abstract

This invention discloses a dual-modal parallel cooperative control system and control method for an intelligent chassis, belonging to the field of vehicle control technology. The dual-modal parallel cooperative control system for the intelligent chassis includes: a rule-based first control module, a learning-based second control module, a calibration module, a scene perception module, and an arbitration module. The rule-based first control module calculates a first control signal based on the vehicle's state information; the learning-based second control module calculates a second control signal and a first confidence signal based on the vehicle's state information; the calibration module outputs a calibrated second confidence signal; the scene perception module outputs a scene classification result based on the vehicle's state information; and the arbitration module determines a fusion weight, performs weighted fusion of the first and second control signals according to the fusion weight, and outputs a fused control signal for controlling the vehicle, thereby improving the safety and stability of driving under various operating conditions.

Description

Technical Field

[0001] This invention relates to the field of vehicle control technology, and in particular to a dual-modal parallel cooperative control system and control method for intelligent chassis. Background Technology

[0002] With the deepening of the wave of automotive intelligence, the intelligent chassis control system, as the core carrier of vehicle motion execution, is facing severe challenges brought about by advanced driver assistance and diversified user scenarios.

[0003] In related technologies, the current technical routes for multi-actuator collaborative control of intelligent chassis are mainly divided into rule-based collaborative control schemes and end-to-end learning-based collaborative control schemes. The choice between rule-based collaborative control schemes and end-to-end learning-based collaborative control schemes is generally one of the two, resulting in relatively limited and one-sided control decisions. Summary of the Invention

[0004] This invention aims to at least solve the technical problems existing in related technologies. To this end, this invention proposes a dual-modal parallel cooperative control system for an intelligent chassis. Through a dynamic weight fusion mechanism, it can fully leverage the performance advantages of the learning-based second control module under normal operating conditions, significantly improving control accuracy, while ensuring driving safety under boundary conditions through a rule-based first control module.

[0005] The present invention also proposes a control method.

[0006] The first aspect of the present invention discloses a dual-modal parallel cooperative control system for an intelligent chassis, applied to a vehicle with an intelligent chassis. The dual-modal parallel cooperative control system includes: a rule-based first control module configured to calculate a first control signal based on vehicle state information; a learning-based second control module configured to calculate a second control signal and a first confidence signal regarding the second control signal based on vehicle state information; a calibration module configured to calibrate the first confidence signal and output the calibrated second confidence signal; a scene perception module configured to output a scene classification result of the current driving scene based on vehicle state information; and an arbitration module configured to determine a fusion weight based on the first control signal, the second control signal, the scene classification result, and the second confidence signal, perform weighted fusion of the first control signal and the second control signal based on the fusion weight, and output a fused control signal for controlling the vehicle.

[0007] The dual-modal parallel cooperative control system for an intelligent chassis according to embodiments of the present invention has at least the following beneficial effects: The dual-modal parallel cooperative control system for the intelligent chassis has a rule-based first control module and a learning-based second control module. The rule-based first control module is configured to calculate a first control signal based on vehicle state information. The learning-based second control module is configured to calculate a second control signal and a first confidence signal regarding the second control signal based on vehicle state information. A calibration module is configured to calibrate the first confidence signal and output the calibrated second confidence signal to reduce the risk caused by overconfidence in the learning-based second control module. A scene perception module is configured to output a scene classification result of the current driving scene based on vehicle state information. An arbitration module is configured to arbitrate based on the first control signal, the second control signal, and the scene classification result. The fusion weights are determined by the class result and the second confidence signal. The first control signal and the second control signal are weighted and fused according to the fusion weights, and a fused control signal for controlling the vehicle is output. When the scene classification result determined by the scene perception module belongs to the normal working condition, the arbitration module can appropriately increase the weight of the second control signal and decrease the weight of the first control signal to give full play to the performance advantages of the learning-based second control module and improve the control accuracy. When the scene classification result determined by the scene perception module belongs to the boundary working condition, the arbitration module can appropriately decrease the weight of the second control signal and increase the weight of the first control signal to ensure driving safety through the rule-based first control module. That is, the dual-modal parallel cooperative control system of this intelligent chassis can combine the advantages of the rule-based first control module and the learning-based second control module to improve the safety and stability of the vehicle under various working conditions.

[0008] According to some embodiments of the present invention, the dual-modal parallel cooperative control system of the intelligent chassis further includes a startup management module; the startup management module is configured to control the vehicle only through a rule-based first control module during the vehicle startup phase, and to control the vehicle with a fused control signal that is a weighted fusion of the first control signal and the second control signal after the vehicle startup phase has passed. And / or, the dual-modal parallel cooperative control system of the intelligent chassis also includes a training module, which is configured to collect a first control signal, the corresponding vehicle state, the scene classification result and the second control signal, and train and update the learning-based second control module based on the collected first control signal, the corresponding vehicle state, the scene classification result and the second control signal.

[0009] The control method of the second aspect of the present invention is applied to the dual-modal parallel cooperative control system of the intelligent chassis of the first aspect, and the control method includes: The vehicle's status information is acquired and a first control signal is calculated based on a rule-based first control module. The first control signal has a safety boundary. The vehicle's status information is acquired, and a second control signal and a first confidence signal regarding the second control signal are calculated based on the learning-based second control module. The first confidence signal is calibrated, and the calibrated second confidence signal is output. Obtain vehicle status information and output the scene classification result of the current driving scenario; The fusion weights are determined based on the first control signal, the second control signal, the scene classification result, and the second confidence signal. The first control signal and the second control signal are then weighted and fused according to the fusion weights, and a fused control signal for controlling the vehicle is output.

[0010] According to some embodiments of the present invention, a weighted fusion of a first control signal and a second control signal is performed based on a fusion weight, and a fused control signal for controlling a vehicle is output, including: The product of the first control signal and the fusion weight is used as the first calculated value, and the product of the difference between 1 and the fusion weight and the second control signal is used as the second calculated value. The fusion control signal is the sum of the first calculated value and the second calculated value.

[0011] According to some embodiments of the present invention, acquiring vehicle state information and calculating a second control signal based on a learning-based second control module, and a first confidence signal regarding the second control signal, includes: The system acquires environmental information and uses a deep neural network to output a second control signal and a first confidence signal based on the environmental and state information.

[0012] According to some embodiments of the present invention, the security boundary includes a soft security boundary and a hard security boundary; determining the fusion weight based on a first control signal, a second control signal, a scene classification result, and a second confidence signal includes: Construct a correspondence between basic weights and scene classification results; Based on the scene classification results and corresponding relationships output by the scene perception module, the corresponding basic weights are output; When the second control signal is within the soft security boundary and the hard security boundary, the basic weight is used as the fusion weight.

[0013] According to some embodiments of the present invention, a fusion weight is determined based on a first control signal, a second control signal, a scene classification result, and a second confidence signal; the first control signal and the second control signal are weighted and fused according to the fusion weight; and a fused control signal for controlling the vehicle is output, including: When the second control signal exceeds the hard safety boundary, the fusion weight is set to 1, and the first control signal is used as the fusion control signal to control the vehicle. Alternatively, if the second control signal exceeds the soft safety boundary but not the hard safety boundary, the base weight is adjusted according to the value of the excess part and the fusion weight is output. If the value of the excess part is less than or equal to the preset value, the first control signal and the second control signal are weighted and fused using the fusion weight, and the fused control signal is output. Alternatively, if the second control signal exceeds the soft safety boundary but not the hard safety boundary, and the value of the exceeding part is greater than the preset value, the fusion weight is set to 1, and the first control signal is used as the fusion control signal to control the vehicle.

[0014] According to some embodiments of the present invention, the control method further includes: Arbitrate the fused control signals based on vehicle dynamics coupling constraints; If the fused control signal violates the vehicle dynamics coupling constraint, the fused control signal is corrected according to the vehicle dynamics coupling constraint.

[0015] According to some embodiments of the present invention, the control method further includes: When the vehicle is in the vehicle start-up phase, the first control module based on rules operates, sets the fusion weight to 1, and uses the first control signal as the fusion control signal to control the vehicle. The second control module, based on learning, is controlled to operate and sends a ready signal, a second control signal, and a first confidence signal to the arbitration module. When the arbitration module receives the ready signal, it determines that the vehicle has passed the vehicle start-up phase and recalculates the fusion weight as the target value based on the second control signal and the first confidence signal. Within a preset number of control cycles, the fusion weight of the vehicle is gradually reduced from 1 to the target value.

[0016] According to some embodiments of the present invention, the control method further includes: Collect the first control signal, the corresponding vehicle status, the scene classification result, and the second control signal; The second control module is trained and updated based on the first control signal, the corresponding vehicle state, the scene classification result, and the second control signal.

[0017] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments, wherein: Figure 1 This is a system architecture block diagram of a dual-modal parallel cooperative control system for an intelligent chassis according to an embodiment of the present invention; Figure 2This is a flowchart of the punching module of a dual-modal parallel collaborative control system for an intelligent chassis according to an embodiment of the present invention. Figure 3 This is a flowchart of the startup management module of a dual-modal parallel collaborative control system for an intelligent chassis according to an embodiment of the present invention; Figure 4 This is a flowchart of a control method according to an embodiment of the present invention; Figure 5 This is a flowchart illustrating the output of a second control signal and a first confidence signal, and the calculation of fusion weights, according to an embodiment of the control method of the present invention. Figure 6 This is a flowchart illustrating the calculation of fusion weights based on basic weights in a control method according to an embodiment of the present invention. Figure 7 This is a flowchart illustrating a control method according to an embodiment of the present invention for controlling a vehicle based on a comparison of a second control signal with soft and hard safety boundaries; Figure 8 This is a flowchart illustrating a control method according to an embodiment of the present invention for controlling a vehicle based on a comparison of a fused control signal and a hard safety boundary. Figure 9 This is a flowchart illustrating the startup management of a control method according to an embodiment of the present invention; Figure 10 This is a flowchart illustrating the training and updating of a control method according to an embodiment of the present invention. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0020] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, etc., are based on the orientation or positional relationship shown in the drawings and are only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0021] In the description of this invention, "multiple" refers to two or more. The use of "first" and "second" is for distinguishing technical features only and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features or their sequential relationship.

[0022] In the description of this invention, unless otherwise explicitly defined, terms such as "set up," "install," and "connect" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0023] With the deepening of the automotive intelligence trend, the intelligent chassis control system, as the core carrier of vehicle motion execution, is facing severe challenges brought about by advanced driver assistance systems and diversified user scenarios. The chassis domain controller needs to coordinate multiple heterogeneous actuators such as steering, braking, drive, and suspension to complete complex dynamic control within milliseconds, while simultaneously meeting the functional safety ASIL-D (Automotive Safety Integrity Level D) requirements. Currently, the technical approaches to multi-actuator collaborative control of intelligent chassis mainly fall into two categories: The first category is rule-based collaborative control schemes. This type of scheme combines expert experience with physical models, using algorithms such as lookup tables, PID (Proportional-Integral-Derivative) control, and model predictive control (MPC) to calculate control commands for each actuator based on vehicle state information (vehicle speed, yaw rate, sideslip angle, etc.). Its advantages include high determinism, strong interpretability, and ease of functional safety certification; its disadvantages include difficulty in covering all edge cases, a large workload for parameter calibration, and limited performance when dealing with unmodeled dynamics.

[0024] The second category is collaborative control schemes based on end-to-end learning. This type of scheme uses deep neural networks or reinforcement learning to directly map sensor inputs to chassis actuator outputs, achieving integrated "perception-decision-control." Its advantages include the ability to learn complex nonlinear mappings from data, achieving performance superior to rule-based control under normal operating conditions. Its disadvantages include a lack of interpretability, uncertainty under out-of-distribution (OOD) samples, and difficulty in passing functional safety certification independently.

[0025] Reference Figures 1 to 3 As shown, an embodiment of the present invention discloses a dual-modal parallel cooperative control system for an intelligent chassis, applied to a vehicle equipped with an intelligent chassis. The intelligent chassis is a system capable of recognizing, predicting, and controlling the interaction between the wheels and the ground, and managing its own operating state, specifically realizing the intelligent driving task of the vehicle. Specifically, the dual-modal parallel cooperative control system for this intelligent chassis includes a rule-based first control module, a learning-based second control module, a scene perception module, a calibration module, and an arbitration module.

[0026] Reference Figure 1 , Figure 2 and Figure 3As shown, the rule-based safety control module employs deterministic algorithms (PID control, MPC, etc.) to calculate the first control signal u_rule in real time based on the vehicle's state information. The first control signal has a clear and dynamically adjustable safety boundary.

[0027] Reference Figure 1 , Figure 2 and Figure 3 As shown, the core of this rule-based safety margin control module is to provide a first control signal with deterministic and verifiable safety boundaries. Specifically, the rule-based first control module can employ a model predictive control (MPC) algorithm.

[0028] Reference Figure 1 , Figure 2 and Figure 3 As shown, the output of the rule-based first control module is the first control signal u_rule, which contains the expected values ​​of each actuator in the smart chassis (such as steering wheel angle, braking torque of each wheel, drive torque, suspension damping force, etc.).

[0029] Reference Figure 1 , Figure 2 and Figure 3 As shown, the learning-based second control module runs in parallel with the rule-based first control module. It calculates the second control signal u_learn in real time based on the vehicle's state information and simultaneously outputs the first confidence signal c_raw of the second control signal. The first confidence signal is used to characterize the credibility of the second control signal. If the first confidence signal is high, it indicates that the credibility of the second control signal is high. If the first confidence signal is low, it indicates that the credibility of the second control signal is low.

[0030] Reference Figure 1 , Figure 2 and Figure 3 As shown, the learning-based second control module uses a deep neural network. The inputs are vehicle state information (vehicle speed, yaw rate, center of gravity sideslip angle, steering wheel angle, wheel speed, etc.) and environmental information (such as optional perception inputs, including lane line curvature, relative position and speed of target obstacles, etc.). The output is the second control signal u_learn, and the first confidence signal c_raw is also output, where c_raw∈[0,1].

[0031] In related technologies, existing arbitration mechanisms primarily rely on external scenarios or preset rules for decision-making, failing to fully utilize the uncertainty information output by the learning-based second control module itself. Research shows that when neural networks encounter out-of-distribution (OOD) samples, their output confidence significantly decreases. This internal information, reflecting the model's "self-awareness," could serve as a crucial basis for arbitration. After in-depth consideration, the inventors concluded that automatically increasing the weight of the first control signal when the learning-based second control module is "uncertain" would greatly enhance the system's security in unknown scenarios, but existing solutions have not utilized this approach.

[0032] Reference Figure 1 , Figure 2 and Figure 3 As shown, the dual-modal parallel cooperative control system of the intelligent chassis provided in this embodiment of the invention calculates a first confidence signal c_raw, similar to the confidence level of the estimation result in a traditional state estimator. When the vehicle is driving under normal operating conditions with good training data coverage, the model output variance is small, and the first confidence signal is high; when the vehicle encounters an OOD sample (outside the model's distribution), the model's extrapolation ability decreases, the output variance increases, and the first confidence signal is low. The arbitration module utilizes this information to rely more on the rule-based first control module when the learning-based second control module has high uncertainty, in order to ensure the safety and stability of vehicle driving.

[0033] Reference Figure 1 , Figure 2 and Figure 3 As shown, the first confidence signal can be calculated using the Monte Carlo dropout method: during forward inference, some neurons are randomly dropped, and this is repeated N times (N=20~50). The mean μ and variance σ² of the N outputs are calculated. The first confidence signal c_raw = 1 / (1 + σ²). The larger the variance, the lower the first confidence signal.

[0034] Reference Figure 1 , Figure 2 and Figure 3 As shown, the calibration module performs post-processing calibration on the first confidence signal output by the learning-based second control module, outputting a calibrated second confidence signal c for use by the arbitration module. The scene perception module outputs the scene classification result (scene) of the current operating scene in real time based on vehicle status information.

[0035] Reference Figure 1 , Figure 2 and Figure 3As shown, the dual-modal parallel cooperative control system for an intelligent chassis provided in this embodiment of the invention allows the learning-based second control module to output a first confidence signal, which is then processed by a calibration module to obtain a second confidence signal for arbitration. This enables the system to perceive the uncertainty of the learning-based second control module in real time. When the learning-based second control module is "uncertain" (low confidence), the arbitration module automatically increases the weight of the first control signal to enhance safety. By introducing and calibrating the first confidence signal, the system can automatically tilt towards the rule-based first control module when encountering scenarios outside the training distribution, based on the "uncertainty" of the model itself. Simulation verification shows that in scenarios where the confidence of the learning-based second control module decreases, by dynamically adjusting the weights, some contributions from the learning-based second control module are retained while ensuring safety, achieving a dynamic balance between safety and performance.

[0036] In related technologies, some existing solutions employ a "two-choice" discrete selection mechanism, which, in principle, cannot achieve the complementary advantages of rules and learning, potentially leading to abrupt changes in control signals or performance loss. Some existing solutions use a time-sharing switching mode, requiring algorithm initialization and state synchronization during switching, which easily results in discontinuities in control output and struggles to meet real-time coordination requirements under sudden operating conditions. Neither of these existing solutions achieves real-time fusion of the safety of rules and the performance advantages of learning within each control cycle.

[0037] Reference Figure 1 , Figure 2 and Figure 3 As shown, the arbitration module of the dual-modal parallel collaborative control system of the intelligent chassis provided in this embodiment of the invention is configured to determine the fusion weight based on the first control signal, the second control signal, the scene classification result, and the second confidence signal, to perform weighted fusion of the first control signal and the second control signal based on the fusion weight, and to output a fusion control signal for controlling the vehicle.

[0038] Reference Figure 1 , Figure 2 and Figure 3As shown, when the scene classification result determined by the scene perception module belongs to the normal operating condition, the arbitration module can appropriately increase the weight of the second control signal and decrease the weight of the first control signal when performing weighted fusion of the first and second control signals. This fully leverages the performance advantages of the learning-based second control module and improves control accuracy. Conversely, when the scene classification result determined by the scene perception module belongs to the boundary operating condition, the arbitration module can appropriately decrease the weight of the second control signal and increase the weight of the first control signal when performing weighted fusion of the first and second control signals. This ensures driving safety through the rule-based first control module. In other words, the dual-modal parallel collaborative control system of this intelligent chassis can fully utilize the advantages of the rule-based first control module and the learning-based second control module to improve the safety and stability of the vehicle under various operating conditions.

[0039] Reference Figure 1 , Figure 2 and Figure 3 As shown, it is understandable that the calibration module performs post-processing calibration on the first confidence signal of the learning-based second control module, which often exhibits calibration bias (such as overconfidence). Specifically, the calibration module of the dual-modal parallel cooperative control system of this intelligent chassis can employ confidence calibration suitable for regression tasks: Isotonic Regression. On the validation set, using the first confidence signal c_raw as input and a measure of the model prediction error (such as the negative value of the absolute error) as the supervision signal, a monotonically increasing calibration function f is learned. In online applications, c = f(c_raw).

[0040] Reference Figure 1 , Figure 2 and Figure 3 As shown, the dual-modal parallel collaborative control system of the intelligent chassis should periodically (e.g., during each OTA model update) re-optimize the calibration function using the latest validation set to adapt to changes in data distribution. This dual-modal parallel collaborative control system of the intelligent chassis employs confidence calibration suitable for regression tasks due to its stronger ability to fit nonlinear calibration relationships. The calibrated second confidence signal c more accurately reflects the actual accuracy of the model and is used by the arbitration module.

[0041] It should be understood that, in some other embodiments, specifically, the calibration module of the dual-modal parallel collaborative control system of the smart chassis may employ confidence calibration suitable for classification tasks: temperature scaling.

[0042] If the output of the learning-based second control module is a discrete scene category or decision category (such as "turn left", "go straight", "turn right"), its output layer is usually a softmax function, outputting the probability of each category. In this case, temperature scaling can be used for calibration: p_calibrated = softmax( (logit) / T ); Here, logit is the input vector before the Softmax layer, and T is the temperature parameter (>0). T is obtained by minimizing the negative log-likelihood (NLL) or expected calibration error (ECE) on the validation set. The calibrated probability p_calibrated more accurately reflects the actual frequency of the model's predictions being true.

[0043] It should be understood that, in some other embodiments, specifically, the calibration module of the dual-modal parallel collaborative control system of the smart chassis may employ variance-based parametric calibration.

[0044] Assuming the model prediction error follows a certain distribution (such as the Laplace distribution), a linear or nonlinear mapping can be learned to map the output variance σ² to the calibrated confidence level, for example, c = exp(-α·σ²), where α is the parameter to be learned.

[0045] Reference Figure 1 , Figure 2 and Figure 3 As shown, it can be understood that in this embodiment, the scene perception module outputs the scene classification result of the current driving scene in real time based on the vehicle state information. The scene perception module adopts a lightweight decision tree model, and the input parameters include: vehicle speed v, steering wheel angle δ and its rate of change dδ / dt, yaw rate ω and its deviation from the expected value Δω, longitudinal acceleration ax, lateral acceleration ay, and road adhesion coefficient estimate μ.

[0046] Reference Figure 1 , Figure 2 and Figure 3 As shown, the scene classification results are scene ∈ {high-speed cruising, urban congestion, emergency obstacle avoidance, low-friction road surface, extreme conditions, startup phase}. The correspondence between scene classification results and basic weights is as follows:

[0047] Comparison table of scene classification results and basic weights Reference Figure 1 , Figure 2 and Figure 3As shown, it is understandable that after receiving the first control signal, the second control signal, the scene classification result, and the second confidence signal, the arbitration module can calculate the basic weight based on the scene classification result. By querying the lookup table between the scene classification result and the basic weight, the basic weight w_base is obtained. The lookup table between the scene classification result and the basic weight can be based on statically preset domain expert knowledge, defining the basic tendencies for safety and performance under different operating conditions.

[0048] Reference Figure 1 , Figure 2 and Figure 3 As shown, it is understandable that in related technologies, existing solutions do not clearly address the control strategy problem during the loading delay of the learning-based second control module, which may result in the vehicle not being able to immediately achieve the best control effect after startup.

[0049] Reference Figure 1 , Figure 2 and Figure 3 As shown, the dual-modal parallel cooperative control system of the intelligent chassis also includes a start-up management module. The start-up management module is configured to control the vehicle with a first control signal during the vehicle start-up phase, and to control the vehicle with a fusion control signal that is a weighted fusion of the first and second control signals after the vehicle start-up phase has passed. The startup management module is responsible for coordinating the various modules during the vehicle startup phase, enabling a smooth connection. The process is as follows: Figure 3 As shown.

[0050] Reference Figure 1 , Figure 2 and Figure 3 As shown, considering the high real-time requirements of chassis control, the startup management module of the dual-modal parallel collaborative control system of the intelligent chassis provided in this embodiment of the invention preferably adopts a lightweight or medium-sized model architecture and uses UFS 3.0 storage to reduce loading latency.

[0051] Specific procedures: The moment the vehicle is powered on, the startup management module immediately activates the rule-based first control module, ensuring that it outputs the first control signal within 5ms, and the vehicle can start moving immediately.

[0052] Meanwhile, the loading process of the learning-based second control module is asynchronously started in the background: reading model parameters from flash memory, initializing the inference engine, and completing the first inference warm-up.

[0053] After the learning-based second control module is loaded, it sends a ready signal to the arbitration module and outputs the first second control signal and its first confidence signal.

[0054] Reference Figure 1 , Figure 2 and Figure 3As shown, after receiving the ready signal from the learning-based second control module, the arbitration module no longer forces the scenario to the "startup phase," but instead calculates the fusion weight w_target as the target value based on the actual scenario. To avoid abrupt control changes, a smooth transition mechanism is introduced: within M control cycles (e.g., M=10, each cycle 10ms), the weight is linearly decreased from 1.0 to the target value; after the transition is complete, it enters normal operation mode.

[0055] Reference Figure 1 , Figure 2 and Figure 3 As shown, the startup management module provided in this embodiment of the invention is designed with a phased startup strategy. Before the model is loaded, the rule-based first control module completely dominates to ensure that the vehicle can start and drive immediately. After the learning-based second control module is ready, its output is introduced through a smooth transition mechanism (gradual weight change) to achieve seamless access and avoid control abrupt changes. This ensures both immediate availability and avoids control abrupt changes, thereby improving the user experience.

[0056] It should be understood that, in some other embodiments, the startup management module may be designed with a two-level model architecture, preloading a lightweight, fast-start model (parameters < 1M, loading time < 20ms) to provide basic performance during the loading of the full model.

[0057] Understandably, in this embodiment, while existing solutions in related technologies collect rule data for training, this process is unidirectional (rules → data → model) and does not form a closed loop for continuous optimization. The high-quality security data generated by the rule-based first control module during operation is not systematically used for the continuous evolution of the model.

[0058] Reference Figure 1 , Figure 2 and Figure 3 As shown, the dual-modal parallel collaborative control system of the intelligent chassis provided in this embodiment of the invention also includes a training module. The training module is configured to collect a first control signal, the corresponding vehicle state, the scene classification result, and a second control signal, and to train and update the learning-based second control module based on the collected first control signal, the corresponding vehicle state, the scene classification result, and the second control signal.

[0059] It should be understood that in some other embodiments, the training module can be updated periodically in the cloud: when the vehicle is parked and charging, it uploads rule data to the cloud via Wi-Fi, and after the cloud completes the training, it periodically issues new models via OTA.

[0060] Reference Figures 4 to 10 As shown, the control methods of some embodiments of the present invention are applied to a dual-modal parallel cooperative control system for an intelligent chassis as illustrated in any of the above embodiments. The control method includes the following steps: Step S100: Obtain vehicle status information and calculate a first control signal based on the rule-based first control module. The first control signal has a safety boundary. Step S200: Obtain vehicle status information and calculate a second control signal and a first confidence signal about the second control signal based on the learning-based second control module; Step S300: The first confidence signal is calibrated, and the calibrated second confidence signal is output. Step S400: Obtain the vehicle's status information and output the scene classification result of the current driving scenario; Step S500: Determine the fusion weight based on the first control signal, the second control signal, the scene classification result, and the second confidence signal; perform weighted fusion of the first control signal and the second control signal based on the fusion weight; and output the fused control signal for controlling the vehicle.

[0061] Reference Figure 1 , Figure 2 and Figure 4 As shown, in step S100, the rule-based safety control module can use a deterministic algorithm (PID control, MPC, etc.) to calculate the first control signal u_rule in real time based on the vehicle's state information. The first control signal has a clear and dynamically adjustable safety boundary.

[0062] Reference Figure 1 , Figure 2 and Figure 4 As shown, the core of this rule-based safety margin control module is to provide a first control signal with deterministic and verifiable safety boundaries. Specifically, the rule-based first control module can employ a model predictive control (MPC) algorithm, whose optimization problem can be formulated as: min J = Σ(y_ref - y_pred)^TQ (y_ref - y_pred) + Σu^TR u; st x(k+1) = A x(k) + B u(k) / / Vehicle dynamics model; y(k) = C x(k) / / Output equation; u_min(μ) ≤ u(k) ≤ u_max(μ) / / Actuator physical constraints (hard safety boundary); y_min(μ) ≤ y(k) ≤ y_max(μ) / / Vehicle state safety boundary (hard safety boundary); In the first formula, the first term is the deviation between the penalized predicted output y_ref and the reference trajectory y_pred, where Q is a weight matrix that determines the degree of importance attached to the deviation between the penalized predicted output and the reference trajectory; the second term is the magnitude of the penalized control quantity u, where R is the weight matrix of the control quantity, used to save energy and reduce actuator wear, and avoid drastic, high-frequency operations.

[0063] In the second formula, x(k) is the current state (such as sideslip angle, yaw rate, wheel speed, etc.), u(k) is the control input (such as braking pressure, front wheel angle), and A and B together determine the next state x(k+1). C maps the state to the output y(k) (such as lateral acceleration, yaw rate, etc.).

[0064] The rule-based first control module predicts the output at several future time points based on MPC, and then finds the optimal u. The more accurate the model, the more reliable the prediction.

[0065] The third and fourth formulas represent constraints on hard safety boundaries, specifically physical constraints on the actuators: for example, braking force cannot exceed physical limits, and steering motor speed cannot be infinitely high. Here, μ represents the road adhesion coefficient (dry / wet / icy / snowy roads). The limit values ​​will vary with μ; for example, lower μ results in lower maximum braking force. Vehicle state safety boundaries include, for example, the sideslip angle cannot exceed a certain value (otherwise, fishtailing and loss of control), and the yaw rate cannot exceed the adhesion ellipse boundary. Similarly, μ determines the size of these boundaries.

[0066] Among them, the hard safety boundary y_min(μ) ≤ y(k) ≤ y_max(μ) is determined by functional safety analysis and vehicle dynamics theory. It represents the physical limit for stable vehicle operation and cannot be exceeded. The hard safety boundary is dynamically changing and needs to be adaptively adjusted based on the real-time estimated road adhesion coefficient μ. For example, the upper limit of yaw rate ω_max = f(μ, v) and the upper limit of the center of gravity sideslip angle also need to be dynamically changed with μ according to the phase plane method. Their functional relationship is jointly determined by the vehicle dynamics model and a large amount of real vehicle calibration test data.

[0067] Reference Figure 1 , Figure 2 and Figure 4 As shown, the output of the rule-based first control module is the first control signal u_rule, which contains the expected values ​​of each actuator in the smart chassis (such as steering wheel angle, braking torque of each wheel, drive torque, suspension damping force, etc.).

[0068] Reference Figure 1 , Figure 2 and Figure 4As shown, in step S200, the learning-based second control module and the rule-based first control module run in parallel. The second control signal u_learn is calculated in real time according to the vehicle's state information, and the first confidence signal c_raw of the second control signal is output at the same time. The first confidence signal is used to characterize the credibility of the second control signal. If the first confidence signal is high, it indicates that the credibility of the second control signal is high. If the first confidence signal is low, it indicates that the credibility of the second control signal is low.

[0069] Reference Figure 1 , Figure 2 and Figure 4 As shown, in steps S300 and S400, the control method can perform post-processing calibration on the first confidence signal output by the learning-based second control module through the calibration module, and output the calibrated second confidence signal c for use by the arbitration module. The control method can also output the scene classification result (scene) of the current operating scene in real time based on the vehicle status information through the scene perception module.

[0070] Reference Figure 1 , Figure 2 and Figure 4 As shown, the control method provided in this embodiment of the invention allows the learning-based second control module to output a first confidence signal, which is then processed by a calibration module to obtain a second confidence signal for arbitration. This enables real-time perception of the uncertainty of the learning-based second control module. When the learning-based second control module is "uncertain" (low confidence), the control method automatically increases the weight of the first control signal to enhance safety. By introducing and calibrating the first confidence signal, the system can automatically tilt towards the rule-based first control module when encountering scenarios outside the training distribution, based on the model's own "uncertainty." Simulation verification shows that in scenarios where the confidence of the learning-based second control module decreases, by dynamically adjusting the weights, some of the contribution of the learning-based second control module is retained while ensuring safety, achieving a dynamic balance between safety and performance.

[0071] Reference Figure 1 , Figure 2 and Figure 4 As shown, in step S500, the control method determines the fusion weights based on the first control signal, the second control signal, the scene classification result, and the second confidence signal through the arbitration module, performs weighted fusion of the first control signal and the second control signal based on the fusion weights, and outputs a fusion control signal for controlling the vehicle.

[0072] Reference Figure 1 , Figure 2 and Figure 4As shown, when the scene classification result determined by the scene perception module belongs to a normal operating condition, the arbitration module can appropriately increase the weight of the second control signal and decrease the weight of the first control signal when performing weighted fusion of the first and second control signals. This fully leverages the performance advantages of the learning-based second control module and improves control accuracy. Conversely, when the scene classification result determined by the scene perception module belongs to a boundary operating condition, the arbitration module can appropriately decrease the weight of the second control signal and increase the weight of the first control signal when performing weighted fusion of the first and second control signals. This ensures driving safety through the rule-based first control module. In other words, this control method can fully utilize the advantages of both the rule-based first control module and the learning-based second control module to improve the safety and stability of the vehicle under various operating conditions.

[0073] Reference Figure 1 , Figure 2 and Figure 4 As shown, the control method provided in this embodiment of the invention introduces a dynamic weighted fusion mechanism, enabling the first and second control signals to be continuously fused and output according to their weights. This control method, while ensuring safety boundaries, allows the performance advantage of the learning-based second control module to continuously contribute in each control cycle, fundamentally overcoming the limitations of "discrete selection" or "time-sharing switching" in existing schemes. Compared with the "two-choice" mechanism of existing schemes, the control method provided in this embodiment of the invention retains part of the performance contribution of the learning-based second control module through weighted fusion, achieving higher control accuracy under normal operating conditions. Compared with the "time-sharing switching" mechanism of existing schemes, the control method provided in this embodiment of the invention eliminates switching transients in principle through parallel operation and real-time fusion, achieving smooth control.

[0074] It should be understood that in some other embodiments, the rule-based first control module is not limited to MPC, and may also employ other control algorithms with deterministic output and verifiable safety boundaries, such as linear quadratic regulators (LQR) or feedforward and feedback control based on vehicle dynamics models for coordinated control.

[0075] Reference Figure 1 , Figure 2 and Figure 5 As shown, it can be understood that in step S500, the control method performs weighted fusion of the first control signal and the second control signal according to the fusion weight, and outputs a fused control signal for controlling the vehicle, including the following steps: Step S510: The product of the first control signal and the fusion weight is used as the first calculated value, and the product of the difference between 1 and the fusion weight and the second control signal is used as the second calculated value. The fusion control signal is the sum of the first calculated value and the second calculated value.

[0076] The formula for calculating the fused control signal is: u_temp = w_final × u_rule + (1 - w_final) × u_learn; Where w_final is the fusion weight, u_rule is the first control signal, and u_learn is the second control signal.

[0077] Reference Figure 1 , Figure 2 and Figure 5 As shown, the control method can calculate the fusion weight corresponding to the first control signal based on the first control signal, the second control signal, the scene classification result, and the second confidence signal, and then obtain the adaptation weight corresponding to the second control signal. The adaptation weight is the difference between 1 and the fusion weight, thereby realizing the weighted fusion of the first control signal and the second control signal.

[0078] Reference Figure 1 , Figure 2 and Figure 5 As shown, it can be understood that in step S200, the control method acquires the vehicle's state information and calculates a second control signal and a first confidence signal regarding the second control signal based on the learning-based second control module, including the following steps: Step S210: Obtain environmental information, and use a deep neural network to output a second control signal and a first confidence signal based on the environmental information and state information.

[0079] Reference Figure 1 , Figure 2 and Figure 5 As shown, the learning-based second control module uses a deep neural network. The inputs are vehicle state information (vehicle speed, yaw rate, center of gravity sideslip angle, steering wheel angle, wheel speed, etc.) and environmental information (such as optional perception inputs, including lane line curvature, relative position and speed of target obstacles, etc.). The output is the second control signal u_learn, and the first confidence signal c_raw is also output, where c_raw∈[0,1].

[0080] Reference Figure 1 , Figure 2 and Figure 5As shown, the control method provided in this embodiment of the invention calculates a first confidence signal c_raw, similar to the confidence level of the estimation result in a traditional state estimator. When the vehicle is traveling under normal operating conditions with good training data coverage, the model output variance is small, and the first confidence signal is high; when the vehicle encounters an OOD sample (outside the model's distribution), the model's extrapolation ability decreases, the output variance increases, and the first confidence signal is low. The arbitration module utilizes this information to rely more on the rule-based first control module when the learning-based second control module has high uncertainty, thus ensuring the safety and stability of vehicle operation.

[0081] Reference Figure 1 , Figure 2 and Figure 5 As shown, the first confidence signal can be calculated using the Monte Carlo dropout method: during forward inference, some neurons are randomly dropped, and this is repeated N times (N=20~50). The mean μ and variance σ² of the N outputs are calculated. The first confidence signal c_raw = 1 / (1 + σ²). The larger the variance, the lower the first confidence signal.

[0082] It should be understood that, in some other embodiments, considering the computational overhead of Monte Carlo dropout, in practical engineering applications, the calculation of the first confidence signal can adopt a more computationally efficient method, such as directly having the model output a variance head, which is then trained and predicted together with the control signal; or the calculation of the first confidence signal can also use the maximum Softmax probability of the model output layer, or train an independent confidence evaluation network.

[0083] Reference Figure 1 , Figure 2 and Figure 6 As shown, it can be understood that in this embodiment, the control method, in step S500, includes soft security boundaries and hard security boundaries; determining the fusion weights based on the first control signal, the second control signal, the scene classification result, and the second confidence signal includes the following steps: Step S520: Construct the correspondence between the basic weights and the scene classification results; Step S530: Based on the scene classification results and corresponding relationships output by the scene perception module, output the corresponding basic weights; Step S540: When the second control signal is within the soft security boundary and the hard security boundary, the basic weight is used as the fusion weight.

[0084] Reference Figure 1 , Figure 2 and Figure 6As shown, in step S520, the scene perception module outputs the scene classification result of the current driving scene in real time based on the vehicle state information. The scene perception module adopts a lightweight decision tree model, and the input parameters include: vehicle speed v, steering wheel angle δ and its rate of change dδ / dt, yaw rate ω and its deviation from the expected value Δω, longitudinal acceleration ax, lateral acceleration ay, and road adhesion coefficient estimate μ.

[0085] Reference Figure 1 , Figure 2 and Figure 6 As shown in the table above, referring to the comparison between the scene classification results and the basic weights, the scene classification result is scene ∈ {high-speed cruise, urban congestion, emergency obstacle avoidance, low-friction road surface, extreme working conditions, start-up phase}.

[0086] Reference Figure 1 , Figure 2 and Figure 6 As shown, in step S530, after the arbitration module receives the first control signal, the second control signal, the scene classification result, and the second confidence signal, the arbitration module can calculate the basic weight based on the scene classification result. By querying the lookup table between the scene classification result and the basic weight, the basic weight w_base is obtained. The lookup table between the scene classification result and the basic weight can be based on statically preset domain expert knowledge, defining the basic tendencies of safety and performance under different working conditions.

[0087] Reference Figure 1 , Figure 2 and Figure 6 As shown, in step S540, when the second control signal conforms to both soft and hard safety boundaries, the basic weight is used as the fusion weight. When the scene classification result determined by the scene perception module belongs to a normal operating condition, the weight of the second control signal is appropriately increased to fully leverage the performance advantages of the learning-based second control module and improve control accuracy. Conversely, when the scene classification result determined by the scene perception module belongs to a boundary operating condition, the weight of the second control signal can be appropriately reduced, and driving safety is determined by the rule-based first control module. In other words, this control method leverages the advantages of both the rule-based first control module and the learning-based second control module to improve the safety and stability of the vehicle under various operating conditions.

[0088] Reference Figure 1 , Figure 2 and Figure 6 As shown, the control method dynamically adjusts the weights using the calibrated second confidence signal c through the arbitration module: w_conf = w_base + β × (1 - c) × (w_max - w_base); Where w_max is the preset maximum weight (e.g., 0.95), and β is the confidence level influence factor. When the confidence level is low, 1-c increases, and the weight shifts towards the rule-based first control module.

[0089] The control method compares the second control signal u_learn, output by the learning-based second control module, with the soft safety boundary (e.g., a maximum steering wheel angle change rate of 200° / s based on a comfort definition) of the rule-based first control module through an arbitration module: If u_learn is entirely within the hard and soft safety boundaries, then w_conf remains unchanged.

[0090] Reference Figure 1 , Figure 2 and Figure 7 As shown, it can be understood that in step S500, the control method determines the fusion weights based on the first control signal, the second control signal, the scene classification result, and the second confidence signal; it then performs weighted fusion of the first control signal and the second control signal based on the fusion weights and outputs a fused control signal for controlling the vehicle, including the following steps: Step S550: When the second control signal exceeds the hard safety boundary, set the fusion weight to 1 and use the first control signal as the fusion control signal to control the vehicle. Reference Figure 1 , Figure 2 and Figure 7 As shown, in step S550, if u_learn exceeds the hard safety boundary (such as the maximum steering wheel angle or the maximum yaw rate), or exceeds the soft safety boundary by a preset threshold (such as 30%), then forced takeover is triggered. w_final = 1.0 / / Use the first control signal exclusively; At the same time, the frame data (including state, scene, learning output, degree of excess, etc.) is stored in the Corner Case database for subsequent training.

[0091] Step S560: When the second control signal exceeds the soft safety boundary but not the hard safety boundary, the base weight is adjusted according to the value of the excess part and the fusion weight is output. If the value of the excess part is less than or equal to the preset value, the first control signal and the second control signal are weighted and fused with the fusion weight and the fused control signal is output. Reference Figure 1 , Figure 2 and Figure 7As shown, in step S560, the second control signal u_learn output by the learning-based second control module is compared with the soft safety boundary of the rule-based first control module (e.g., the upper limit of the steering wheel angle change rate of 200° / s based on the comfort definition): If u_learn exceeds the soft safety boundary, the weights are dynamically increased based on the degree of exceedance: w_final = min(1.0, w_conf + γ × Δ); Where Δ is the degree of normalization beyond the boundary (Δ = (|u_learn| - boundary_soft) / (boundary_hard - boundary_soft)), and γ is the gain coefficient (typical value 0.2-0.5).

[0092] γ can be a fixed constant (typically 0.2-0.5), or it can be dynamically determined based on factors such as the overspeed and actuator type. For example:

[0093] Step S570: When the second control signal exceeds the soft safety boundary but not the hard safety boundary, if the value of the exceeding part is greater than the preset value, the fusion weight is set to 1, and the first control signal is used as the fusion control signal to control the vehicle.

[0094] Reference Figure 1 , Figure 2 and Figure 7 As shown, in step S570, when the second control signal exceeds the soft safety boundary but not the hard safety boundary, if the value of the exceeding part is greater than a preset value, a forced takeover is triggered. The arbitration module compares the second control signal u_learn, output by the learning-based second control module, with the hard safety boundary of the rule-based first control module: If u_learn exceeds hard safety boundaries (such as maximum steering wheel angle, maximum yaw rate), or exceeds soft safety boundaries by a preset value (such as 30%), then forced takeover is triggered: w_final = 1.0 / / Use the first control signal exclusively; At the same time, the frame data (including state, scene, learning output, degree of excess, etc.) is stored in the Corner Case database for subsequent training.

[0095] Reference Figure 1 , Figure 2 and Figure 8As shown, it is understandable that existing solutions only focus on the safety boundaries at the single actuator level, without considering the coupling relationships between multiple actuator commands after fusion. The actuator commands output by the rule-based first control module satisfy vehicle dynamics constraints (such as yaw moment balance and tire friction circle limits), but the output of the learning-based second control module does not guarantee satisfaction of these constraints. When the two are weighted and fused, the final output of multiple actuator commands may violate these coupling constraints, leading to vehicle dynamics instability or actuator conflicts. Therefore, the control method further includes the following steps: Step S700: Arbitrate the fused control signal according to the vehicle dynamics coupling constraints; Step S800: If the fused control signal violates the vehicle dynamics coupling constraint, the fused control signal is corrected according to the vehicle dynamics coupling constraint.

[0096] Reference Figure 1 , Figure 2 and Figure 8 As shown, to further ensure safety, this control method can adopt a priority arbitration strategy for the fused control signal u_temp, and sequentially verify and correct the stability constraints, actuator physical limits, tire friction circle constraints, and rate of change constraints, and output the final control signal u_final.

[0097] Where u_temp is the temporary control command vector. For multi-dimensional control signals, weights can be calculated for each dimension separately, or a uniform weight can be used. For example, steering control has a significant impact on safety and can be assigned a higher rule weight; drive torque control has a significant impact on fuel economy and can be assigned a higher learning weight.

[0098] To meet the real-time and deterministic requirements of the ASIL-D functional safety level, the control method provided in this embodiment of the invention employs a priority arbitration strategy for the fused control signals. The priority arbitration strategy verifies and corrects u_temp sequentially according to a preset priority order.

[0099] The priority order and processing method are as follows:

[0100] After the above arbitration, the final control signal u_final, which satisfies the multi-actuator coupling constraint, is output and sent to each chassis actuator.

[0101] The control method provided in this embodiment of the invention adds a final output coordination and verification step after weighted fusion of the first control signal and the second control signal. It adopts a priority arbitration strategy to ensure that the final output actuator commands meet the vehicle dynamics coupling constraints and functional safety requirements, solve the problem of command incoordination after fusion, avoid the risk of instability caused by command incoordination, and meet high functional safety requirements.

[0102] Reference Figure 1 , Figure 2 and Figure 9 As shown, it is understandable that the control method also includes: Step S810: When the vehicle is in the vehicle start-up phase, control the operation of the rule-based first control module, set the fusion weight to 1, and use the first control signal as the fusion control signal to control the vehicle. Step S820: Control the operation of the learning-based second control module and send a ready signal, a second control signal and a first confidence signal to the arbitration module; Step S830: When the arbitration module receives the ready signal, it determines that the vehicle has passed the vehicle start-up phase, and recalculates the fusion weight as the target value based on the second control signal and the first confidence signal. Step S840: Control the vehicle to gradually decrease the fusion weight from 1 to the target value within a preset number of control cycles.

[0103] Reference Figure 1 , Figure 2 and Figure 9 As shown, in steps S810 to S840, the dual-modal parallel cooperative control system of the intelligent chassis further includes a start-up management module. The start-up management module is configured to control the vehicle with a first control signal during the vehicle start-up phase, and to control the vehicle with a fusion control signal that is a weighted fusion of the first and second control signals after the vehicle start-up phase has passed.

[0104] Reference Figure 1 , Figure 2 and Figure 9 As shown, the startup management module is responsible for coordinating the various modules during the vehicle startup phase to achieve smooth integration. The process is as follows: Figure 3 As shown.

[0105] Reference Figure 1 , Figure 2 and Figure 9 As shown, considering the high real-time requirements of chassis control, the control method provided in this embodiment of the invention preferably uses a lightweight or medium-sized model architecture startup management module and adopts UFS 3.0 storage to reduce loading latency.

[0106] Specific procedures: The moment the vehicle is powered on, the control method startup management module immediately activates the rule-based first control module, ensuring that it outputs the first control signal within 5ms, and the vehicle can start moving immediately.

[0107] Meanwhile, the control method asynchronously starts the loading process of the learning-based second control module in the background: reads model parameters from flash memory, initializes the inference engine, and completes the first inference warm-up.

[0108] After the learning-based second control module is loaded, it sends a ready signal to the arbitration module and outputs the first second control signal and its first confidence signal.

[0109] Reference Figure 1 , Figure 2 and Figure 9 As shown, after receiving the ready signal from the learning-based second control module, the arbitration module no longer forces the scenario to the "startup phase," but instead calculates the fusion weight w_target as the target value based on the actual scenario. To avoid abrupt control changes, a smooth transition mechanism is introduced: within M control cycles (e.g., M=10, each cycle 10ms), the weight is linearly decreased from 1.0 to the target value. w_final(k) = 1.0 - (1 - w_target) × (k / M), k = 1..M; After the transition is complete, it will enter normal operation mode.

[0110] Reference Figure 1 , Figure 2 and Figure 9 As shown, the control method provided in this embodiment of the invention has a phased startup strategy for its startup management module. Before the model is loaded, the rule-based first control module completely dominates to ensure that the vehicle can start and drive immediately. After the learning-based second control module is ready, its output is introduced through a smooth transition mechanism (weight gradient) to achieve seamless access and avoid control abrupt changes. This ensures both immediate availability and avoids control abrupt changes, thus improving the user experience.

[0111] It should be understood that this control method can be designed with a two-level model architecture, preloading a lightweight, fast-start model (parameter count < 1M, loading time < 20ms) to provide basic performance during the loading of the full model.

[0112] Reference Figure 1 , Figure 2 and Figure 10 As shown, it can be understood that in this embodiment, the control method further includes the following steps: Step S910: Collect the first control signal, the corresponding vehicle status, the scene classification result, and the second control signal; Step S920: Train and update the learning-based second control module based on the first control signal, the corresponding vehicle state, the scene classification result, and the second control signal.

[0113] The control method provided in this embodiment of the invention collects a first control signal, the corresponding vehicle state, the scene classification result, and a second control signal, and trains and updates the learning-based second control module based on the collected first control signal, the corresponding vehicle state, the scene classification result, and the second control signal.

[0114] The control method employs offline training, primarily based on the following considerations: 1) Functional safety requirements: Online training may lead to unpredictable model behavior, making it difficult to pass ISO 26262 certification.

[0115] 2) Computing resource limitations: The computing power of the vehicle platform is limited, and online training affects real-time control.

[0116] 3) Data quality control: Rule data needs to be screened and cleaned, and online training may introduce noise.

[0117] Workflow of the training module for the control method: 1) Data acquisition: Collect the first control signal u_rule, the corresponding vehicle state x, the scene classification result scene, and the boundary exceedance data when the forced takeover is triggered from the operation log of the rule-based first control module.

[0118] 2) Data Filtering: Only retain data that meets the following conditions: vehicle status is stable (e.g., lateral acceleration change rate is less than a threshold). The rule-based first control module has not triggered any anomalies. Diverse scenarios are covered (sampling is balanced according to scenario classification results).

[0119] Data that triggers mandatory takeover is marked as "high-value samples in edge scenarios" and given higher weight during training.

[0120] 3) Model Training: Using x as input and u_rule as the supervision label, fine-tune the learning-based second control module. During training, it is important to employ strategies such as Elastic Weight Consolidation (EWC) to prevent the original performance from being forgotten when fine-tuning on rule-based data.

[0121] 4) Hybrid training: Mix rule data with original training data in proportion to balance security and performance.

[0122] 5) Deployment and Update: The newly trained model is deployed back to the learning-based second control module via OTA.

[0123] Reference Figure 1 , Figure 2 and Figure 10As shown, the control method provided in this embodiment of the invention introduces an offline training module. This module collects and filters high-quality security data (especially edge scenario data triggered by forced takeover) generated during the operation of the rule-based first control module. This data is then used to continuously fine-tune the learning-based second control module. This forms a positive closed loop of "rule-based baseline → data collection → model optimization → performance improvement," enabling the control method to self-evolve. This allows the learning-based second control module to gradually internalize rule-based security knowledge, reducing its real-time dependence on the rule-based first control module and laying the foundation for a future transition to a more efficient pure learning solution.

[0124] To further illustrate the control method provided in the embodiments of the present invention, several embodiments in different scenarios are described below: Example 1: Principle verification under high-speed cruise conditions Scenario description: The vehicle is cruising at 120km / h on a highway with clear lane markings and a dry road surface (μ≈0.8). A sinusoidal sweep steering input (frequency 0.2Hz, amplitude ±1°) is executed to simulate lane keeping.

[0125] Test path: High-speed curves with curvature of 0.001-0.002 1 / m, conforming to the ISO 11270 lane keeping test standard.

[0126] Input parameters: vehicle speed v=120km / h, steering wheel angle δ varies sinusoidally between ±1°.

[0127] Scene classification: The scene perception module outputs scene="high-speed cruise", and the lookup table shows w_base=0.2.

[0128] The second control module output based on learning: The steering wheel angle output by the second control module based on learning (using variance head confidence estimation) follows the sine change, with a root mean square error of 0.08° and a confidence level of c=0.95.

[0129] Confidence correction: w_conf = 0.2 + 0.5×(1-0.95)×(0.95-0.2) = 0.2 + 0.01875 = 0.21875; Safety boundary verification: The soft safety boundary (based on comfort settings) of the rule-based first control module is a steering wheel angle change rate of ≤200° / s, and the output of the learning-based second control module is much lower than this threshold.

[0130] Weighted fusion and coordination verification: u_temp = 0.22×u_rule + 0.78×u_learn. Since u_temp does not violate any coupling constraints, priority arbitration directly outputs u_final = u_temp.

[0131] Simulation results (CarSim / Simulink co-simulation):

[0132] Technical effect: Theoretically, this control method can achieve a balance between high precision (lateral error of 0.10m, which is 33% better than pure rule control) and high comfort (angle change rate of 1.6° / s), and no safety boundary intervention is triggered throughout the process.

[0133] Example 2: Principle verification under the scenario of decreased confidence Scenario Description: A vehicle enters a rainstorm area; the road surface is slippery (μ=0.4), and lane lines are partially blurred. The vehicle is performing a lane keeping task. This scenario represents an out-of-distribution (OOD) sample for the model.

[0134] Input parameters: vehicle speed v = 80 km / h, yaw rate ω fluctuates more, and the standard deviation increases from 0.5° / s to 1.2° / s.

[0135] Scene classification: The scene perception module outputs scene="low-adhesion road surface", w_base=0.7.

[0136] The output of the learning-based second control module shows significant fluctuations due to the limited number of rainstorm scenarios in the training data, with a root mean square error (RMS) of 0.25° for the steering wheel angle. The first confidence signal, c_raw, estimated using the variance head, is 0.60. After calibration by the calibration module (using ordinal-preserving regression), c=0.55, correcting the model's overconfidence.

[0137] Confidence correction: w_conf = 0.7 + 0.5×(1-0.55)×(0.95-0.7) = 0.7 + 0.05625 = 0.75625; Safety boundary settings: Hard safety boundary: steering wheel angle ≤ 35° (corresponding to a lateral acceleration of approximately 0.4g, which is the physical limit for wet and slippery roads); Soft safety boundary: steering wheel angle ≤ 22° (corresponding to a lateral acceleration of approximately 0.25g, which is the upper limit for comfort on wet and slippery roads).

[0138] Safety boundary verification: The maximum value of the fluctuation in the output of the learning-based second control module is the steering wheel angle of 25°, which exceeds the soft boundary of 22°, but is still within the hard boundary of 35°.

[0139] Exceeding the limit: Δ = (25-22) / (35-22) = 3 / 13 ≈ 0.23; Let the gain coefficient γ = 0.3, then: w_final = min(1.0, 0.75625 + 0.3×0.23) =0.75625 + 0.069 = 0.82525; Weighted fusion and coordination verification: Weighted fusion yields u_temp = 0.825×u_rule + 0.175×u_learn. The coupling constraints of u_temp are checked. The braking force of the left front wheel does not violate the friction circle, and the rate of change is normal. Stability at P0 is passed, and P1~P3 are all passed. P4 records a slight exceedance of the lateral acceleration limit. The final output is u_final = u_temp. Simulation results (CarSim / Simulink co-simulation):

[0140] Technical Results: Under low model confidence, pure learning control exceeded the safety boundary twice. The confidence-guided arbitration mechanism of this control method automatically increased the rule weights to approximately 0.825. While ensuring safety (0 instances of exceeding the hard safety boundary), it still retained approximately 17.5% of the contribution from the learning-based second control module, resulting in a lateral tracking error (0.31m) that is superior to pure rule control (0.35m), achieving a balance between safety and performance. Furthermore, the corrected safety boundary matches the physical characteristics of the slippery road surface, ensuring the dynamic rationality of the data.

[0141] Example 3: Gradual Adjustment Beyond Soft Security Boundaries Scenario Description: The vehicle is cruising at 100 km / h on a highway, maintaining lane keeping. Ahead, a curve with gradually increasing curvature appears, with the radius of curvature transitioning from 2000m to 500m.

[0142] Input parameters: vehicle speed v = 100 km / h, steering wheel angle requirement increases with curve curvature.

[0143] Safety boundary settings (based on μ≈0.8): Soft safety boundary (based on the arbitration reference line set for comfort): Steering wheel angle ≤ 30° (corresponding to a lateral acceleration of approximately 0.35g). Hard safety boundary (based on the physical limit MPC hard constraint): Steering wheel angle ≤ 45° (corresponding to a lateral acceleration of approximately 0.52g).

[0144] The rule-based first control module (MPC) outputs the following: Based on a simplified linear bicycle model, the MPC predicts that a 28° steering wheel angle is needed to track the lane line in this curve. However, because the model does not consider practical factors such as tire nonlinearity and suspension kinematics, the MPC's prediction suffers from model mismatch, and its output of 28° is not the true optimal value.

[0145] The learning-based second control module outputs: By learning from a large amount of historical data (including vehicle response under MPC control), the learning-based second control module has captured the deviation between the MPC model and the actual system, and outputs a 33° steering wheel angle, which is closer to the true optimal control value.

[0146] Confidence: The model has a high confidence level for this curve scenario. The first confidence signal estimated using the variance head is calibrated to c=0.88.

[0147] Scene classification: The scene perception module outputs scene="high-speed cruise", and the lookup table shows w_base=0.2.

[0148] Confidence adjustment: w_conf = 0.2 + 0.5×(1-0.88)×(0.95-0.2) = 0.2 + 0.045 = 0.245; Safety boundary verification: The learned output of 33° exceeds the soft boundary of 30°, but is still within the hard boundary of 45°.

[0149] Exceeding the limit: Δ = (33-30) / (45-30) = 3 / 15 = 0.2; Let the gain coefficient γ = 0.3, then: w_final = min(1.0, 0.245 + 0.3×0.2) = 0.245 +0.06 = 0.305; Weighted fusion and coordination verification: Weighted fusion yields u_temp = 0.305×28° + 0.695×33° = 31.48°. The coupling constraints of u_temp are checked. The steering wheel angle of 31.48° remains within the hard boundary, the rate of change is within the limit, and other actuators are normal. P0~P3 pass, P4 records a lateral acceleration of 0.36g slightly exceeding 0.35g. The final output is u_final = u_temp. Simulation results (CarSim / Simulink co-simulation):

[0150] Technical Results: This embodiment demonstrates the effectiveness of the incremental adjustment mechanism when the output of the learning-based second control module exceeds the soft safety boundary but remains within the hard boundary. This control method, while ensuring that the hard boundary is not exceeded, increases the weight from 0.245 to 0.305, resulting in a final control value of 31.48° that falls between the two. The lateral error (0.12m) is superior to that of pure rule control (0.15m), and the lateral acceleration of 0.36g only slightly exceeds the comfort boundary of 0.35g, achieving a balance between performance and comfort.

[0151] Example 4: Security Boundary Triggering and Data Acquisition Scenario description: The vehicle is traveling at 60km / h and suddenly encounters an obstacle in front, requiring emergency obstacle avoidance.

[0152] Input parameters: The steering wheel angular velocity dδ / dt increases sharply to 300° / s.

[0153] Scene classification: The scene perception module outputs scene="emergency obstacle avoidance", w_base=0.8.

[0154] The second control module based on learning outputs a steering wheel angle of 80° in order to achieve the obstacle avoidance effect. However, the first control module based on rules calculates a hard safety boundary of 60° based on the current vehicle speed and road surface adhesion (μ≈0.8) (based on the maximum lateral force constraint).

[0155] Forced takeover: When the learning signal exceeds the hard boundary (exceeding the limit Δ = (80-60) / 60 ≈ 0.33 > 30%), forced takeover is triggered, w_final = 1.0, and the 60° control of the rules module is fully adopted. u_temp = u_rule.

[0156] Coordination and Verification: The priority arbitration module receives u_temp. Since it is already the first control signal of the rule-based first control module, all constraints are satisfied, and it directly outputs u_final = u_temp.

[0157] Data Acquisition: The data in this frame (state: v=60km / h, δ=80°, μ=0.8; scenario: emergency obstacle avoidance; learning output: 80°; exceedance level: 33%) is stored in the Corner Case database.

[0158] Offline training: After accumulating a certain amount of such data, the learning-based second control module can be fine-tuned to gradually learn the safety boundaries of the rules. Theoretically, with sufficient training, the model's output in the same scenario can approach the safety boundary (e.g., reduced to approximately 65°), reducing the occurrence of forced takeovers.

[0159] Principle Analysis: This embodiment demonstrates the principle of the closed-loop evolution mechanism of the control method: safety is ensured by forced takeover, while safety events are transformed into training data to provide material for subsequent model optimization.

[0160] Example 5: Multi-actuator coupling constraint verification (priority arbitration) Scenario description: The vehicle is traveling at 60 km / h on a wet and slippery road surface (μ=0.4) while performing emergency obstacle avoidance and braking.

[0161] The first control module output based on rules (MPC) is: left front wheel braking force 2000N, right front wheel braking force 1800N, front wheel rotation angle 5°, and each command satisfies the tire friction circle constraint F_x²+F_y² ≤ (μ·F_z)² (the maximum allowable left front wheel is about 2400N).

[0162] The second control module, based on learning, outputs the following to achieve the shortest braking distance: a braking force of 2800N for the left front wheel, a braking force of 2500N for the right front wheel, and a front wheel steering angle of 6°. The 2800N braking force for the left front wheel exceeds the friction circle constraint.

[0163] Weighted fusion (w_final=0.4): In u_temp, the braking force of the left front wheel = 0.4×2000 + 0.6×2800 = 2480N, the braking force of the right front wheel = 0.4×1800+0.6×2500 = 2220N, and the front wheel steering angle = 0.4×5+0.6×6 = 5.6°.

[0164] Priority arbitration process: P0 (Stability): Yaw rate error not exceeded the limit → No action.

[0165] P1 (Physical Limit): All braking forces within [0, 3000N] → No action.

[0166] P2 (Friction Circle): Left front wheel: Fx=2480N, Fy=300N (estimated), Fz=3500N, μ=0.4, max_force=1400N, Resultant force = sqrt(2480²+300²)=2498N>1400N → Proportional scaling: scale=1400 / 2498≈0.56, after correction: Fx=2480×0.56=1389N, Fy=300×0.56=168N. Right front wheel: Fx=2220N, Fy=200N, max_force=1400N, Resultant force =2229N>1400N → scale≈0.628, after correction: Fx=1394N, Fy=126N.

[0167] P3 (rate of change): The change in braking force relative to the previous cycle is approximately 200N, less than 5000N / s → no action.

[0168] P4 (Soft Constraint): Lateral acceleration 0.39g > 0.35g → Log.

[0169] Finally: All instructions satisfy hard boundaries → Output u_final.

[0170] The corrected u_final: left front wheel braking force 1389N, right front wheel braking force 1394N, front wheel rotation angle 5.6° (unchanged), satisfying the friction circle constraint.

[0171] Simulation results:

[0172] Technical effect: This embodiment demonstrates the key role of the priority arbitration strategy in preventing multi-actuator instructions from violating coupling constraints, ensuring that the fused instructions retain the performance contribution of the learning-based second control module while also satisfying vehicle dynamics feasibility.

[0173] Example 6: Smooth Entry During Vehicle Start-up Scenario description: The vehicle is powered on and started for the first time in the morning. The model is of medium size (15M parameters, 60MB storage size, UFS 3.0 loading time of about 120ms).

[0174] Process deduction: t=0ms: The vehicle is powered on, and the startup management module immediately activates the rule-based first control module. The rule-based first control module outputs the first control signal within 5ms.

[0175] t=0~120ms: The rule-based first control module continuously outputs the first control signal, while the learning-based second control module is asynchronously loaded in the background.

[0176] t=120ms: The learning-based second control module completes loading and outputs the first frame of the second control signal u_learn=0.2°. After calibration, the second confidence level c=0.9 is obtained. The scene perception module outputs scene="high-speed cruise" and the target weight w_target=0.2.

[0177] t=120~220ms: Entering the smooth transition period, the weight decreases linearly from 1.0 to 0.2. During this period, u_temp is merged according to the decreasing weight, and the coordination module performs normal verification.

[0178] After t=220ms: Enter normal operation mode.

[0179] Performance Analysis: Users can start the vehicle and drive immediately without waiting for the learning-based second control module to load; the transition process is smooth, theoretically avoiding abrupt changes in the control signal.

[0180] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A dual-modal parallel cooperative control system for an intelligent chassis, characterized in that, Applied to vehicles with intelligent chassis, the dual-modal parallel cooperative control system of the intelligent chassis includes: The rule-based first control module is configured to calculate a first control signal based on the vehicle's state information; The learning-based second control module is configured to calculate a second control signal and a first confidence signal about the second control signal based on the vehicle's state information. The calibration module is configured to calibrate the first confidence signal and output the calibrated second confidence signal. The scene perception module is configured to output the scene classification result of the current driving scene based on the vehicle's state information; The arbitration module is configured to determine the fusion weight based on the first control signal, the second control signal, the scene classification result, and the second confidence signal, perform weighted fusion of the first control signal and the second control signal based on the fusion weight, and output a fusion control signal for controlling the vehicle.

2. The dual-modal parallel cooperative control system for an intelligent chassis according to claim 1, characterized in that, The dual-modal parallel collaborative control system of the intelligent chassis also includes a startup management module; the startup management module is configured to control the vehicle only through the rule-based first control module during the vehicle startup phase, and to control the vehicle with the fused control signal obtained by weighted fusion of the first control signal and the second control signal after the vehicle startup phase is completed; And / or, the dual-modal parallel collaborative control system of the intelligent chassis further includes a training module, which is configured to collect the first control signal, the corresponding vehicle state, the scene classification result and the second control signal, and train and update the learning-based second control module based on the collected first control signal, the corresponding vehicle state, the scene classification result and the second control signal.

3. A control method, characterized in that, The dual-modal parallel cooperative control system applied to the intelligent chassis according to claim 1 or 2, wherein the control method includes: The vehicle's status information is acquired and a first control signal is calculated based on the rule-based first control module, wherein the first control signal has a safety boundary. The system acquires the vehicle's state information and calculates a second control signal and a first confidence signal regarding the second control signal based on the learning-based second control module. The first confidence signal is calibrated, and the calibrated second confidence signal is output. Obtain the vehicle's status information and output the scene classification result of the current driving scenario; The fusion weight is determined based on the first control signal, the second control signal, the scene classification result, and the second confidence signal. The first control signal and the second control signal are weighted and fused according to the fusion weight, and a fusion control signal for controlling the vehicle is output.

4. The control method according to claim 3, characterized in that, The step of weightedly fusing the first control signal and the second control signal according to the fusion weight, and outputting a fused control signal for controlling the vehicle, includes: The product of the first control signal and the fusion weight is used as the first calculated value, and the product of the difference between 1 and the fusion weight and the second control signal is used as the second calculated value. The fusion control signal is the sum of the first calculated value and the second calculated value.

5. The control method according to claim 3, characterized in that, The step of acquiring the vehicle's state information and calculating a second control signal and a first confidence signal regarding the second control signal based on the learning-based second control module includes: The system acquires environmental information and uses a deep neural network to output the second control signal and the first confidence signal based on the environmental information and the state information.

6. The control method according to claim 4, characterized in that, The security boundary includes a soft security boundary and a hard security boundary; determining the fusion weight based on the first control signal, the second control signal, the scene classification result, and the second confidence signal includes: Construct a correspondence between the basic weights and the scene classification results; Based on the scene classification result and the corresponding relationship output by the scene perception module, the corresponding basic weight is output; When the second control signal is within the soft security boundary and the hard security boundary, the basic weight is used as the fusion weight.

7. The control method according to claim 6, characterized in that, The step of determining fusion weights based on the first control signal, the second control signal, the scene classification result, and the second confidence signal, weighting and fusing the first control signal and the second control signal according to the fusion weights, and outputting a fused control signal for controlling the vehicle includes: When the second control signal exceeds the hard safety boundary, the fusion weight is set to 1, and the first control signal is used as the fusion control signal to control the vehicle. Alternatively, when the second control signal exceeds the soft safety boundary but not the hard safety boundary, the base weight is adjusted according to the value of the excess portion and the fusion weight is output. If the value of the excess portion is less than or equal to a preset value, the first control signal and the second control signal are weighted and fused using the fusion weight, and the fused control signal is output. Alternatively, if the second control signal exceeds the soft safety boundary but not the hard safety boundary, and the value of the exceeding portion is greater than a preset value, the fusion weight is set to 1, and the first control signal is used as the fusion control signal to control the vehicle.

8. The control method according to claim 7, characterized in that, The control method further includes: Arbitrate the fused control signal based on vehicle dynamics coupling constraints; If the fusion control signal violates the vehicle dynamics coupling constraint, the fusion control signal is corrected according to the vehicle dynamics coupling constraint.

9. The control method according to claim 4, characterized in that, The control method further includes: When the vehicle is in the vehicle start-up phase, the rule-based first control module is controlled to operate, the fusion weight is set to 1, and the first control signal is used as the fusion control signal to control the vehicle. Control the operation of the learning-based second control module and send a ready signal, the second control signal, and the first confidence signal to the arbitration module; When the arbitration module receives the ready signal, it determines that the vehicle has passed the vehicle start-up phase, and recalculates the fusion weight as the target value based on the second control signal and the first confidence signal. The vehicle is controlled to gradually decrease the fusion weight from 1 to the target value within a preset number of control cycles.

10. The control method according to claim 3, characterized in that, The control method further includes: Collect the first control signal, the corresponding vehicle status, the scene classification result, and the second control signal; The learning-based second control module is trained and updated based on the first control signal, the corresponding vehicle state, the scene classification result, and the second control signal.

Images