CN-122008763-A - Active suspension pre-aiming control method based on road surface sensing and hybrid synchronization strategy

CN122008763ACN 122008763 ACN122008763 ACN 122008763ACN-122008763-A

Abstract

The invention relates to the technical field of intelligent control of vehicles and provides an active suspension pre-aiming control method based on road surface sensing and hybrid synchronization strategies, which comprises the steps of acquiring image data of a road ahead and laser radar point cloud data; the method comprises the steps of processing image data and point cloud data by adopting a layered road sensing module, extracting structured road state information, splicing the road state information with current physical state information of a vehicle suspension system to form an augmented observation state, and outputting optimal active control force according to the augmented observation state based on a hybrid synchronous strategy reinforcement learning control module so as to regulate and control a suspension actuator in real time. The invention can efficiently and stably utilize the pre-aiming information of the front road, further construct an active suspension control strategy and remarkably improve the riding comfort of the vehicle under various complex road conditions.

Inventors

HAN SHIYUAN
Chang Keyao
XIN DONGJIN
JIA CHEN
MA NAN
ZHANG XIAOFANG
WANG RUI

Assignees

山东女子学院
山东凌宇通信技术有限公司

Dates

Publication Date: 20260512
Application Date: 20260311

Claims (9)

1. The active suspension pre-aiming control method based on the road surface sensing and the hybrid synchronization strategy is characterized by comprising the following steps of: acquiring image data of a road ahead and laser radar point cloud data; Processing the image data and the point cloud data by adopting a layered road perception module, and extracting structured road state information; Splicing the road state information with the current physical state information of the vehicle suspension system to form an augmented observation state; based on the mixed synchronization strategy reinforcement learning control module, outputting optimal active control force according to the augmented observation state so as to regulate and control the suspension actuator in real time; The hybrid synchronous strategy reinforcement learning control module adopts a dynamically coupled same-strategy learning and different-strategy learning method, and adopts a synchronous mechanism to periodically align learning targets of the same-strategy learning and different-strategy learning method.
2. The method of claim 1, wherein the road status information includes at least a road category label and a forward road elevation value; the current physical state information of the vehicle suspension system includes sprung mass acceleration, velocity, displacement, and suspension dynamic deflection.
3. The method of claim 1, wherein the hierarchical road perception module comprises a road classification unit and a road elevation estimation unit arranged in parallel; The road classification unit is used for processing the image data by adopting an improved lightweight convolutional neural network and outputting a road class label; the road elevation estimation unit is used for processing the laser radar point cloud data by adopting a Kalman filtering enhanced random sampling consistency algorithm, fitting a local road plane and calculating a front road elevation value.
4. A method according to claim 3, wherein the improved lightweight convolutional neural network employs a hybrid attention-directed one-dimensional feature extraction mechanism, for multi-scale features, first fusing channels with spatial attention to generate a unified mask, then performing attention-directed spatial max pooling, retaining high-discriminant local information, and outputting road class labels in combination with a multi-head classifier.
5. A method according to claim 3, characterized in that the kalman filter enhanced random sample consistency algorithm process is specifically: And constructing a state space model taking a road plane equation coefficient as a state vector, fusing point cloud data among time sequence frames by using a Kalman filter, taking plane parameter estimation at the previous moment as a priori, taking plane parameters obtained by fitting a current frame through a RANSAC algorithm as an observation value, carrying out recursive state update to obtain local road plane estimation, and analyzing and calculating a road elevation value according to the local road plane estimation.
6. The method of claim 1, wherein the hybrid synchronization policy reinforcement learning control module implements dynamic coupling learning by: setting a round buffer zone and a replay buffer zone which are respectively used for storing a complete interaction track and a single-step transfer sample crossing rounds; in the same strategy learning stage, calculating weighted multi-round expected returns based on the complete track in the round buffer area, and updating a strategy network and a value network based on a strategy gradient method; In the different strategy learning stage, periodically sampling historical data from the replay buffer area, updating a Q function integration and strategy network, adopting Q network integration and optimizing the most conservative Q value, and introducing a Q distribution network to apply distribution constraint so as to relieve overestimation; And (3) aligning the randomly selected Q network estimated value with the multi-round expected return calculated in the same strategy stage in each training period through a synchronization mechanism so as to realize synchronization of two learning paradigm learning targets.
7. The method of claim 6, wherein the weighted multi-round expected return The calculation formula of (2) is as follows: , Wherein, the 、、 Respectively representing k-step rewards calculated at time step t for the current round and the first two historical rounds, , , Is a weight parameter that can be learned.
8. The method according to claim 1, wherein in the step S4, the action space is adaptively divided according to the road class label, specifically: compressing the original action output by the strategy network through a tanh function, and multiplying the original action by a scaling factor related to the current road class to obtain the control force finally acting on the actuator; For a severe road, the value of the scaling factor is larger, and the force output by the actuator is larger correspondingly; for a gentle road, the value of the scaling factor is smaller, and the force output by the actuator is smaller.
9. The method of claim 1, wherein the hybrid synchronization strategy reinforcement learning control module designs the bonus function using a soft normalization method, comprising: and normalizing the state variables by taking the response extremum of the passive suspension under the typical excitation as a normalization reference, wherein the reward function comprises overflow penalty applied to states beyond the response range of the passive system, smooth negative rewards applied to key performance indexes and sparse positive rewards applied to acceleration and suspension dynamic deflection below a specific threshold.

Description

Active suspension pre-aiming control method based on road surface sensing and hybrid synchronization strategy Technical Field The invention relates to the technical field of intelligent control of vehicles, in particular to an active suspension pre-aiming control method based on road surface sensing and hybrid synchronization strategies. Background The vehicle suspension system is a key subsystem of the chassis, and has the main functions of damping the vibration of the vehicle body caused by road surface unevenness, and plays a decisive role in the steering stability, the running smoothness and the riding comfort of the vehicle. Traditional suspension control strategies (e.g., PID, LQR) and robust control methods (e.g., H-infinity control, sliding mode control) rely primarily on feedback of the current suspension state for passive or semi-active adjustment. The method can not utilize the front road information, and has inherent hysteresis when responding to sudden or periodic road surface excitation, so that the further improvement of riding comfort is limited. To overcome the limitations described above, pretightening control techniques have been introduced in the active suspension field. The technology obtains the front road outline information in advance through a vehicle-mounted sensor (such as a camera and a laser radar), and performs prospective adjustment according to the front road outline information. The existing pretightening control method generally adopts Model Predictive Control (MPC), but the performance of the pretightening control method is highly dependent on an accurate vehicle dynamics model, and the computing complexity is high, so that the pretightening control method is difficult to adapt to actual working conditions of parameter time variation and strong nonlinearity. In recent years, deep Reinforcement Learning (DRL) has been widely used for active suspension control due to its strong model independent nature and the ability to autonomously make decisions in complex environments. However, most of the existing DRL methods make reactive decisions based only on the current suspension state, and cannot effectively utilize pre-aimed road information. Even if there were studies attempting to splice road elevation data directly to the status input, the lack of efficient characterization and fusion of multi-modal road information (e.g., image semantics and point cloud geometry) resulted in limited performance gains. In addition, the DRL algorithm often faces the challenge that the sample efficiency and the training stability are difficult to be compatible in the training process, namely, an on-poll algorithm (such as PPO) is stable but the sample efficiency is low, and an off-poll algorithm (such as TD3 and SAC) is high in sample efficiency but is easy to be influenced by function approximation errors to generate overestimation, so that the training is unstable. Therefore, there is a need for a reinforcement learning control framework that can efficiently fuse multi-source road perception information and design a reinforcement learning control framework that can simultaneously achieve training stability and sample efficiency, so as to fully exploit the potential of pre-aiming information in active suspension control. Disclosure of Invention In view of the above, the invention provides an active suspension pre-aiming control method based on road surface sensing and hybrid synchronization strategies, which aims to remarkably improve the driving smoothness and riding comfort of a vehicle under complex road conditions. In order to achieve the above purpose, the invention provides an active suspension pre-aiming control method based on road surface sensing and hybrid synchronization strategy, which is characterized by comprising the following steps: acquiring image data of a road ahead and laser radar point cloud data; Processing the image data and the point cloud data by adopting a layered road perception module, and extracting structured road state information; Splicing the road state information with the current physical state information of the vehicle suspension system to form an augmented observation state; based on the mixed synchronization strategy reinforcement learning control module, outputting optimal active control force according to the augmented observation state so as to regulate and control the suspension actuator in real time; The hybrid synchronous strategy reinforcement learning control module adopts a dynamically coupled same-strategy learning and different-strategy learning method, and adopts a synchronous mechanism to periodically align learning targets of the same-strategy learning and different-strategy learning method. Further, the road status information at least comprises a road category label and a front road elevation value; the current physical state information of the vehicle suspension system includes sprung mass acceleration, velocity, displacement, and suspens