Vehicle suspension control method and apparatus

Abstract

The application provides a vehicle suspension control method and device, relates to the technical field of vehicle suspension control, and first synchronously collects road environment, vehicle body dynamics and vehicle driving data through a multi-source vehicle-mounted sensor, extracts a multimodal feature vector through pretreatment, inputs a time sequence deep learning network mixed with a time sequence convolution network and a long short-term memory network, and predicts a road disturbance time sequence matrix in a future time window; then, a suspension dynamics state equation is constructed based on the disturbance matrix, a double hard constraint is combined with an actuator physical limit and a suspension safety threshold to generate a quadratic programming control optimization target; and finally, an optimal control force sequence is obtained through lightweight solving, a suspension actuator is driven to complete feedforward disturbance cancellation control, so that the pain point of response lag of traditional suspension control is solved, and ride comfort and driving stability are considered.

Description

Technical Field

[0001] This invention relates to the technical field of vehicle suspension control, and in particular to a vehicle suspension control method and apparatus. Background Technology

[0002] The vehicle suspension system is a key chassis system that ensures ride comfort, handling stability, and occupant safety. With the development of intelligent driving and electrification, users have higher requirements for dynamic responsiveness, multi-condition robustness, and energy efficiency. Traditional control paradigms based on real-time feedback are no longer sufficient to meet the needs of forward-looking regulation under complex road conditions.

[0003] Current mainstream solutions fall into three categories: passive suspensions lack adjustment capabilities; semi-active systems (such as MR dampers) rely on single sensor feedback and lack disturbance prediction; and active systems (such as linear motor suspensions) have strong intervention capabilities, but mostly employ PID or fuzzy control without modeling spatiotemporal evolution characteristics. In recent years, AI-driven solutions have emerged, with some studies attempting to use neural networks to predict road surface height, introduce fuzzy adaptive strategies, or fuse V2X information for road condition prediction. However, these solutions are essentially still limited to instantaneous state feedback or discrete event triggering.

[0004] The fundamental limitation of existing technologies lies in the fact that they cannot achieve continuous, low-latency prediction of the evolution trend of disturbances within 0.5–1.0 seconds in the time dimension, nor can they correlate the terrain undulations within a few meters to more than ten meters in front of the vehicle in the spatial dimension. More importantly, the prediction output is often directly input into the open-loop controller, lacking a dynamic constraint optimization mechanism that coordinates with the force limit, displacement limit, energy consumption limit of the suspension actuator and the stability boundary of vehicle dynamics. This results in the prediction value not being able to be converted into reliable actions. Ultimately, this manifests as control oscillation, actuator saturation, response lag or low energy efficiency, making it difficult to balance prediction accuracy, real-time performance and engineering robustness under automotive-grade computing power constraints. Summary of the Invention

[0005] The purpose of this invention is to provide a vehicle suspension control method and device to alleviate the technical problem of vehicle suspension control response lag or low energy efficiency, making it difficult to achieve prediction accuracy under automotive-grade computing power constraints.

[0006] In a first aspect, the present invention provides a vehicle suspension control method, comprising: Based on multimodal feature vectors corresponding to multi-source sensor data and a temporal deep learning network, predict the temporal matrix of road disturbance within a future time window; Based on the road disturbance time series matrix, the dynamic state equation of the suspension system is constructed, and the constraint boundary is set by combining the physical limit and safety threshold of the suspension actuator to generate a constrained control optimization objective. Solving the control optimization objective yields the optimal control force sequence, which drives the suspension actuator to perform disturbance cancellation control.

[0007] In an optional implementation, the step of predicting the road disturbance time series matrix within a future time window based on the multimodal feature vectors corresponding to multi-source sensor data and a temporal deep learning network includes: Road environment data, vehicle dynamic data and vehicle driving data are collected synchronously by multi-source vehicle-mounted sensors, and multimodal feature vectors are extracted after preprocessing. Based on the multimodal feature vectors, a temporal deep learning network is used to predict the road disturbance time series matrix within a future time window.

[0008] In an optional implementation, the step of predicting the road disturbance time-series matrix within a future time window using a temporal deep learning network based on the multimodal feature vector includes: The multimodal feature vectors are input into a pre-trained temporal deep learning network; wherein the temporal deep learning network is a hybrid of a temporal convolutional network and a long short-term memory network; The spatial distribution and short-term temporal dependence features of road disturbances are extracted using the temporal convolutional network, and then long-term temporal correlation features are mined based on the long short-term memory network. The features are then spliced ​​and fused to obtain a feature vector that maps road spatial features to vehicle driving time. Based on the sequence-to-sequence decoding structure in the Long Short-Term Memory network, the feature vector is used to generate an initial sequence of perturbation amplitude for each wheel at each time step within the future time window. By combining the physical limits of vehicle suspension and the dynamic characteristics of vehicle body, the initial sequence of disturbance amplitude is constrained, corrected and standardized to obtain a road disturbance time series matrix that meets the requirements of suspension control.

[0009] In an optional implementation, the steps of constructing the dynamic state equation of the suspension system based on the road disturbance time series matrix, setting constraint boundaries by combining the physical limits and safety thresholds of the suspension actuators, and generating constrained control optimization objectives include: Based on the road disturbance time series matrix, combined with the vertical dynamic characteristics of the vehicle suspension and the real-time transmitted operating status data, a suspension system motion state equation is constructed to characterize the dynamic mapping relationship between the output force of the suspension actuator and the vertical displacement, vertical velocity and acceleration of the vehicle body. Minimizing the deviation between the vehicle's vertical attitude and the target reference attitude, and minimizing the vehicle's vertical motion speed are the two optimization objectives. A secondary planning optimization objective covering the future time window is constructed. The upper and lower limits of the output force of the suspension actuator under physical limit conditions and the suspension stroke displacement constraint corresponding to the safety threshold are superimposed on the quadratic programming optimization objective to generate a constrained control optimization objective for the suspension actuator.

[0010] In an optional implementation, the step of solving the control optimization objective to obtain the optimal control force sequence includes: Based on the dynamic state equation of the suspension system, an approximate linearization process is performed at the current driving condition point of the vehicle to construct a linear time-varying model predictive control framework. The control optimization objective is transformed into a recursively solvable quadratic programming matrix form. The calculation results of the previous cycle are reused by a recursive matrix decomposition algorithm to obtain the optimal control force sequence of the suspension actuator covering the future time window.

[0011] In an optional implementation, prior to the step of driving the suspension actuator for disturbance cancellation control, the method further includes: Based on the road disturbance time series matrix, the future disturbance level is identified. Combined with the vehicle's real-time speed, vehicle body vertical attitude, and suspension travel displacement in the multimodal feature vector, the weight coefficients of the deviation between the vehicle body vertical attitude and the target reference attitude and the vehicle body vertical motion speed in the control optimization target are dynamically adjusted. Based on the adjusted weighting coefficients, the optimal control force sequence is modified to adapt to different operating conditions, and the future disturbance gradient decay weights are superimposed to generate the final control command with active pre-control attributes.

[0012] In an optional implementation, the step of driving the suspension actuator for disturbance cancellation control includes: According to the final control command corresponding to the optimal control force sequence, the suspension actuator is driven to output control force in advance to feed forward and cancel out future road disturbances. By collecting real-time operating status data of the suspension actuator, including actual output force, stroke and vehicle vertical acceleration, the final control command is corrected in real-time using a closed loop, and the final execution command is generated to drive the suspension actuator to move. Simultaneously, the operational status data is transmitted back to the stage of constructing the dynamic state equation of the suspension system.

[0013] In a second aspect, the present invention provides a vehicle suspension control device, comprising: The prediction module, based on multimodal feature vectors corresponding to multi-source sensor data and a temporal deep learning network, predicts the road disturbance time series matrix within a future time window; The generation module constructs the dynamic state equation of the suspension system based on the road disturbance time series matrix, sets constraint boundaries by combining the physical limits and safety thresholds of the suspension actuators, and generates a constrained control optimization objective. The control module solves the control optimization objective to obtain the optimal control force sequence, and drives the suspension actuator to perform disturbance cancellation control.

[0014] Thirdly, the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the method as described in any of the foregoing embodiments.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed, implements the method described in any of the foregoing embodiments.

[0016] This invention provides a vehicle suspension control method and device. By fusing multi-modal feature vectors from multiple sources of sensors with a temporal deep learning network, it achieves accurate prediction of future road disturbances in advance. This overcomes the limitations of traditional suspension systems that rely solely on real-time vehicle feedback, and solves the inherent response lag problem of traditional feedback control. Using the predicted road disturbance time-series matrix as the sole core input, a dynamic state equation of the suspension system that fits the actual operating state of the vehicle is constructed. By combining the hardware physical limits of the suspension actuators and the safety threshold of the suspension system to set dual hard constraint boundaries, a feasible and automotive-grade constrained control optimization target is generated. This ensures the physical feasibility of the control scheme while completely avoiding the safety risks of actuator saturation and suspension overtravel. A lightweight numerical algorithm is used to quickly solve the control optimization target to obtain a pre-controlled optimal control force sequence. This drives the suspension actuators to output the corresponding control force in advance before the road disturbance arrives, achieving feedforward cancellation of future road disturbances. Ultimately, this achieves a deep integration of active feedforward pre-control and real-time closed-loop feedback, significantly improving vehicle ride comfort and driving stability.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 A flowchart of a vehicle suspension control method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the functional modules of a vehicle suspension control device provided in an embodiment of the present invention; Figure 3 A schematic diagram of the hardware architecture of an electronic device provided in an embodiment of the present invention. Detailed Implementation

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

[0022] Research has found that it is currently difficult to achieve time-series adjustments under different road conditions (such as potholes, slopes, and bumps); existing control strategies mainly rely on feedback from the current instantaneous state and lack dynamic constraint mechanisms to prevent suspension saturation, hysteresis, or oscillation.

[0023] Based on this, the vehicle suspension control method and device provided in this embodiment of the invention can achieve early prediction and active cancellation of road disturbances by the suspension system through multi-sensor fusion spatiotemporal disturbance prediction, optimized control with hard constraints and adaptive weight adjustment, which greatly improves the vehicle's driving comfort and handling stability, while reducing the algorithm's computational complexity and system energy consumption, and can operate stably in real time on ordinary automotive-grade hardware.

[0024] To facilitate understanding of this embodiment, a vehicle suspension control method disclosed in this embodiment of the invention will first be described in detail.

[0025] Figure 1 This is a flowchart of a vehicle suspension control method provided in an embodiment of the present invention.

[0026] Reference Figure 1 The vehicle suspension control method may include the following steps: Step S102: Based on the multimodal feature vectors corresponding to multi-source sensor data and the temporal deep learning network, predict the road disturbance temporal matrix within the future time window; Among them, the multimodal feature vector refers to a standardized feature set that integrates different types and dimensions of sensor data. In this solution, it is specifically a multidimensional vector that integrates road environment image features, vehicle dynamic inertial features, and vehicle speed features. This can solve the problems of insufficient data dimensions from a single sensor and poor robustness in harsh environments. For example, by integrating road image data from a camera with vehicle vibration data from a 6-axis inertial measurement unit (IMU), information on the road conditions ahead and the current state of the vehicle can be obtained simultaneously. Temporal deep learning networks refer to deep learning network architectures specifically designed for time series data, which can effectively uncover the correlation patterns between different time series data. The temporal deep learning network in this solution is a hybrid cascaded architecture of temporal convolutional networks and long short-term memory networks, which is suitable for the scenario requirements of road disturbance temporal prediction. The road disturbance time series matrix is ​​a two-dimensional matrix used to characterize the vertical undulation amplitude of the road surface that each wheel of the vehicle will encounter at each time step within a future preset time window. The row dimension represents the continuous control time step, and the column dimension represents the four wheels of the vehicle. It is the core input basis for the active pre-control of the suspension. For example, a value of 50mm in the matrix represents a 50mm high road surface protrusion that the corresponding wheel will encounter at the corresponding time step. In this scheme, it is uniformly denoted as _____. , The length of the future time window; The future time window refers to the continuous time period after the current moment that allows for effective road condition prediction. In this solution, the typical value of the future time window is 0.5~1.0s, which provides sufficient response time for suspension control. For example, a prediction window of 0.8s can detect road disturbances about 22 meters ahead when the vehicle is traveling at a high speed of 100km / h.

[0027] This step addresses the environmental adaptability limitations of single sensors by fusing multi-source sensor data. Through a deep learning network specifically optimized for time-series data, it simultaneously mines the spatial distribution characteristics of the road surface and their temporal correlation patterns, enabling accurate early prediction of future road disturbances. This overcomes the limitations of traditional suspension feedback control logic, which relies on the principle of first experiencing impact and then compensating. It can predict road undulations ahead even at high speeds, allowing sufficient response time for active suspension control and fundamentally solving the problem of lag in traditional suspension control.

[0028] For example, step S102 can be refined through the following steps, including: Step 1.1: Simultaneously collect road environment data, vehicle dynamic data, and vehicle driving data through multi-source vehicle sensors, and extract multimodal feature vectors after performing preprocessing operations; Among them, the multi-source vehicle sensors include four core types of sensors: forward-looking monocular camera, millimeter-wave / LiDAR, 6-axis inertial measurement unit (IMU) and wheel speed sensor. These sensors respectively collect road environment data, vehicle dynamic data and vehicle driving data, achieving complementary fusion of multi-dimensional data and solving the perception limitations of a single sensor.

[0029] Here, multi-source sensors synchronously collect comprehensive data covering road conditions, vehicle status, and driving conditions. Preprocessing addresses issues such as inconsistent sampling frequencies and data noise from different sensors. Finally, a standardized feature vector integrating multi-dimensional information is extracted, solving the problems of poor robustness and easy failure of single sensor data in harsh environments. This achieves complementary fusion of multi-source data, ensuring that even if a single sensor fails, the system can still maintain a prediction accuracy of over 85%, significantly improving the algorithm's environmental robustness and adaptability to different operating conditions.

[0030] Step 1.2: Based on multimodal feature vectors, predict the road disturbance time series matrix within the future time window using a temporal deep learning network.

[0031] Here, the multimodal feature vectors extracted in the preceding sequence are used as input. Through a deep learning network specifically optimized for time series data, the spatial distribution pattern and time series correlation features of road disturbances are simultaneously mined. Finally, the continuous road disturbance time series matrix within the future time window is output, providing accurate pre-control basis for subsequent suspension control. It can achieve end-to-end accurate prediction of future road disturbances. The mean absolute error (MAE) of prediction on the test set is less than 3mm, which fully meets the accuracy requirements of active suspension control.

[0032] In practical applications, step S102 of this embodiment can be implemented through the following exemplary process, including: First, a forward-facing camera installed inside the rearview mirror on the inside of the windshield and a millimeter-wave / LiDAR sensor in the center of the front bumper simultaneously collect road environment data within a preset range of 30-50 meters in front of the vehicle, acquiring images and point cloud information of road surface undulations, potholes, and speed bumps. A 6-axis inertial measurement unit installed at the vehicle's center of gravity collects three-dimensional acceleration and angular velocity signals of the vehicle body, obtaining dynamic data reflecting the real-time dynamics of the vehicle. Wheel speed sensors installed at the four wheel hubs collect wheel speed data and calculate the vehicle's real-time speed, acquiring vehicle driving data, thus completing the synchronous acquisition of multi-source data. Then, standardized preprocessing operations are performed on the collected multi-source data, using timestamp alignment technology to separate different data sources. Sensor data with the same sampling frequency are uniformly synchronized to a reference frequency such as 100Hz to avoid prediction errors caused by data misalignment. Kalman filtering is used to filter out vehicle vibration noise in inertial measurement data, Gaussian filtering is used to remove environmental interference in image and point cloud data, and interpolation algorithms are used to complete sparse areas of radar point clouds. In addition, road image features, point cloud elevation features, vehicle inertial features, and driving speed features are extracted from the preprocessed multi-source data and integrated to obtain a standardized multimodal feature vector. Finally, the extracted multimodal feature vector is input into a pre-trained temporal deep learning network, which performs in-depth feature mining and temporal prediction, and finally outputs a road disturbance temporal matrix within a preset future time window.

[0033] In some embodiments, step 1.2 can be further refined by the following steps, including: Step 2.1: Input the multimodal feature vectors into the pre-trained temporal deep learning network; Among them, the temporal deep learning network is composed of a hybrid cascade of temporal convolutional networks and long short-term memory networks. The temporal convolutional network (TCN) is a convolutional neural network specifically optimized for temporal data. Through causal convolution and dilated convolution structures, it can efficiently extract local spatial features and short-term temporal dependencies of temporal data. It has the advantages of parallel computing and low latency, and is suitable for in-vehicle real-time prediction scenarios. The long short-term memory network (LSTM) is an advanced improvement of recurrent neural networks. Through the gating mechanism of input gate, forget gate and output gate, it solves the gradient vanishing problem of long sequence training in traditional recurrent neural networks. It can accurately mine the long-term dependency patterns of temporal data and is suitable for long-term road condition prediction needs.

[0034] This step clarifies the hybrid cascaded architecture of the temporal deep learning network, employing a layered design of TCN and LSTM instead of a single network structure. This fully leverages the complementary advantages of the two types of networks, addressing the limitation of a single network in simultaneously handling spatial feature extraction, short-term temporal capture, and long-term pattern discovery. The hybrid network architecture balances the efficiency of feature extraction with the accuracy of long-sequence prediction, while its parallel computing capabilities significantly reduce prediction latency, meeting the real-time operation requirements of 100Hz in automotive applications.

[0035] Step 2.2: Use a temporal convolutional network to extract the spatial distribution and short-term temporal dependency features of road disturbances, and then use a long short-term memory network to mine long-term temporal correlation features. Then, splice and fuse them to obtain a feature vector that maps road surface spatial features to vehicle driving time sequence. Here, the specific responsibilities of the hybrid network are clearly defined. The TCN is specifically responsible for extracting the spatial distribution features of road disturbances (such as speed bump width and pothole depth) and short-term temporal dependency features (such as the time interval between adjacent disturbances). The LSTM is specifically responsible for mining the long-term temporal correlation patterns of similar road conditions during historical driving. The two types of features are then concatenated and fused to achieve accurate mapping and alignment between spatial features and vehicle driving time sequence. This scenario-customized division of labor design enables the network to simultaneously possess accurate road spatial feature recognition capabilities and long-term temporal road condition prediction capabilities, significantly improving prediction accuracy under complex road conditions.

[0036] Step 2.3: Based on the sequence-to-sequence decoding structure in the Long Short-Term Memory network, the feature vector is used to generate an initial sequence of perturbation amplitude for each wheel at each time step within the future time window; Sequence-to-sequence (Seq2Seq) decoding structure refers to a network structure that can map the input historical feature sequence to the output future prediction sequence. It can achieve multi-step continuous prediction rather than single-point prediction, which is perfectly suited to the continuous temporal requirements of suspended pre-control. For example, it can output the continuous prediction values ​​of the next 80 time steps at once, corresponding to a prediction window of 0.8s.

[0037] Here, this step uses the sequence-to-sequence decoding structure of LSTM to directly decode the fused feature vector into an initial sequence of continuous disturbance amplitudes for each wheel and time step within the future time window. The output is a continuous prediction result that can be directly used for suspension control, rather than a single-point discrete prediction value, which differs from the limitation of existing technologies that can only achieve single-point prediction. It can directly output a continuous disturbance sequence covering the entire prediction time window, providing a complete timing basis for multi-step advance control of the suspension, perfectly matching the control frequency requirements of the suspension system.

[0038] Step 2.4: Combine the physical limits of vehicle suspension and the dynamic characteristics of vehicle body to constrain and standardize the initial sequence of disturbance amplitudes, and obtain the road disturbance time series matrix that meets the requirements of suspension control.

[0039] Here, by combining the physical limits of the suspension system and the dynamic characteristics of the vehicle body, outliers in the initial sequence that exceed the physical reasonable range are removed. At the same time, the sequence is standardized to unify the dimensions and time step intervals, and finally a standardized road disturbance time series matrix that fully meets the requirements of suspension control is generated. This can ensure the physical reasonableness and control adaptability of the prediction results, avoid the failure of subsequent control optimization solutions due to abnormal prediction values, and further improve the operational stability of the algorithm.

[0040] In some embodiments, the determination of the road disturbance time series matrix in step 1.2 can be implemented through the following example process, including: First, the preprocessed multimodal feature vectors are input into a pre-trained temporal deep learning network. This network employs a hybrid cascaded architecture of temporal convolutional networks and long short-term memory networks. The pre-training process utilizes both publicly available road datasets and a self-collected full-scene road condition dataset, with a total sample size exceeding 1 million, demonstrating strong generalization capabilities. Then, the temporal convolutional network layer at the network's front end extracts the spatial distribution features and short-term temporal dependencies of road disturbances from the multimodal feature vectors, capturing core information such as the physical dimensions of road surface undulations and the time intervals between adjacent disturbances. The temporal convolutional network uses a three-layer causal convolutional structure with kernel sizes of 3, 5, and 7, expanding the receptive field through increasing dilation rates while introducing residual connections to prevent gradient vanishing. Furthermore, the long short-term memory network layer at the network's back end further mines historical features based on the output of the temporal convolutional network. By analyzing long-term temporal correlation patterns in historical road condition data and capturing the recurring cycles of similar road conditions and vehicle response characteristics, a Long Short-Term Memory (LSTM) network with a two-layer structure and 128 hidden units per layer is used. A gating mechanism is employed to selectively retain key historical information. Finally, the features output from the two networks are concatenated and fused to obtain a fused feature vector that maps road spatial features to vehicle driving time sequence. Then, through a fully connected sequence-to-sequence decoding structure at the end of the LSM network, the fused feature vector is decoded into an initial sequence of disturbance amplitudes for each wheel at each control time step within a future time window. Finally, considering the physical limits of the vehicle suspension actuators and the vehicle's vertical dynamics, outliers in the initial sequence are constrained and corrected. Simultaneously, the sequence is standardized to unify the time step interval and data dimensions, ultimately generating a standardized road disturbance time sequence matrix that fully meets suspension control requirements.

[0041] Step S104: Construct the dynamic state equation of the suspension system based on the road disturbance time series matrix, set the constraint boundary by combining the physical limit and safety threshold of the suspension actuator, and generate the constrained control optimization objective; Among them, the dynamic state equation of the suspension system refers to the mathematical equation describing the dynamic relationship between the vertical motion state of the vehicle body and the output force of the suspension actuators under road disturbance. It is the core physical foundation of suspension control, and can accurately quantify the correspondence between the amount of force output by the actuators and the amount of vertical vibration generated by the vehicle body. The core equation is:

[0042] In the formula, For equivalent vehicle body mass, For equivalent suspension damping, For equivalent suspension stiffness, The vertical acceleration of the vehicle body. The vertical speed of the vehicle body. This refers to the vertical displacement of the vehicle body. For the output force of the suspension actuator, For tire stiffness, This is the time series matrix of road disturbances predicted in the previous example; Quadratic programming control optimization refers to a class of constrained mathematical optimization problems where the objective function is a quadratic function of the optimization variables, and the constraints are linear inequalities / equalities. In this scheme, it is used to solve for the optimal suspension control force that satisfies physical constraints. It has the advantages of fast solution speed and strong stability, making it suitable for real-time vehicle control scenarios. The core optimization objective is: Objective function: ; Constraints: ; ; In the formula, The vehicle body is used as a reference posture. To optimize the term weight coefficients, , These are the upper and lower limits of the actuator output force. The maximum permissible travel displacement of the suspension; the core characteristic of the vehicle's vertical attitude is the vertical displacement of the vehicle's center of gravity; The physical limits of a suspension actuator refer to the maximum tensile and thrust range that the actuator hardware itself can output. These are insurmountable physical boundaries to prevent actuator saturation and runaway. For example, in this solution, the output force limit of the electromagnetic active suspension actuator is -5000N to 5000N, where a negative sign represents tensile force and a positive sign represents thrust, corresponding to the constraints mentioned above. , ; The suspension safety threshold refers to the maximum allowable compression / tension stroke limit of the suspension mechanical structure. It is an inviolable safety boundary to prevent the suspension from overtraveling and touching the top / bottom, which could lead to component damage. For example, in this solution, the suspension travel safety threshold is ±80mm, corresponding to the constraints mentioned above. .

[0043] Here, the road disturbance time series matrix predicted in the aforementioned embodiment is used as the sole core input to construct a state equation that fits the actual dynamic characteristics of the vehicle, and to establish a precise mapping between the actuator output force and the vehicle body motion state. At the same time, dual hard constraints are set with hardware physical limits and driving safety requirements to construct a solvable and implementable control optimization target. Unlike the shortcomings of existing technologies that are unconstrained or have soft constraints that are prone to control failure, the constructed control optimization target takes into account vehicle stability, ride comfort and hardware safety. The dual hard constraints completely avoid the risks of suspension actuator saturation and overtravel, thus significantly improving the stability of the suspension system.

[0044] In some embodiments, step S104 may be refined to include the following steps: Step 3.1: Based on the road disturbance time series matrix, combined with the vertical dynamic characteristics of the vehicle suspension and the real-time transmitted operating state data, construct the suspension system motion state equation to characterize the dynamic mapping relationship between the output force of the suspension actuator and the vertical displacement, vertical velocity and acceleration of the vehicle body. Here, the road disturbance time series matrix predicted in the previous embodiment is used. Using the vehicle's inherent vertical dynamics as the core input, and incorporating real-time feedback of suspension operating status data, the equivalent vehicle mass in the equations is dynamically updated. Suspension damping Suspension stiffness The core parameter, the constructed state equation, directly defines the output force of the suspension actuator. Vertical motion state of the vehicle body , , The bidirectional dynamic mapping relationship between them distinguishes it from the poor adaptability of existing technologies that use fixed-parameter models. The constructed dynamic state equations can fit the vehicle's current real-time operating state, solving the problem of decreased control accuracy of fixed-parameter models when loads and operating conditions change, and providing a precise physical basis for subsequent control optimization.

[0045] Step 3.2: With minimizing the deviation between the vehicle's vertical attitude and the target reference attitude, and minimizing the vehicle's vertical motion speed as the dual optimization objectives, construct a secondary planning optimization objective covering future time windows; Here, based on the core state variables of the preceding state equations, the vehicle body attitude stability (corresponding to...) Minimize) and ride comfort (corresponding) Using minimization as the dual core, a quadratic programming optimization objective covering the entire forecast time window is constructed, and the integration interval of this quadratic programming optimization objective is... ~ The prediction window is completely consistent with the preceding road disturbance time series matrix, ensuring that the optimization objective fully covers all predicted road disturbances. The dual optimization objectives simultaneously take into account vehicle handling stability and ride comfort. The design covering the entire prediction window ensures that the optimization results can achieve disturbance cancellation throughout the entire time period, avoiding the control discontinuity problem caused by single-point optimization.

[0046] Step 3.3: In the quadratic programming optimization objective, superimpose the upper and lower limits of the output force corresponding to the physical limit of the suspension actuator and the suspension stroke displacement constraint corresponding to the safety threshold as double hard boundaries to generate a constrained control optimization objective for the suspension actuator.

[0047] Here, we add two insurmountable hard constraints to the quadratic programming optimization objective. The first constraint is the upper and lower limits of the actuator output force. The first layer ensures that the calculated control force is within the hardware's physical capabilities; the second layer is the suspension travel displacement constraint. This ensures that the vehicle body displacement remains within the safe suspension travel range, ultimately generating a constrained control optimization target that meets automotive-grade safety requirements. This dual hard constraint completely avoids the safety risks of actuator saturation and suspension overtravel, ensuring the physical feasibility and driving safety of the control scheme, unlike the shortcomings of existing soft constraints that are prone to control failure.

[0048] In practical applications, step S104 can be implemented using the following example flow, including: First, based on the road disturbance time-series matrix obtained from prior prediction, combined with the vertical dynamic characteristics of the vehicle's active suspension, and incorporating real-time transmitted data on the actual output force, travel, and vertical acceleration of the vehicle body, the core dynamic parameters of the suspension system—equivalent vehicle mass, equivalent suspension damping, and equivalent suspension stiffness—are identified and updated in real time, thus constructing the vertical dynamic state equation of the suspension system. This equation directly establishes a quantifiable two-way dynamic mapping relationship between the output control force of the suspension actuator and the vertical displacement, velocity, and acceleration of the vehicle body. Then, taking the minimization of the deviation between the vertical displacement of the vehicle body output by the completed dynamic state equation and the target reference attitude, and the minimization of the vertical velocity of the vehicle body as the dual core optimization objectives, a quadratic programming optimization objective function that fully covers the entire cycle of the future disturbance prediction time window is constructed. Simultaneously, adaptively adjustable weight coefficients are set for the two optimization terms to achieve a flexible balance between handling stability and ride comfort. Furthermore, the constructed quadratic programming optimization objective is superimposed with two insurmountable hard constraint boundaries: the first being the upper and lower limits of the output force corresponding to the physical limits of the suspension actuators. The second layer is the maximum permissible travel displacement constraint corresponding to the suspension safety threshold. Finally, a direct solvable optimization objective for suspension actuator control is generated with clearly defined hard constraint boundaries.

[0049] Step S106: Solve for the control optimization objective to obtain the optimal control force sequence, and drive the suspension actuator to perform disturbance cancellation control.

[0050] Here, a lightweight numerical algorithm is used to quickly solve the control optimization objective, resulting in a continuous control force sequence covering the entire prediction time window. This sequence drives the suspension actuator to output the corresponding control force in advance, thereby feeding forward to cancel out road disturbances that will arrive in the future. This achieves an active control effect that allows for advance action before the disturbance arrives and precise cancellation when the impact occurs, thus completely solving the lag problem of compensation after the impact of traditional suspension.

[0051] In some embodiments, the operation of solving the control optimization objective to obtain the optimal control force sequence in step S106 can be refined through the following steps, including: Step 4.1: Based on the dynamic state equation of the suspension system, perform approximate linearization processing at the current driving condition point of the vehicle to construct a linear time-varying model predictive control framework; Among them, the operating point refers to the set of vehicle operating state parameters that are directly related to the dynamic characteristics of the suspension system within the current control cycle, including real-time vehicle speed, real-time suspension travel, vertical motion state of the vehicle body, and vehicle load distribution, which are used to characterize the current operating conditions of the vehicle; for example, the vehicle is cruising at 100km / h at high speed and going over a speed bump at 30km / h, which correspond to two completely different operating points. Model predictive control (MPC) is a model-based closed-loop optimization control strategy that can predict the changes in system state over a future period based on the system model and solve for the optimal control sequence through rolling optimization. It is very suitable for constrained multivariable control scenarios such as suspension systems. In this solution, a lightweight linear time-varying MPC framework is optimized for vehicle scenarios.

[0052] Here, to address the nonlinear characteristics of the original suspension dynamics equations, a first-order Taylor expansion is performed at the vehicle driving condition point in the current control cycle to approximate linearization, transforming the nonlinear equations into a linear time-varying model. This constructs a lightweight model predictive control framework, avoiding the high computational overhead caused by nonlinear solutions. The approximate linearization process significantly reduces the computational complexity of the solution while ensuring control accuracy, laying the foundation for subsequent real-time solutions on the vehicle side.

[0053] Step 4.2: Transform the control optimization objective into a recursively solvable quadratic programming matrix form. Reuse the calculation results of the previous control cycle through the recursive matrix decomposition algorithm to complete the fast solution and obtain the optimal control force sequence of the suspension actuator covering the future time window.

[0054] Among them, recursive matrix decomposition refers to a numerical calculation method that reuses the core intermediate results of matrix decomposition from the previous control cycle during the quadratic programming solution process, and only performs low-complexity rank-1 correction on the numerically changed parts. This method avoids the high computational overhead caused by repeating full matrix decomposition in each cycle and is the core means of reducing computational complexity in this scheme.

[0055] Here, this step is the core of lightweight solution. It transforms the constrained control optimization objective into a standard quadratic programming matrix form. By using a recursive matrix factorization algorithm, it reuses the intermediate values ​​and optimal initial values ​​from the matrix factorization process in the previous control cycle. It eliminates the need for full matrix factorization and full iterative solution in each control cycle, thus achieving rapid solution of the optimization objective. Finally, it outputs the optimal control force sequence covering the entire prediction time window.

[0056] In practical applications, as an example, the operation of solving the control optimization objective to obtain the optimal control force sequence may include: First, based on the previously constructed nonlinear dynamic state equations of the suspension system, a first-order Taylor expansion is performed at the vehicle driving condition point corresponding to the current control cycle to approximate linearization, transforming the nonlinear dynamic equations into a linear time-varying model. This forms the basis for a lightweight model predictive control framework, avoiding the high computational overhead of nonlinear solutions. Then, the previously constructed control optimization objective with dual hard constraints is transformed into a standard quadratic programming matrix form that is universally applicable in vehicle control and can be recursively solved. Furthermore, through a recursive matrix factorization algorithm, the core intermediate quantities of matrix factorization generated during the solution process of the previous control cycle are reused. Simultaneously, the optimal solution result of the previous cycle is used as the initial iteration value for the current cycle, eliminating the need for full matrix factorization and full iteration calculations in each control cycle, significantly reducing computational overhead and enabling rapid solution of the quadratic programming problem. Finally, the optimal control force sequence of the suspension actuators, covering the entire future disturbance prediction time window and divided into time steps, is obtained.

[0057] In some embodiments, step S106 precedes the step of driving the suspension actuator to perform disturbance cancellation control, and the method further includes: Step 5.1: Identify the future disturbance level based on the road disturbance time series matrix, and dynamically adjust the weight coefficients of the deviation term between the vehicle vertical attitude and the target reference attitude and the vehicle vertical motion speed term in the control optimization target by combining the real-time vehicle speed, vehicle vertical attitude and suspension travel displacement in the multimodal feature vector. Here, the disturbance level of the future road surface is identified based on the road disturbance time series matrix. At the same time, combined with multi-dimensional operating parameters such as real-time vehicle speed, vehicle posture, and suspension travel, the weight coefficients of the two optimization terms in the quadratic programming optimization objective are dynamically adjusted through an adaptive weight adjustment formula. This approach achieves adaptive balance of control objectives under different operating conditions. Unlike existing technologies that use fixed weight coefficients and cannot adapt to all operating conditions, the core formula for adaptive weight adjustment in this scheme is: ; In the formula, Based on the weighting coefficient, This is the gain adjustment coefficient. For the vehicle's real-time speed, For reference speed.

[0058] This weight adjustment enables adaptive control under all operating conditions. In high-speed conditions, it prioritizes ensuring vehicle body stability, in low-speed conditions, it prioritizes reducing actuator energy consumption, and in large-disturbance conditions, it prioritizes offsetting impacts. This solves the problem of high energy consumption at low speeds and poor comfort at high speeds caused by fixed weights, thereby reducing the average energy consumption of actuators in real vehicle tests.

[0059] Step 5.2: Based on the adjusted weight coefficients, perform condition adaptation correction on the optimal control force sequence, and simultaneously add future disturbance gradient decay weights to generate the final control command with active pre-control attributes.

[0060] Here, the future disturbance gradient decay weight refers to the gradient weight set for the control force sequence based on the arrival time of the disturbance. The closer the disturbance arrives, the higher the weight, and the more accurate the control force output. The further the disturbance arrives, the lower the weight, which can effectively avoid the interference of long-distance prediction errors on the current control. For example, a weight of 1.0 is set for a disturbance arriving 0.1s later, and a weight of 0.3 is set for a disturbance arriving 0.8s later.

[0061] Here, based on dynamically adjusted weight coefficients, the optimal control force sequence is modified to adapt to the operating conditions. Simultaneously, a gradient decay weight for future disturbances is overlaid to gradient-adjust the control force at different time steps, ultimately generating a final control command with active pre-control attributes. This ensures the control command is both adapted to the current operating conditions and possesses strong anti-interference capabilities. The generated final control command is fully adapted to the vehicle's current driving conditions and future road disturbances. The gradient decay weight effectively reduces the impact of long-distance prediction errors on control accuracy, further improving the stability and precision of the control.

[0062] In practical applications, the operation of dynamically adjusting the weighting coefficients can be implemented using the following examples: First, based on the road disturbance time series matrix obtained from prior prediction, the road disturbance level within the future time window is identified, distinguishing different road condition levels such as straight roads, minor undulations, and large disturbances like speed bumps / potholes. Simultaneously, by combining real-time vehicle speed, vehicle vertical attitude data, and real-time suspension travel displacement data from the multimodal feature vector, the complete current driving condition of the vehicle is obtained. Then, based on the identified disturbance level and the vehicle's current driving condition, a preset adaptive weight adjustment formula is used. In the quadratic programming control optimization objective, the weight coefficients of the vehicle vertical attitude deviation term and the vehicle vertical motion velocity term are dynamically adjusted. Under high-speed conditions, the weight of the attitude deviation term is increased to ensure driving stability; under conditions with large disturbances, the weight of the vibration velocity term is increased to ensure ride comfort; and under low-speed, flat road conditions, the weights of both terms are reduced to decrease actuator energy consumption. Furthermore, based on the adjusted weight coefficients, the optimal control force sequence obtained from the previous solution is modified to adapt to the operating conditions. Simultaneously, a future disturbance gradient decay weight is added to the control force sequence, and lower weights are set for control forces corresponding to prediction steps further away from the current time step to avoid interference from long-distance prediction errors on the current control. Finally, a final control command for the suspension actuator with active pre-control attributes, adapted to the current operating conditions, is generated.

[0063] In some embodiments, the operation of driving the suspension actuator to perform disturbance cancellation control in step S106 includes the following detailed steps: Step 6.1: Based on the final control command corresponding to the optimal control force sequence, drive the suspension actuator to output control force in advance to feed forward and cancel out future road disturbances; Here, based on the final control command with pre-control attributes, the drive suspension actuator outputs a corresponding amount of control force before road disturbances reach the wheels, thus feeding forward to cancel out the upcoming road disturbances. This achieves an active control effect that anticipates disturbances and precisely cancels them out when the impact occurs, unlike existing technologies that only provide feedback compensation for impacts that have already occurred. This fundamentally solves the response lag problem of traditional suspension control, significantly reducing the impact on the vehicle body caused by road disturbances. In real-world testing, when passing over speed bumps, the peak vertical acceleration of the vehicle body decreased from 2.5 m / s² to 1.8 m / s², significantly improving ride comfort.

[0064] Step 6.2: By collecting real-time data on the actual output force, stroke, and vertical acceleration of the suspension actuator, the final control command is corrected in real-time using a closed-loop method to generate the final execution command to drive the suspension actuator. Here, the actual operating status data of the suspension is collected in real time through the sensor unit built into the actuator. The deviation between the actual operating status and the target status is compared, and the final control command is corrected in real time through a closed-loop control algorithm to generate the final execution command to drive the suspension actuator. This solves the problems of model error and actuator response deviation in feedforward pre-control. It realizes dual closed-loop control of feedforward pre-control and feedback correction, which not only ensures the anticipation of active control, but also ensures control accuracy, keeping the control error within 5%.

[0065] Step 6.3: Simultaneously, the operating status data is transmitted back to the construction stage of the dynamic state equation of the suspension system and the road disturbance prediction stage to complete the rolling iterative optimization of the entire system.

[0066] Here, real-time collected operational status data is fed back to the front-end dynamic state equation construction and road disturbance prediction stages. This data is used to update the core parameters of the dynamic equations in real time, and to incrementally train and optimize the disturbance prediction model, achieving a full-link rolling iterative closed loop for predictive control execution optimization. This allows the system's prediction accuracy and control performance to continuously optimize over time, further improving the algorithm's adaptability to different operating conditions and its long-term operational stability.

[0067] In practical applications, controlling the suspension actuator to cancel out disturbances can be implemented using the following examples: First, based on the final control command corresponding to the optimal control force sequence generated earlier, the suspension actuators corresponding to the four wheels of the vehicle are driven to output control forces of the appropriate magnitude in advance before future road disturbances reach the corresponding wheels, thus feeding forward to cancel out the upcoming road disturbances and preventing road impacts from being transmitted to the vehicle body at the source. Then, through the force sensors and displacement sensors built into the suspension actuators, as well as the inertial measurement unit at the center of gravity of the vehicle body, the actual output force of the suspension, the real-time travel of the suspension, and the vertical acceleration of the vehicle body are collected in real time. The deviation between the actual operating state and the target stable state is compared, and the final control command is corrected in real time through a closed-loop feedback control algorithm to generate a deviation-free final execution command, driving the suspension actuators to complete precise actions. In addition, the real-time collected suspension operating state data is synchronously transmitted back to the suspension system dynamic state equation construction stage to update the equivalent mass, damping, and stiffness core parameters of the equation in real time, and also transmitted back to the road disturbance prediction stage for incremental training and optimization of the prediction model. Finally, the entire system rolling iterative closed loop of predictive control execution optimization is completed, so that the control effect of the system is continuously optimized as the vehicle operates.

[0068] In a preferred embodiment of practical application, firstly, multiple types of onboard sensors simultaneously collect comprehensive data covering the road environment ahead, real-time vehicle dynamics, and vehicle driving conditions. After standardizing and preprocessing the collected multi-source data, multimodal feature vectors integrating multi-dimensional information are extracted, providing a high-quality input foundation for subsequent road disturbance prediction. Then, the preprocessed multimodal feature vectors are input into a pre-trained temporal deep learning network. The network deeply mines the spatial distribution patterns and temporal series correlation features of road disturbances to predict the road disturbance temporal matrix within a preset time window. This allows for accurate early detection of road disturbances ahead. Furthermore, based on the predicted road disturbance time-series matrix and the vertical dynamics of the vehicle suspension itself, a dynamic state equation for the suspension system is constructed, clarifying the mapping relationship between the actuator output force and the vehicle's vertical motion state. Simultaneously, considering the hardware physical limits of the suspension actuators and the travel safety threshold of the suspension system, dual insurmountable hard constraint boundaries are set, ultimately generating a constrained quadratic programming control optimization objective that meets automotive-grade control requirements. Finally, a lightweight numerical algorithm adapted to in-vehicle scenarios is used to quickly solve this control optimization objective, obtaining an optimal control force sequence for the suspension actuators covering the entire prediction time window. Based on this sequence, the suspension actuators are driven to output corresponding control forces, thus proactively offsetting upcoming road disturbances and achieving active pre-control of the vehicle suspension.

[0069] As an optional implementation, the hybrid model of temporal convolutional network and long short-term memory network in the embodiments of the present invention can be replaced with a temporal Transformer structure to further improve the prediction accuracy of long time series; Gaussian process regression can also be used to probabilistically model road disturbances to achieve uncertainty estimation of prediction results; the control constraint part can adopt a controller that combines reinforcement learning and quadratic programming to achieve self-learning adjustment of control parameters; for low-speed conditions such as off-road vehicles, an adaptive damping switching module can be introduced to dynamically change control parameters based on terrain type; for pure electric vehicles, an energy recovery mechanism can be combined to incorporate the feedback energy of the suspension actuator into the vehicle's electric drive system to further improve the overall vehicle energy efficiency.

[0070] In some embodiments, such as Figure 2 As shown, an embodiment of the present invention provides a vehicle suspension control device, comprising: The prediction module, based on multimodal feature vectors corresponding to multi-source sensor data and a temporal deep learning network, predicts the road disturbance time series matrix within a future time window; The generation module constructs the dynamic state equation of the suspension system based on the road disturbance time series matrix, sets constraint boundaries by combining the physical limits and safety thresholds of the suspension actuators, and generates a constrained control optimization objective. The control module solves the control optimization objective to obtain the optimal control force sequence, and drives the suspension actuator to perform disturbance cancellation control.

[0071] The present invention provides an embodiment for implementing an electronic device. In this embodiment, the electronic device may be, but is not limited to, a personal computer (PC), a laptop computer, a monitoring device, a server, or other computer device with analysis and processing capabilities.

[0072] As an exemplary embodiment, see [link to example]. Figure 3The electronic device 110 includes a communication interface 111, a processor 112, a memory 113, and a bus 114. The processor 112, the communication interface 111, and the memory 113 are connected via the bus 114. The memory 113 is used to store a computer program that supports the processor 112 in executing the above-described method. The processor 112 is configured to execute the program stored in the memory 113.

[0073] The machine-readable storage medium mentioned in this article can be any electronic, magnetic, optical, or other physical storage device that can contain or store information such as executable instructions, data, etc. For example, machine-readable storage media can be: RAM (Random Access Memory), volatile memory, non-volatile memory, flash memory, storage drives (such as hard disk drives), any type of storage disk (such as optical discs, DVDs, etc.), or similar storage media, or combinations thereof.

[0074] Non-volatile media can be non-volatile memory, flash memory, storage drives (such as hard disk drives), any type of storage disk (such as optical discs, DVDs, etc.), or similar non-volatile storage media, or combinations thereof.

[0075] It is understood that the specific operation methods of each functional module in this embodiment can be referred to the detailed description of the corresponding steps in the above method embodiment, and will not be repeated here.

[0076] The computer-readable storage medium provided in the embodiments of the present invention stores a computer program. When the computer program code is executed, it can implement the method described in any of the above embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0077] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0078] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0079] In the description of this invention, it should be noted that the terms center, up, down, left, right, vertical, horizontal, inner, and outer, indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms first, second, and third are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0080] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention.

Claims

1. A vehicle suspension control method, characterized in that, include: Based on multimodal feature vectors corresponding to multi-source sensor data and a temporal deep learning network, predict the temporal matrix of road disturbance within a future time window; Based on the road disturbance time series matrix, the dynamic state equation of the suspension system is constructed, and the constraint boundary is set by combining the physical limit and safety threshold of the suspension actuator to generate a constrained control optimization objective. Solving the control optimization objective yields the optimal control force sequence, which drives the suspension actuator to perform disturbance cancellation control.

2. The method according to claim 1, characterized in that, The steps for predicting the road disturbance time series matrix within a future time window based on multimodal feature vectors corresponding to multi-source sensor data and a temporal deep learning network include: Road environment data, vehicle dynamic data and vehicle driving data are collected synchronously by multi-source vehicle-mounted sensors, and multimodal feature vectors are extracted after preprocessing. Based on the multimodal feature vectors, a temporal deep learning network is used to predict the road disturbance time series matrix within a future time window.

3. The method according to claim 2, characterized in that, Based on the multimodal feature vectors, the step of predicting the road disturbance time series matrix within a future time window using a temporal deep learning network includes: The multimodal feature vectors are input into a pre-trained temporal deep learning network; wherein the temporal deep learning network is a hybrid of a temporal convolutional network and a long short-term memory network; The spatial distribution and short-term temporal dependence features of road disturbances are extracted using the temporal convolutional network, and then long-term temporal correlation features are mined based on the long short-term memory network. The features are then spliced ​​and fused to obtain a feature vector that maps road spatial features to vehicle driving time. Based on the sequence-to-sequence decoding structure in the Long Short-Term Memory network, the feature vector is used to generate an initial sequence of perturbation amplitude for each wheel at each time step within the future time window. By combining the physical limits of vehicle suspension and the dynamic characteristics of vehicle body, the initial sequence of disturbance amplitude is constrained, corrected and standardized to obtain a road disturbance time series matrix that meets the requirements of suspension control.

4. The method according to claim 1, characterized in that, The steps of constructing the dynamic state equation of the suspension system based on the road disturbance time series matrix, setting constraint boundaries by combining the physical limits and safety thresholds of the suspension actuators, and generating constrained control optimization objectives include: Based on the road disturbance time series matrix, combined with the vertical dynamic characteristics of the vehicle suspension and the real-time transmitted operating status data, a suspension system motion state equation is constructed to characterize the dynamic mapping relationship between the output force of the suspension actuator and the vertical displacement, vertical velocity and acceleration of the vehicle body. Minimizing the deviation between the vehicle's vertical attitude and the target reference attitude, and minimizing the vehicle's vertical motion speed are the two optimization objectives. A secondary planning optimization objective covering the future time window is constructed. The upper and lower limits of the output force of the suspension actuator under physical limit conditions and the suspension stroke displacement constraint corresponding to the safety threshold are superimposed on the quadratic programming optimization objective to generate a constrained control optimization objective for the suspension actuator.

5. The method according to claim 1, characterized in that, The steps to solve for the control optimization objective and obtain the optimal control force sequence include: Based on the dynamic state equation of the suspension system, an approximate linearization process is performed at the current driving condition point of the vehicle to construct a linear time-varying model predictive control framework. The control optimization objective is transformed into a recursively solvable quadratic programming matrix form. The calculation results of the previous cycle are reused by a recursive matrix decomposition algorithm to obtain the optimal control force sequence of the suspension actuator covering the future time window.

6. The method according to claim 1, characterized in that, Prior to the step of driving the suspension actuator for disturbance cancellation control, the method further includes: Based on the road disturbance time series matrix, the future disturbance level is identified. Combined with the vehicle's real-time speed, vehicle body vertical attitude, and suspension travel displacement in the multimodal feature vector, the weight coefficients of the deviation between the vehicle body vertical attitude and the target reference attitude and the vehicle body vertical motion speed in the control optimization target are dynamically adjusted. Based on the adjusted weighting coefficients, the optimal control force sequence is modified to adapt to different operating conditions, and the future disturbance gradient decay weights are superimposed to generate the final control command with active pre-control attributes.

7. The method according to claim 1 or 6, characterized in that, The steps of driving the suspension actuator for disturbance cancellation control include: According to the final control command corresponding to the optimal control force sequence, the suspension actuator is driven to output control force in advance to feed forward and cancel out future road disturbances. By collecting real-time operating status data of the suspension actuator, including actual output force, stroke and vehicle vertical acceleration, the final control command is corrected in real-time using a closed loop, and the final execution command is generated to drive the suspension actuator to move. Simultaneously, the operational status data is transmitted back to the stage of constructing the dynamic state equation of the suspension system.

8. A vehicle suspension control device, characterized in that, include: The prediction module, based on multimodal feature vectors corresponding to multi-source sensor data and a temporal deep learning network, predicts the road disturbance time series matrix within a future time window; The generation module constructs the dynamic state equation of the suspension system based on the road disturbance time series matrix, sets constraint boundaries by combining the physical limits and safety thresholds of the suspension actuators, and generates a constrained control optimization objective. The control module solves the control optimization objective to obtain the optimal control force sequence, and drives the suspension actuator to perform disturbance cancellation control.

9. An electronic device, characterized in that, It includes a memory, a processor, and a program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed, implements the method described in any one of claims 1-7.

Images