Vehicle control method, electronic device, and vehicle

Abstract

The application provides a vehicle control method, an electronic device and a vehicle, and relates to the technical field of vehicle shock absorption adjustment. The method comprises the following steps: acquiring terrain data and vehicle state data in the process of vehicle driving, and generating a shock risk score through a shock risk identification model based on the terrain data and the vehicle state data. A vehicle state vector is generated through a pre-trained autoencoder based on the vehicle state data. A stiffness adjustment value is generated through a pre-trained shock adjustment model based on the shock risk score, the vehicle state vector, the vehicle state data and a historical behavior feature vector corresponding to a current driver. The shock risk score, the vehicle state vector and the stiffness adjustment value are fused to obtain a control vector. The control vector is sent to a shock absorber control unit to adjust the stiffness of the shock absorber according to the control vector through the shock absorber control unit, so that the accuracy and timeliness of stiffness adjustment are realized, and a more comfortable and personalized shock absorption experience is provided for the driver.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle shock absorption adjustment, and in particular to a vehicle control method, an electronic device and a vehicle. BACKGROUND

[0002] Current mainstream vehicle shock absorption control systems mostly still rely on preset fixed rules or real-time feedback data from a single sensor. Such control strategies show obvious insufficient adaptability when facing changing road conditions, significantly different driving styles and diversified vehicle platforms. Due to the lack of prediction ability of the system for the road conditions ahead, the system can only respond passively after the bumping or impact occurs, resulting in unavoidable lag in adjustment action and relatively harsh parameter switching, lacking smooth transition. In complex scenes such as continuous undulating road or sudden potholes, this response mode is easy to cause unstable shock absorption effect, and even cause unnecessary shaking of the vehicle body, affecting the ride comfort and control safety. SUMMARY

[0003] Therefore, the present application aims to provide a vehicle control method, an electronic device and a vehicle to solve the problems of response lag, harsh adjustment and unstable shock absorption effect of vehicle shock absorption adjustment.

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

[0005] obtaining terrain data and vehicle state data in the vehicle driving process;

[0006] generating a shock risk score based on the terrain data and the vehicle state data through a pre-trained shock risk identification model;

[0007] generating a vehicle state vector based on the vehicle state data through a pre-trained autoencoder;

[0008] generating a stiffness adjustment value based on the shock risk score, the vehicle state vector, the vehicle state data and a current driver identifier through a pre-trained shock adjustment model;

[0009] fusing the shock risk score, the vehicle state vector and the stiffness adjustment value to obtain a control vector;

[0010] sending the control vector to a shock absorber control unit to adjust the stiffness of the shock absorber according to the control vector through the shock absorber control unit.

[0011] Optionally, the generating of the shock risk score based on the terrain data and the vehicle state data through the pre-built shock risk identification model comprises:

[0012] construct a vehicle state graph according to the terrain data and the vehicle state data;

[0013] input the vehicle state graph into the vibration risk identification model, and output the vibration risk score through the vibration risk identification model.

[0014] Optionally, the vehicle state graph is constructed according to the terrain data and the vehicle state data, comprising:

[0015] generating a terrain node according to the terrain data, and generating a state node according to the vehicle state data;

[0016] constructing a directed edge from the terrain node to the state node according to a physical dependency relationship;

[0017] generating the vehicle state graph based on the terrain node, the state node and the directed edge.

[0018] Optionally, the training method of the autoencoder comprises:

[0019] constructing an initial autoencoder;

[0020] collecting vehicle historical state data of the vehicle in multiple driving scenes, multiple driving styles and multiple vehicle types as first training samples respectively;

[0021] inputting the first training samples into the initial autoencoder, and outputting corresponding state vector samples through the initial autoencoder;

[0022] decoding the state vector samples through a decoder to obtain restoration samples;

[0023] adjusting parameters of the initial autoencoder by minimizing the difference between the restoration samples and the first training samples, to obtain a trained autoencoder.

[0024] Optionally, the training method of the vibration adjustment model comprises:

[0025] constructing an initial vibration adjustment model;

[0026] collecting historical terrain data and historical vehicle state data when different drivers drive the vehicle;

[0027] constructing a second training sample according to the historical terrain data and the historical vehicle state data;

[0028] dividing the second training sample into multiple training tasks, and the training task corresponds to a driver one by one;

[0029] Based on the plurality of training tasks, an initial shock adjustment model is trained by using a meta-learning method to obtain a trained shock adjustment model.

[0030] Optionally, the fusing of the shock risk score, the vehicle state vector and the stiffness adjustment value to obtain the control vector comprises:

[0031] According to historical terrain data and historical vehicle state data within a time length before the current time, weight coefficients corresponding to the shock risk score, the vehicle state vector and the stiffness adjustment value are respectively calculated;

[0032] Based on the weight coefficients, the shock risk score, the vehicle state vector and the stiffness adjustment value are weighted to obtain the control vector.

[0033] Optionally, the fusing of the shock risk score, the vehicle state vector and the stiffness adjustment value to obtain the control vector comprises:

[0034] A vehicle model identifier of a current vehicle is determined, and a preset weight proportion relationship corresponding to the vehicle model identifier is determined;

[0035] Based on the preset weight proportion relationship, the shock risk score, the vehicle state vector and the stiffness adjustment value are weighted to obtain the control vector.

[0036] Optionally, based on the vehicle state data, a vehicle state vector is generated by a pre-trained autoencoder, comprising:

[0037] The vehicle state data is quality evaluated to obtain a quality evaluation result;

[0038] According to the quality evaluation result, weights of channel data in the vehicle state data are determined, and the vehicle state data is preprocessed according to the weights of the channel data to obtain preprocessed vehicle state data;

[0039] Based on the preprocessed vehicle state data, a vehicle state vector is generated by a pre-trained autoencoder.

[0040] Based on the same inventive concept, a second aspect of the present application further provides an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable by the processor, wherein the processor implements the method as described above when executing the computer program.

[0041] Based on the same inventive concept, a third aspect of the present application further provides a vehicle comprising the electronic device as described in the second aspect.

[0042] It can be seen from the above that the vehicle control method, electronic device and vehicle provided by the application, wherein the method comprises: acquiring terrain data and vehicle state data in the vehicle driving process, and generating a vibration risk score based on the terrain data and the vehicle state data through a pre-trained vibration risk identification model. The vibration risk identification model can predict the vibration risk score reflecting the future terrain vibration risk according to the terrain data and the vehicle state data, thereby realizing the prediction of the influence of the future terrain on the vehicle vibration system and improving the timeliness of subsequent shock absorber adjustment. Based on the vehicle state data, a vehicle state vector is generated through a pre-trained autoencoder. The autoencoder can compress the dimension of a large amount of original data, thereby avoiding overload of the vehicle shock absorption system and improving the response rate of vehicle shock absorption adjustment. The autoencoder can also effectively extract features from the original data, thereby improving the control accuracy of shock absorption adjustment and avoiding harsh shock absorption adjustment. Based on the vibration risk score, the vehicle state vector, the vehicle state data and the current driver identifier, a stiffness adjustment value is generated through a pre-trained vibration adjustment model. The vibration risk score, the vehicle state vector and the vehicle state data are further combined with the current driver identifier, so that the stiffness adjustment value generated by the vibration adjustment model can meet the individualized shock absorption needs of the driver and meet the driving habits of different drivers. The vibration risk score, the vehicle state vector and the stiffness adjustment value are fused to obtain a control vector. The generation of the control vector realizes information complementation of output data of different models, eliminates conflicts between different data, ensures the stability and efficiency of the control vector, and is beneficial to vehicle shock absorption adjustment in complex scenarios. The control vector is sent to a shock absorber control unit to adjust the stiffness of the shock absorber according to the control vector through the shock absorber control unit, thereby realizing the accuracy and timeliness of stiffness adjustment and providing more comfortable and personalized shock absorption experience for the driver. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the present application or related art, the following will briefly introduce the drawings needed to be used in the embodiments or related art descriptions. Obviously, the drawings in the following description are only embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0044] Figure 1 The flowchart of the vehicle control method of the embodiment of the present application;

[0045] Figure 2 The structural schematic diagram of the vehicle control device of the embodiment of the present application;

[0046] Figure 3Fig. 1 is a schematic diagram of an electronic device hardware structure according to an embodiment of the present application. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to specific embodiments and drawings.

[0048] It should be noted that, unless otherwise defined, technical terms or scientific terms used in the embodiments of the present application should be understood as their common meanings to those skilled in the art. The terms "first", "second", and similar terms used in the embodiments of the present application do not denote any order, quantity, or importance, but are only used to distinguish different components. The terms "include", "contain", and similar terms mean that the elements or objects before the terms encompass the elements or objects listed after the terms and their equivalents, and do not exclude other elements or objects. The terms "connect" or "connected" and similar terms do not mean physical or mechanical connections, but can include electrical connections, whether direct or indirect. The terms "upper", "lower", "left", "right", and the like only represent relative positional relationships, which can change when the absolute positions of the described objects change.

[0049] As described in the background, the current mainstream vehicle damping control system still mostly relies on preset fixed rules or real-time feedback data from a single sensor. Such control strategies show obvious lack of adaptability when facing changing road conditions, significantly different driving styles, and diversified vehicle platforms. Due to the lack of prediction ability of the system for the road conditions ahead, it can only respond passively after the bump or impact occurs, resulting in unavoidable lag in adjustment actions, and the parameter switching is often harsh and lacks smooth transition. In complex scenes such as continuous undulating road or sudden potholes, this response mode is easy to cause unstable damping effect, and even causes unnecessary shaking of the vehicle body, affecting the ride comfort and control safety. The above problems mainly come from the following three aspects:

[0050] First, most existing technologies only react based on the instant vehicle body posture and wheel-ground contact information, and cannot obtain the characteristics of the road ahead, such as slope change and road unevenness. Without prediction of future road conditions, the suspension system cannot be pre-adjusted before action, resulting in a passive state when responding to continuous excitation or sudden impact, and causing adjustment lag.

[0051] Second, modern vehicles are equipped with a large number of sensors that can collect multi-dimensional signals such as acceleration, height, and pressure in real time. However, these data are often not effectively extracted and compressed for modeling, and are directly processed subsequently. Information overload makes it difficult for the system to make accurate decisions in a very short time, leading to delayed reaction and rough control, especially in complex working conditions, which can easily exceed its processing capacity.

[0052] Third, different drivers have different preferences for comfort and handling, and existing damping strategies are often set as a general setting, which cannot be adjusted adaptively according to the habits, body weight distribution, and common road conditions of specific drivers. Such control logic not only affects the personalization of the riding experience, but also restricts the performance accuracy of the vehicle in different use scenarios.

[0053] Therefore, the present application proposes a vehicle control method, which generates a shock risk score through a shock risk identification model to realize prediction of future impacts. The vehicle state vector is generated through a self-encoder to realize compression and effective feature extraction of the original data, avoiding system overload and improving damping adjustment accuracy. By introducing a driver identifier into the shock adjustment model, a stiffness adjustment value that matches the driving style and control needs of the driver can be generated to improve driving comfort. Finally, by fusing different data through a control vector, information complementation and data conflict elimination are realized to ensure the stability and efficiency of the control vector, which is conducive to the damping adjustment of the vehicle in complex scenarios. The control vector is sent to a shock absorber control unit to adjust the stiffness of the shock absorber according to the control vector, realizing the accuracy and timeliness of stiffness adjustment and providing a more comfortable and personalized damping experience for the driver.

[0054] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0055] The embodiments of the present application provide a vehicle control method applied to a vehicle damping control system, which refers to Figure 1 , comprising the following steps:

[0056] Step 102, acquiring terrain data and vehicle state data during vehicle driving.

[0057] Specifically, the terrain data can be collected by image sensors, which at least include a vehicle-mounted front-view camera and a laser radar. The vehicle-mounted front-view camera can record the road conditions in front of the vehicle in video, and the image frame rate is kept above 30 frames to ensure fine-grained recognition. The laser radar is responsible for constructing a three-dimensional terrain point cloud map to capture the accurate position and shape information of small undulations and obstacles.

[0058] The vehicle state data can be collected by accelerometers, gyroscopes, wheel speed sensors, and shock absorbers. Accelerometers are used to collect real-time vibration data of the vehicle in the vertical, lateral, and longitudinal directions. Gyroscopes synchronously collect the angular velocity changes of the vehicle to obtain the attitude information of the vehicle under different turning and slope conditions. Wheel speed sensors are used to monitor the wheel speed and its fluctuation characteristics, which indirectly reflect the road adhesion and pothole conditions. The shock absorber body integrates a displacement self-encoder to obtain the small changes in compression and rebound travel of the shock absorber, further assisting in analyzing the degree of terrain influence.

[0059] The above data collection frequencies are unified. For example, the data collection frequency can be 200 Hz to ensure that high-frequency vibration characteristics can be captured even in high-speed driving or extreme bumpy road sections. Some low-frequency signals such as GPS and map matching results will be aligned with high-frequency data through interpolation methods to form a unified data structure. All sensor data are synchronized and time-stamped through a central gateway to form a unified time-series data stream. The data collection module supports high-speed controller area network bus and Ethernet hybrid communication protocols to ensure that large bandwidth data can be transmitted in real time to the vehicle-mounted computing platform.

[0060] In addition to terrain data and vehicle state data, driver identification can also be synchronously collected to associate the driver's historical driving behavior characteristics. At the same time, vehicle static configuration parameters such as vehicle model identification, suspension structure type, current load, tire wear degree, tire model, etc. can also be recorded as static condition characteristics for subsequent model input.

[0061] The system integrates a high-precision map module to achieve sub-meter positioning accuracy in combination with the global navigation satellite system and the inertial navigation system. The sensor data collected for each path is bound to the terrain attributes, including but not limited to road grade, material type, slope, curvature, known pothole or repair area information.

[0062] The data collected during vehicle driving can be cached locally or uploaded to the cloud. The vehicle-mounted cache space needs to have a storage capacity of not less than ten hours of driving data, and the data is periodically encrypted and uploaded to the vehicle enterprise cloud training platform. To protect the privacy of the driver, all uploaded data needs to be desensitized to remove personal information and only keep the driving behavior feature code. The uploaded data needs to have complete meta-information index fields, such as collection time, location, weather, driver identification hash value, vehicle model identification, software version, sensor calibration state, etc.

[0063] The collected data also needs to be preprocessed, including denoising, normalization and time window segmentation. The time window segmentation operation can adjust the time window according to the input requirements of different models, for example, using a two-second sliding window with a sliding step of 0.5 seconds. Through the time window segmentation operation, multiple data segments with context association can be constructed for subsequent model training.

[0064] Through the data collection and preprocessing operation in this step, high-quality training samples and real-time data can be provided for subsequent models, which is beneficial to the vehicle to realize precise and reliable shock adjustment operation.

[0065] Step 104, based on the terrain data and the vehicle state data, a shock risk score is generated by a pre-trained shock risk identification model.

[0066] Specifically, the shock risk identification model in this embodiment is a graph neural network. After model training, the graph neural network can well learn the association between terrain and vehicle state. The graph neural network structure includes a multi-layer graph convolution network, which has the ability of local structure perception and context information fusion. During model inference, the latest collected terrain data and vehicle state data are input into the shock risk identification model every certain time length (such as 0.5s), and the shock risk score of the next time window is output by the shock risk identification model. The higher the shock risk score, the stronger the impact the vehicle will face, and the need to pre-press the shock absorber or reduce the equivalent stiffness (or reduce the damping coefficient) to absorb the possible impact force in advance. The lower the shock risk score, the vehicle will travel on a flat, smooth or only slightly bumpy road, at which time the current state of the shock absorber can be maintained or restored to the default state of the shock absorber. The shock risk identification model realizes terrain perception and risk forward analysis, making the shock adjustment proactive and predictive, and greatly reducing the hysteresis of the shock adjustment.

[0067] Step 106, based on the vehicle state data, a vehicle state vector is generated by a pre-trained autoencoder.

[0068] Specifically, the vibration situation faced by the vehicle during driving often contains complex dynamic structures, which include but are not limited to road unevenness, tire and ground impact transmission, inertia response caused by vehicle speed change, and steering disturbance caused by driving behavior. These factors jointly act on the vehicle body structure, causing the vehicle to produce continuous and nonlinear vibration behavior in three-dimensional space, and further generating vehicle state data. Therefore, the vehicle state data contains a large amount of high-frequency multi-channel sensor data, including but not limited to vehicle body vertical acceleration, lateral roll angle rate, wheel suspension compression length, shock absorber oil pressure change, electromagnetic valve opening degree feedback, pitch angle rate, four-wheel speed gradient, vehicle body yaw rate, longitudinal impact response, real-time feedback state of electronically controlled shock absorber, etc. These sensor data form high-dimensional time series data with large data volume. If the high-dimensional time series data is directly input into the subsequent model, it will cause excessive burden on the vehicle end computing resources. At the same time, there are highly redundant data and noise disturbance data in the sensor data, which may cause unstable subsequent shock adjustment operation and even vibration feedback failure.

[0069] Therefore, this step performs dimension compression and effective feature extraction on the vehicle state data by using the pre-trained autoencoder to obtain a low-dimensional state vector. The autoencoder is composed of multiple layers of neural networks. The first layer of neural network is used to receive all channels of vehicle state data, and the time window length of each channel is fixed to form a fixed shape input tensor. The neural network compresses the channel number and dimension layer by layer, and each layer of neural network abstractly represents the features extracted by the previous layer of neural network, and finally obtains a low-dimensional state vector (vehicle state vector) at the end of the autoencoder. The vehicle state vector can concisely represent the overall vibration response mode of the vehicle in this time period. The vehicle state vector not only retains the key vibration change trend, but also reflects the coupling relationship between the driver style and the vehicle structure. For example, under the same road conditions, the vehicle may produce different vibration intensities during driving by different drivers, and these vibration intensity differences are also reflected in the vehicle state vector.

[0070] The dimension of the vehicle state vector can be adjusted according to different vehicle models, control accuracy requirements and computing resources, and is usually between 16 and 32. Each vehicle state vector represents the complete vibration response information of the vehicle within two seconds, and the time window of the vehicle state vector generation can be adaptively adjusted according to the actual road conditions. Multiple consecutive vehicle state vectors can also be combined to build a time series as training data for the subsequent vibration adjustment model, to enhance the time context perception ability of the vibration adjustment model for vibration behavior.

[0071] The autoencoder not only plays a role in dimensionality reduction and compression of vehicle state data, but also captures key structural changes in the vehicle state data through its internal nonlinear feature learning capability, significantly improving the robustness and active control capability of the vehicle shock absorption control system under complex road and personalized driving conditions.

[0072] In step 108, based on the shock risk score, the vehicle state vector, the vehicle state data, and the current driver identifier, a stiffness adjustment value is generated by a pre-trained shock adjustment model.

[0073] Specifically, the dynamic performance of a vehicle under the control of different drivers is significantly different. In particular, when facing the same road conditions, the speed selection, direction correction, braking frequency, and throttle control method of different drivers will directly affect the body shock response. This individualized difference also significantly affects the response effect of the suspension system. For example, the same speed bump may not produce strong impact under the control of a conservative driver, but may cause significant compression and vehicle bouncing under the operation of an aggressive driver. Traditional shock absorption control strategies often only rely on road conditions and vehicle state for judgment, ignoring the influence of driving style on shock perception and adjustment requirements, resulting in poor user experience for some users, and even misoperation of the control logic. Therefore, the present application constructs a shock adjustment model that can quickly adapt to the behavior of individual drivers.

[0074] The shock adjustment model is trained by historical terrain data and historical vehicle state data when different drivers drive the vehicle, and adopts meta-learning, such as Model-Agnostic Meta-Learning (MAML) method. The shock adjustment model is a model that can quickly adapt to the behavior of individual drivers, and has the ability to quickly adapt to the shock preferences of new drivers. Through the shock adjustment model, the purpose of personalized shock absorption adjustment is achieved.

[0075] After the shock risk score, the vehicle state vector, the vehicle state data, and the current driver identifier are spliced, they are input into the shock adjustment model, and the stiffness adjustment value is output by the shock adjustment model. The stiffness adjustment value is a stiffness offset for adjusting the basic stiffness value of the vehicle shock absorber, so that the final shock absorber stiffness meets the expectations of the current driver.

[0076] The model structure of the shock adjustment model is a fully connected neural network, including an input layer, a hidden layer and an output layer. The input layer is used to receive the vector obtained by splicing the shock risk score, the vehicle state vector, the vehicle state data and the current driver identifier. The hidden layer adopts a multi-layer nonlinear mapping structure to extract individual difference features and behavior response relationships. The output layer is used to output the stiffness adjustment value. For example, the numerical range of the stiffness adjustment value is (-1, 1). When the stiffness adjustment value is negative, it indicates that the driver prefers a softer suspension, and the system should reduce the damping on the basis of the stiffness value. If the stiffness adjustment value is zero, it indicates that the basis stiffness value meets the driver's preference at this time, and no offset is needed. If the stiffness adjustment value is positive, it indicates that the driver prefers a harder suspension, and the system should increase the damping on the basis of the stiffness value.

[0077] Step 110, fusing the shock risk score, the vehicle state vector and the stiffness adjustment value to obtain a control vector.

[0078] Specifically, the shock risk score reflects the intensity of the terrain impact that the vehicle body may withstand in the future, the vehicle state vector reflects the current shock state and shock response trend of the vehicle, and the stiffness adjustment value reflects the user's shock adjustment preference. In this step, by fusing the above data, a control vector that can reflect the fusion decision is given. The fusion process not only ensures the information coordination between different data, but also significantly improves the ability of the vehicle shock control system to respond to complex scenarios, and constructs a control decision chain that can be dynamically adjusted and self-adaptively evolved.

[0079] In the fusion process, a weighted gating mechanism is used to dynamically adjust the weights of the shock risk score, the vehicle state vector and the stiffness adjustment value in the control vector, which is beneficial to the vehicle shock control system to maintain stable, efficient and personalized adjustment ability when facing complex roads.

[0080] The control vector is a fixed-length, multi-dimensional vector that includes a core instruction and instruction-related parameters. For example, the core instruction includes a target damping coefficient, which represents the target damping setting as a percentage of the maximum damping. The target damping coefficient can be calculated based on the base stiffness value, the stiffness adjustment value, and the shock risk score. Target damping coefficient = (base stiffness value + stiffness adjustment value) x risk weight coefficient, where the risk weight coefficient is calculated based on the shock risk score. The risk weight coefficient is a value greater than or equal to 1, which amplifies the sum of the base stiffness value and the stiffness adjustment value according to the shock risk score. The greater the shock risk score, the greater the risk weight coefficient. Risk weight coefficient = 1 + a x (shock risk score)^b, where a is a scaling factor and b is an exponential factor. For example, both a and b are adjustment parameters greater than 1. For example, when a = 2.0, b = 2.0, and the shock risk score = 0.2, the risk weight coefficient = 1 + 2.0 x (0.2)^2.0 = 1.08.

[0081] The instruction-related parameters can include a pre-charge parameter, an advance time parameter, an overshoot time parameter, an overshoot curve parameter, and a control source identifier. The pre-charge parameter is a scalar, which can be a Boolean value (0 or 1). When the pre-charge parameter is 0, it indicates that no pre-charge is needed. When the pre-charge parameter is 1, it indicates that standard pre-charge is performed. The advance time parameter indicates how far in advance of the impact the pre-charge operation should be performed. When the advance time parameter is 1200, it indicates that the pre-charge action should be completed within 1.2 seconds. The overshoot time parameter is a scalar in milliseconds, which defines the time required to switch from the current stiffness to the target stiffness. The greater the overshoot time parameter value, the smoother the overshoot. The overshoot curve parameter is used to define the rate curve of stiffness change, which can include linear overshoot, slow-in-slow-out overshoot, or immediate switch overshoot. The control source identifier indicates the main source of the current control vector, such as the shock risk identification model, the autoencoder, or the shock adjustment model.

[0082] In a specific example, assume that the vehicle predicts that it will pass over a pothole in 1.5 seconds and that the current driver prefers comfort. The adjustment action expressed by the control vector can be interpreted as "immediately start an action that reduces the damping to 35% within 100 milliseconds in an S-curve manner, and simultaneously start the shock absorber pre-charge program to ensure that the optimal cushioning stroke is achieved after 1.5 seconds to respond to the upcoming impact."

[0083] Step 112, send the control vector to the shock absorber control unit to adjust the stiffness of the shock absorber according to the control vector through the shock absorber control unit.

[0084] Specifically, the control vector is sent to the shock absorber control unit, and after the shock absorber control unit analyzes each parameter in the control vector, it is converted into a specific electrical signal to act on the magnetorheological fluid channel or electrically controlled valve in the electrically controlled shock absorber, controlling the stiffness change or response delay time to achieve adjustment of the stiffness of the shock absorber.

[0085] The control logic, response curve, electrical interface, and magnetorheological material type of the shock absorber on different vehicle models may differ. The adjustment range of the control vector is not linearly related to the change range of the actual stiffness, and a corresponding mapping model of the control vector needs to be established for each vehicle model to adapt to different shock absorbers and achieve accurate shock absorption adjustment.

[0086] The vehicle shock absorption control system has a millisecond-level response capability. For example, the vehicle shock absorption control system can perform fifty shock absorber stiffness adjustments per second, ensuring that the shock absorption experience remains continuous and smooth even when the road conditions change dramatically, the vehicle is driven at high speed, or the driving style suddenly changes.

[0087] Based on the steps 102 to 112, the embodiment provides a vehicle control method, wherein the method comprises obtaining terrain data and vehicle state data in the vehicle driving process, and generating a vibration risk score based on the terrain data and the vehicle state data through a pre-trained vibration risk identification model. The vibration risk identification model can predict the vibration risk score reflecting the future terrain vibration risk according to the terrain data and the vehicle state data, realize the prediction of the influence of the future terrain on the vehicle vibration system, and is beneficial to improving the timeliness of subsequent shock absorber adjustment. Based on the vehicle state data, a vehicle state vector is generated through a pre-trained autoencoder. The autoencoder can compress the dimension of a large amount of original data, avoid overload of the vehicle shock absorption system, and is beneficial to improving the response rate of vehicle shock absorption adjustment. The autoencoder can also effectively extract features from the original data, which is beneficial to improving the control accuracy of shock absorption adjustment and avoiding the problem of harsh shock absorption adjustment. Based on the vibration risk score, the vehicle state vector, the vehicle state data and the current driver identifier, a stiffness adjustment value is generated through a pre-trained vibration adjustment model. On the basis of the vibration risk score, the vehicle state vector and the vehicle state data, the current driver identifier is also fused, so that the stiffness adjustment value generated through the vibration adjustment model can meet the individualized shock absorption needs of the driver and meet the driving habits of different drivers. The vibration risk score, the vehicle state vector and the stiffness adjustment value are fused to obtain a control vector. The generation of the control vector realizes the information complementation of the output data of different models, eliminates the conflict between different data, ensures the stability and efficiency of the control vector, and is beneficial to the shock absorption adjustment of the vehicle in complex scenes. The control vector is sent to a shock absorber control unit to adjust the stiffness of the shock absorber according to the control vector through the shock absorber control unit, realizes the accuracy and timeliness of stiffness adjustment, and provides a more comfortable and personalized shock absorption experience for the driver.

[0088] In some embodiments, the vibration risk score is generated based on the terrain data and the vehicle state data through a pre-constructed vibration risk identification model, comprising:

[0089] constructing a vehicle state atlas according to the terrain data and the vehicle state data;

[0090] inputting the vehicle state atlas into the vibration risk identification model, and outputting the vibration risk score through the vibration risk identification model.

[0091] Specifically, the shock risk identification model in this embodiment is a graph neural network model (GNN), and the input of the graph neural network model is a graph structure. Therefore, in this step, first, a vehicle state graph is constructed in real time according to the terrain data and the vehicle state data. The vehicle state graph includes multiple nodes and edges. The multiple nodes are constructed according to the terrain data and the vehicle state data, and the edges between the nodes are constructed according to the relationship between the terrain and the vehicle state. The constructed vehicle state graph is input into the shock risk identification model, and forward propagation calculation is performed in the shock risk identification model. Through multiple rounds of graph convolution operations, each node aggregates the information of its neighbor nodes. Finally, the information of the entire vehicle state graph is converged into a global representation vector. The global representation vector is mapped to a scalar value, i.e., a shock risk score, by a regression layer of the shock risk identification model. The value range of the shock risk score can be 0-1. The larger the value, the stronger the impact that the vehicle will encounter in the future.

[0092] Through the shock risk identification model, the impact intensity that may occur in the future for a certain length of time is probabilistically predicted based on the currently perceived environment and vehicle state. When the impact intensity is large, the shock absorber can be adjusted in advance to realize active response of shock absorption adjustment and significantly reduce the hysteresis of shock absorption adjustment. In addition, the shock risk identification model can continuously identify road changes, which is conducive to subsequent model planning to obtain a smooth shock absorption operation adjustment strategy, avoids harsh switching of shock adjustment, and improves driving comfort.

[0093] In some embodiments, the vehicle state graph is constructed according to the terrain data and the vehicle state data, including:

[0094] generating terrain nodes according to the terrain data and generating state nodes according to the vehicle state data;

[0095] constructing directed edges from the terrain nodes to the state nodes according to the physical dependency relationship;

[0096] generating the vehicle state graph based on the terrain nodes, the state nodes, and the directed edges.

[0097] Specifically, the nodes in the vehicle state graph represent specific terrain segments or vehicle dynamic states, and the edges describe the physical correlation and causal transmission relationship between two nodes. Nodes include terrain nodes and state nodes. For example, terrain nodes include pothole areas, speed bumps, manhole covers, gravel roads, slopes, and continuously undulating roads, which are generated by image recognition, laser radar point cloud analysis, and map annotation recognition, and have corresponding geographic coordinates and time sequence positions. The attributes of the terrain nodes can include slope value, texture complexity, point cloud density, type label, etc. State nodes include shock absorber compression distance, vehicle body vertical acceleration, vehicle body pitch angle rate, tire pressure change, vehicle speed change rate, etc. The attributes of the state nodes can include average acceleration in the last n seconds, maximum impact force, wheel speed change, etc.

[0098] According to the physical dependence relationship, directed edges are constructed from terrain nodes to state nodes, and the edges define the causal relationship, time sequence, or physical interaction between nodes. For example, the vehicle first passes through the speed bump (terrain node), and then the vehicle body is lifted (vehicle state node), forming a directed edge from the "speed bump" to the "vehicle body lift". For another example, a pothole (terrain node) will cause tire impact (vehicle state node), which in turn causes vehicle body vibration (another vehicle state node).

[0099] Each edge has a weight that quantifies the strength of the relationship between nodes. The edge weight can be obtained by statistically analyzing the average physical response amplitude between events. For example, a certain type of pothole usually causes an average shock absorber compression of 20 mm. Therefore, "20 mm" can be used as the initial weight of the edge. In the training of the graph neural network, the edge weight is continuously optimized. The relationship between nodes can be maintained by a dynamic adjacency matrix, which is updated in time sequence and constructs a graph snapshot sequence of the vehicle in different time periods, facilitating subsequent time series modeling and prediction. According to all the terrain nodes, state nodes, and directed edges obtained in real time, a vehicle state graph is generated. By inputting the vehicle state graph into the shock risk identification model, the corresponding shock risk score can be predicted.

[0100] In addition, different vehicle models have significant differences in structural layout, curb weight, chassis configuration, suspension type, and electronic control system architecture. These differences determine that the shock response behavior of the vehicle under the same road conditions can be completely different. Therefore, in order to ensure that the system can be migrated between different vehicle models without retraining or major modifications, vehicle model-related features need to be introduced when constructing the vehicle state graph. Specifically, the vehicle model identifier and the chassis parameter vector are introduced as auxiliary features when constructing the vehicle state graph, the edge connection between terrain nodes and state nodes is adjusted, and the expression ability of the vehicle state graph for different vehicle structures is increased, so that the shock risk score derived from the same vehicle state graph in different vehicle models has vehicle model difference expression ability.

[0101] The training process of the shock risk identification model includes: constructing an initial graph neural network model, and when training the initial graph neural network model, the purpose is to let the initial graph neural network model learn to extract a representative feature model from a historical vehicle state graph, establish a mapping relationship between terrain changes and vehicle state responses, generate a shock risk identification model, and provide data support for subsequent shock absorption adjustment. The initial graph neural network model adopts a multi-layer graph convolution network architecture, and each layer updates the state of the center node through an information aggregation mechanism of adjacent nodes, and has the ability of local structure perception and context information fusion.

[0102] The training sample of the initial graph neural network model includes a vehicle historical state graph and a corresponding historical shock risk score. The training sample includes historical state graphs collected under multiple drivers, multiple geographical areas, multiple weather and road conditions, so that the shock risk identification model after training has strong generalization ability across regions and driving behaviors. After training is completed, the shock risk identification model can identify typical terrain-state causal chains, such as "gravel→small vertical shock→increased shock frequency", and "speed bump→shock absorber compression→vehicle body pitch". Similar subgraph structures in the vehicle historical state graph will be identified frequently during the training process, forming structural memory within the model, so that future predictions can be made quickly when similar scenarios are encountered.

[0103] In addition, the shock risk identification model is dynamically updated as the vehicle continues to travel, and the system continuously collects new terrain data and vehicle state data to form a new vehicle state graph to adjust the parameters of the shock risk identification model. The vehicle state graph of each vehicle in different environments will be stored and uploaded, and the vehicle manufacturer backend will form a large amount of structured vehicle state graph, which is used for subsequent retraining and optimization of the shock risk identification model, and further improves the stability and precision of the shock risk identification model in complex road conditions.

[0104] Through the construction of the vehicle state graph in the embodiment, the physical causal relationship between various terrains and vehicle states is mined, so that the shock risk identification model can learn this physical causal relationship, and then make reasonable predictions of shock risk according to this physical causal relationship, and enhance the generalization ability of the shock risk identification model.

[0105] In some embodiments, the training method of the autoencoder includes:

[0106] constructing an initial autoencoder;

[0107] collecting vehicle historical state data of the vehicle under multiple driving scenarios, multiple driving styles, and multiple vehicle types as first training samples, respectively;

[0108] inputting the first training sample into the initial autoencoder, outputting a corresponding state vector sample through the initial autoencoder;

[0109] decoding the state vector sample through a decoder to obtain a restored sample;

[0110] adjusting parameters of the initial autoencoder by minimizing the difference between the restored sample and the first training sample to obtain a trained autoencoder.

[0111] Specifically, first, an initial autoencoder is constructed. The first training sample for training the initial autoencoder is derived from a large amount of vehicle historical state data under real roads, covering typical driving scenes such as expressways, urban roads, mountain roads, gravel roads, speed bumps, and potholes. At the same time, it covers vehicle historical state data under different driving styles, such as aggressive handling, smooth driving, overtaking acceleration, and frequent braking. Since the structural parameters of different vehicle models also affect the vehicle vibration response, the first training sample also includes vehicle historical data of different vehicle models. The structural parameters include suspension form, wheelbase length, track width, vehicle center of gravity height, etc. During the training of the initial autoencoder, introducing vehicle identification can improve the adaptability of the autoencoder to different vehicle models.

[0112] The first training sample is input into the initial autoencoder, and the vehicle historical state data in the first training sample is extracted and compressed layer by layer through the multi-layer neural network in the initial autoencoder to obtain a low-dimensional state vector sample. Then, the state vector sample is decoded through a decoder to obtain a restored sample. The decoder structure is symmetrical to the initial autoencoder structure and is used to restore the state vector sample generated by the initial autoencoder. By comparing the difference between the restored sample and the first training sample, it can be evaluated whether the core data in the first training sample is lost during the encoding process of the initial autoencoder, and the encoding strategy of the initial autoencoder is continuously optimized. The parameters of the initial autoencoder are adjusted to ensure that each dimension of data in the state vector sample contains meaningful information and continuously improve the compression understanding ability of the autoencoder. When the difference between the restored sample and the first training sample is less than a preset threshold, a trained autoencoder is obtained. During the training of the initial autoencoder, there is no need for manual labeling of training samples, and the training can be completed only according to the collected vehicle historical state data, reducing the training cost.

[0113] After the training is completed, only the autoencoder is retained, and the decoder is discarded. After the autoencoder is deployed to the vehicle, the autoencoder can be run in real time in the on-board computing platform, and the latest vehicle state data is encoded every fixed time length (such as 50 ms) to generate a vehicle state vector. Since the dimension of the vehicle state vector is much lower than that of the vehicle state data, the damping adjustment delay is greatly shortened, and the vehicle damping control response is more timely. At the same time, the stability of the encoded vehicle state vector is better than that of the vehicle state data, and it can maintain high interpretability and prediction stability even in the presence of severe interference or slight sensor abnormalities, providing a high-quality feature basis for subsequent model decision-making.

[0114] In some embodiments, the training method of the shock adjustment model comprises:

[0115] constructing an initial shock adjustment model;

[0116] collecting historical terrain data and historical vehicle state data when different drivers drive the vehicle;

[0117] constructing a second training sample according to the historical terrain data and the historical vehicle state data;

[0118] dividing the second training sample into multiple training tasks, one training task corresponding to one driver;

[0119] training the initial shock adjustment model based on the multiple training tasks using a meta-learning method to obtain a trained shock adjustment model.

[0120] Specifically, first, an initial shock adjustment model is constructed, which is a fully connected neural network model. Historical terrain data and historical vehicle state data when different drivers drive the vehicle are collected to construct a second training sample. The second training sample includes shock risk score samples, vehicle state vector samples, historical vehicle state data, driver identification samples, and stiffness adjustment value samples. The second training sample is divided into multiple training tasks according to the driver identification, with each training task corresponding to one driver identification. Each training task is divided into a support set and a query set, and part of the data in the training task is extracted as the support set, and the other data is taken as the query set. Through the support set, the initial shock adjustment model is updated in the inner layer, i.e., the fast gradient update is performed, and then the performance of the updated initial shock adjustment model is evaluated using the query set, the loss is calculated, and the outer layer update is performed on the initial shock adjustment model to optimize the general model parameters, obtaining a trained shock adjustment model. This training method enables the shock adjustment model parameters to have good initialization capability on multiple training tasks. After deployment in a new vehicle environment, only terrain data and vehicle state data collected by the driver driving the vehicle in a short period of time are needed to quickly fine-tune the model parameters, and then the stiffness adjustment value conforming to the current driving style of the driver is output.

[0121] In addition, in order to improve the rapid adaptation of the shock adjustment model to the vehicle model, a vehicle model platform coding mechanism can also be introduced in the training process of the shock adjustment model, so that the meta-learning process not only adapts to the differences of different drivers, but also is compatible with the differences of different vehicle models. For example, the suspension parameters of different vehicle models are different, and the shock response characteristics are different. The vehicle model platform coding captures the structural differences of different vehicle models, so that the stiffness adjustment value generated by the same driving behavior under different vehicle models is different. Through this mechanism, the universality of the shock adjustment model under the multi-vehicle platform of the vehicle enterprise is enhanced, and the deployability and expansibility of the system are improved.

[0122] When deployed, the vehicle is pre-installed with a general shock adjustment model in the offline stage. When the user drives for the first time, the vehicle system automatically collects the terrain data and vehicle state data generated by the driver driving the vehicle, and completes the local rapid fine-tuning of the shock adjustment model in the background. Generally, after the vehicle drives for a period of time for the first time, the shock adjustment model can generate stable and reliable stiffness adjustment values, and the shock adjustment model parameters at this time are cached as individual adjustment parameters in the vehicle-mounted controller. After each vehicle power-on, the system automatically loads the individual adjustment parameters corresponding to the current driver, and generates the stiffness adjustment value preferred by the current driver in real time through the shock adjustment model. If the driver changes or the driving style changes significantly, the system will trigger the local update process of the shock adjustment model again, to ensure that the shock adjustment model always fits the real intention of the driver.

[0123] The vehicle enterprise can collect large-scale vehicle use data from different users through a remote platform, and periodically update the general shock adjustment model. Each time the remote update is performed, the upgrade package contains the improved meta-parameters of the shock adjustment model and the newly added vehicle model platform adaptation strategy, so that the individual ability iteration can be completed without user intervention, forming a continuous self-evolution process. With the accumulation of data, the vehicle shock absorption control system can realize the perception of more complex behavior patterns, such as the auxiliary identification of non-linear behaviors such as fatigue driving and emotional driving, and further expand to the active safety control system, forming a complete driver state perception and intervention mechanism.

[0124] By introducing meta-learning in the model training process of the present embodiment, a driver-centered shock adjustment model that can adapt itself is constructed, so that the vehicle shock absorption control system can be flexibly adjusted when facing diversified drivers, improving the dynamic performance of the whole vehicle while taking into account the individual driver preferences, significantly enhancing the intelligence, flexibility and accuracy of the driving experience.

[0125] In some embodiments, the fusing of the shock risk score, the vehicle state vector, and the stiffness adjustment value to obtain a control vector comprises:

[0126] According to the historical terrain data and the historical vehicle state data within a time period before the current time, weight coefficients corresponding to the vibration risk score, the vehicle state vector and the stiffness adjustment value are respectively calculated; and the vibration risk score, the vehicle state vector and the stiffness adjustment value are weighted based on the weight coefficients to obtain the control vector.

[0127] Specifically, the weight coefficients are dynamically changed and are calculated according to the historical terrain data and the historical vehicle state data within a time period before the current time. Through the historical terrain data and the historical vehicle state data, the accuracy of the vibration risk score, the accuracy of the vehicle state vector and the accuracy of the stiffness adjustment value can be respectively evaluated. The weight coefficients corresponding to the vibration risk score, the vehicle state vector and the stiffness adjustment value are adjusted according to the accuracy evaluation results.

[0128] For example, if it is determined that the accuracy of the vibration risk score generated by the vibration risk identification model within a time period in the past decreases, the weight coefficient of the vibration risk score is correspondingly reduced, and the weight coefficients corresponding to the vehicle state vector and the stiffness adjustment value are adaptively increased. For another example, if the vehicle state vector generated by the autoencoder jumps sharply, it may indicate that the sensor is disturbed and the vehicle state vector is inaccurate, and the weight coefficient of the vehicle state vector is correspondingly reduced, and the weight coefficients corresponding to the vibration risk score and the stiffness adjustment value are adaptively increased. For another example, if the driver frequently manually adjusts the shock absorption mode after the shock absorber control unit executes a certain control vector, it indicates that the user is not satisfied with the active shock absorption adjustment of the vehicle shock absorption control system, and the weight coefficient corresponding to the stiffness adjustment value is correspondingly reduced, and the weight coefficients corresponding to the vibration risk score and the vehicle state vector are adaptively increased.

[0129] The vibration risk score, the vehicle state vector and the stiffness adjustment value are weighted according to the weight coefficients to obtain a final control vector. Through the dynamic weight coefficient adjustment mechanism of the embodiment, the vehicle shock absorption control system can intelligently allocate the attention of each model according to the current environment and vehicle state, make a reasonable decision, and generate a reasonable control vector. When a certain model is temporarily invalid (such as prediction error of the vibration risk identification model or failure of a certain sensor), the dynamic weight coefficient mechanism reduces the influence of unreliable data sources, and ensures the basic performance and safety of the vehicle shock absorption control system.

[0130] In some embodiments, the fusing of the vibration risk score, the vehicle state vector and the stiffness adjustment value to obtain the control vector comprises:

[0131] A vehicle model identifier of a current vehicle is determined, a preset weight proportion relationship corresponding to the vehicle model identifier is determined, and the vibration risk score, the vehicle state vector and the stiffness adjustment value are weighted based on the preset weight proportion relationship to obtain the control vector.

[0132] Specifically, in addition to dynamically calculating each weight coefficient through the foregoing embodiments, a corresponding weight configuration file can also be pre-set according to different vehicle models. The weight configuration file specifies the preset weight proportion relationship between the shock risk score, the vehicle state vector and the stiffness adjustment value under the current vehicle model. The weight configuration file is uploaded by the test center according to the real vehicle calibration results before the vehicle is delivered, and the vehicle shock absorption control system loads the corresponding weight configuration file by identifying the vehicle model. For example, for a vehicle model including a high-order suspension platform, it may rely more on the shock risk score output by the shock risk identification model, and therefore the preset weight proportion of the shock risk score is higher. While the entry-level vehicle model focuses more on the vehicle state vector and the stiffness adjustment value, and therefore the preset weight proportion of the vehicle state vector and the stiffness adjustment value is higher. By distributing the weight proportion coefficient through the weight configuration file, not only is the vehicle model platform customized control realized, but also the weight configuration file can be fine-tuned and optimized through remote upgrading in the later stage, without affecting the stability of the core algorithm structure.

[0133] In some embodiments, based on the vehicle state data, a vehicle state vector is generated by a pre-trained autoencoder, comprising:

[0134] The quality of the vehicle state data is evaluated to obtain a quality evaluation result;

[0135] The weight of each channel data in the vehicle state data is determined according to the quality evaluation result, the vehicle state data is preprocessed according to the weight of each channel data, and preprocessed vehicle state data is obtained;

[0136] Based on the preprocessed vehicle state data, a vehicle state vector is generated by a pre-trained autoencoder.

[0137] Specifically, the types, quantities, arrangement methods and resolutions of sensors mounted on different vehicle models differ significantly. For example, high-end vehicles are equipped with high-precision inertial measurement units and laser radars, which can capture extremely weak terrain undulations and construct accurate point cloud maps, while economy-type vehicles may only rely on basic accelerometers and camera images. If the autoencoder trained on a high-end vehicle is directly applied to an economy vehicle, the performance of the autoencoder will decrease significantly due to the sharp drop in the quality of the collected vehicle state data, resulting in poor accuracy of the generated vehicle state vector and invalidation of subsequent shock absorption adjustment.

[0138] To solve this problem, multiple input channels are set in the autoencoder, which can meet the data collected by high-end sensors of all possible vehicle models. After the autoencoder receives the vehicle state data, quality assessment of the vehicle state data is needed, mainly to assess the number of channels involved in the vehicle state data and the quality of the data transmitted in each channel, and to determine the weight of each channel data according to the channel data involved and the data quality. The channel with high transmission data and high transmission data quality is assigned a high weight, and the channel with no transmission data or low transmission data quality is assigned a low weight, or the weight is directly set to zero. According to the weight assigned to each channel, the vehicle state data transmitted by each channel is preprocessed, that is, the weight of the data in the subsequent encoding process is increased or decreased according to the weight response.

[0139] For example, if a laser radar is configured in the current vehicle model, the weight of the channel for transmitting the laser radar point cloud feature will be increased accordingly, indicating that this channel belongs to a reliable information source. For channels with no transmission data or poor transmission data quality, the corresponding weight is reduced or set to zero to avoid introducing noise or false information.

[0140] In addition, for sensor data missing in the current vehicle model, the system can also complete the missing data according to the estimation method, and then input the completed data into the autoencoder to maintain the integrity of the channel data.

[0141] Further, for the problem of sensor noise, the system can dynamically adjust the corresponding filtering algorithm according to the sensor type. For sensors with high noise, a stronger filtering algorithm is used, and for sensors with low noise, a weaker filtering algorithm is used to retain more real data. The vehicle state data after weighting and filtering is input into the autoencoder to generate a vehicle state vector. The method of the embodiment can ensure the consistency of the vehicle state vector output by the autoencoder, and provide a high-quality data basis for subsequent models.

[0142] It should be noted that the vibration risk identification model, the autoencoder, and the vibration adjustment model involved in the vehicle shock control system are all deep learning structures, and their performance is highly dependent on the quality of the training data, the diversity of the sample distribution, the setting of the hyperparameters, and the feedback mechanism of the evaluation indicators. In order to obtain stable and reliable modeling effect in the vehicle-grade product, the whole model training and parameter adjustment process must be systematically designed and verified multiple times to ensure that each type of model not only has the best performance alone, but also can still work cooperatively in a multi-model fusion structure, outputting a control vector with strong consistency, low response delay, and strong generalization ability.

[0143] The training data is provided by a vehicle enterprise test vehicle collection platform, covering various main pin vehicle platforms, multiple suspension structure variants, typical urban roads, highways, rural roads, complex work area scenarios, and is labeled according to different driving behavior styles. Each vehicle collects sensor data according to a unified standard to form a structured training sample library. The training sample division uses a proportional splitting method, dividing the data into a training set, a validation set, and a test set according to the time sequence, with a ratio of 7:1.5:1.5, ensuring that the samples are independent in the time dimension and avoiding future information leakage. The sample format is a unified multi-modal time series structure, facilitating the sharing of vibration risk identification models, autoencoders, and vibration adjustment models.

[0144] The entire training process is completed on a vehicle enterprise backend big data platform, supporting distributed parallel training and using GPU resources for high-frequency training and iteration of large-scale samples. During the training phase, a multi-vehicle model labeling system is established, and sample data is labeled according to vehicle platforms. Vehicle information is introduced as a filtering condition during the data loading phase. The model automatically learns the response patterns under different vehicle models during the training process, enabling hierarchical abstraction of cross-platform vibration behavior. In the later training phase, a transfer learning mechanism is introduced. In the case of a small amount of sample data for a certain vehicle model, fine-tuning can be performed on a small amount of target vehicle data by freezing part of the model layers, achieving rapid adaptation and efficient convergence, avoiding the need to train each vehicle model from scratch, and saving data and time costs.

[0145] After each training is completed, multiple model versions are output, and the performance of each model is tested on a specific evaluation set. The vibration risk identification model uses mean absolute error, peak error, and path consistency as evaluation indicators. The autoencoder uses reconstruction error curve and feature distribution stability as evaluation indicators. The vibration adjustment model uses task adaptation rate, parameter convergence step number, and final control response consistency as evaluation indicators. The three models are jointly verified during the fusion process to ensure that the combined model remains stable and accurate.

[0146] The model testing process needs to include different driving styles, different road conditions, and different vehicle models in cross combinations, covering extreme working conditions and boundary behaviors, including severe acceleration, emergency turning, intermittent potholes, gravel slopes, heavy load turning, and other scenarios. The system needs to stably output controllable responses under all test conditions to meet the requirements of vehicle adjustment consistency. The fused control vector is sent to a virtual simulation platform for vehicle dynamic response evaluation during the test phase, verifying the adjustment delay, vibration suppression capability, comfort indicators, and maximum suspension travel utilization rate, and other key performance parameters. The results are used to assist in the optimization direction of the parameter tuning process.

[0147] After training, each model will be quantized and compressed into an embedded version, significantly reducing memory usage and computational resource consumption without affecting prediction performance, facilitating deployment to the vehicle domain controller platform. The inference speed of each model must pass the vehicle standard verification, with a control response time of no more than fifty milliseconds, and the overall inference time of all models must be kept below one hundred milliseconds to ensure real-time shock absorption adjustment. All trained models and parameter configurations will be recorded and archived for subsequent version tracing, model hot updating, and cloud retraining, forming a closed-loop controllable adjustment and maintenance mechanism to support long-term sustainable algorithm deployment and product upgrade system for vehicle manufacturers.

[0148] The deployment phase builds a vehicle model matching guidance mechanism. The system identifies the vehicle platform by vehicle identification code during the first power-on and loads the corresponding model configuration and parameters, while recording local driver behavior preferences for initializing personalized shock absorption adjustment logic. The vehicle accumulates data generated during driving in the subsequent use process, which can be uploaded to the vehicle manufacturer's backend platform for model performance optimization. At the same time, the model iteration cycle of different vehicle models can maintain independence, facilitating model version management according to the product life cycle, forming a maintainable, traceable, and extensible multi-platform model adaptation system.

[0149] Deploying the trained vehicle shock absorption control system containing multiple models to actual vehicles requires full consideration of engineering factors such as computational resource limitations, power stability, real-time constraints, interface compatibility, functional safety levels, and others. Shock absorption adjustment is a dynamic closed-loop control system with extremely low delay tolerance, especially under complex road conditions and high vehicle speed. Any response delay exceeding milliseconds can lead to a decline in ride experience or uncontrolled vibration superposition. Therefore, the model deployment process is not only an algorithm capability migration process, but also a high-precision engineering integration process of the entire inference and control chain, requiring systematic design from software and hardware adaptation, inference structure optimization, control signal decoupling, communication channel stability, and other aspects.

[0150] The running carrier of the model is the vehicle domain controller, which has high-performance processing capability and supports multi-core heterogeneous computing structure. The typical architecture includes a high-frequency CPU, a low-power MCU, and an AI processing unit with neural network acceleration capability. The shock risk identification model, autoencoder, and shock adjustment model are all inferred in the AI acceleration core. The terrain data and vehicle state data are preprocessed by the main processing core and then batched into the inference module. The inference results are analyzed by the main processing core or control-specific MCU to generate control vectors. Each inference cycle is between twenty milliseconds and fifty milliseconds, and the specific frequency is adaptively adjusted according to vehicle speed, road complexity, and algorithm power consumption.

[0151] The multi-model inference structure adopts a pipeline architecture, and all models are not sequentially and serially run, but the current period data is processed in parallel. The inference task scheduling is managed by a local task distribution module, and the core resources are allocated according to the calculation complexity of each model. After the input data preparation is completed, the system sends the terrain data and vehicle state data into the node graph constructor to generate a real-time graph, and then inputs into the shock risk identification model to complete the shock risk scoring. At the same time, the auto-encoder receives the vehicle state data of the same time period to complete the vehicle state coding, and the shock adjustment model receives the shock risk score, the vehicle state vector, the vehicle state data and the current driver identifier to generate a stiffness adjustment value. The three models are independently calculated, and the output results are combined in the control fusion module. The fusion module resides in a high-priority thread to ensure that the output control vector is completed within milliseconds.

[0152] To reduce the model running delay and memory occupation, all models are quantized and compressed before deployment. The shock risk identification model uses structure pruning and weight quantization to compress floating-point parameters to eight-bit integers, reducing storage space while preserving prediction accuracy. The compressed auto-encoder state vector length is limited to between 16 and 32 dimensions to ensure that the input dimension control is within an acceptable range. The shock adjustment model uses a lightweight structure design, which trims the original shock adjustment model to the inference version, retaining only the main level for stiffness adjustment value generation and removing irrelevant auxiliary network structures, significantly improving running efficiency.

[0153] The collection of terrain data and vehicle state data is completed by the vehicle body sensor system, and the sensor data is real-time aggregated to the domain controller through the vehicle Ethernet and controller area network. The data preprocessing module is responsible for completing basic processing such as timestamp alignment, unit normalization and outlier removal. The data is structured as a tensor before entering the model, ensuring that no repeated conversion is needed during the inference phase. The model inference result is decoupled from the original vehicle electronic control architecture, and the conversion between the model output and the electronic control shock absorber driving signal is realized through the control intermediate layer. This intermediate layer adapts to the corresponding driving protocol according to different vehicle platforms, and can output voltage type, PWM type, CAN signal type and other instructions to meet the interface requirements of magnetorheological shock absorbers, electromagnetic valve type shock absorbers, active hydraulic shock absorbers and other devices.

[0154] A redundancy checking mechanism is provided between the control vector from inference to execution to avoid adjustment out of control caused by communication abnormalities. The shock absorber control unit performs legality detection on the control vector before each execution, including numerical range, jump amplitude, change trend and other indicators. Once an abnormal value appears, the safety mode is triggered immediately, switching to the default suspension configuration to ensure that the vehicle still has basic safety performance in the algorithm failure state. All inference data, model output and control vector are stored in the local cache log for subsequent diagnosis and analysis and remote model optimization.

[0155] After deployment, the model supports remote debugging and version management. Through a remote upgrade channel, it enables algorithm upgrades, model replacements, and parameter updates. The system supports a partitioned update mechanism, allowing for hot replacement of model parameters without interrupting the main workflow. Each remote upgrade automatically enters an observation period, where the system monitors the consistency between the model's output trend and the vehicle's response. If a deviation occurs, it can roll back to the previous stable version, ensuring the long-term stable operation of the vehicle's control logic.

[0156] The entire deployment solution undergoes multiple stages in the automotive manufacturer's engineering verification process, including hardware-in-the-loop testing, closed-track dynamic testing, and multi-vehicle integration testing, ensuring the model operates stably, calculates in a timely manner, controls precisely, and responds continuously under various operating conditions. Each module is designed according to automotive-grade standards, meets functional safety requirements, and is guaranteed for stability under extreme temperature, high humidity, high vibration, and high interference conditions. The system ultimately possesses end-to-end online operation capabilities, completing the entire process from terrain recognition, vehicle state modeling, driving style decoding, strategy fusion, control vector generation to hardware execution, all locally on the vehicle. It has autonomous edge computing capabilities, does not rely on cloud inference, and adapts to complex, real-time, and unpredictable road environments. This capability has become an important component of automotive manufacturers' differentiated intelligent chassis control systems, providing the algorithmic execution foundation for vehicles to proactively adapt and evolve in the intelligent era.

[0157] The vehicle damping control system in this application not only achieves algorithmic structural compatibility between different vehicle platforms, but also provides technical support for large-scale deployment by automakers. It significantly reduces the integration cost and debugging complexity of the algorithm in actual implementation, forming a highly engineered algorithmic platform architecture, laying a solid foundation for the subsequent promotion of this intelligent damping adjustment system in more vehicle models, more sub-brands, and more market environments.

[0158] It should be noted that during actual vehicle operation, the vehicle damping control system not only needs to output accurate control vectors in each control cycle, but also must possess dynamic self-feedback capabilities. This means it must continuously monitor the adjustment effect, analyze the difference between the output and the actual vehicle response, and fine-tune the local model based on error changes. By constructing a complete dynamic closed-loop feedback mechanism, the control logic is transformed from a unidirectional command mode to a bidirectional response system, achieving a continuous cycle of prediction, adjustment, data acquisition, and correction. This ensures that the vehicle damping control system always maintains its optimal response state, adapting to changes in vehicle condition, environmental disturbances, and system deviations caused by long-term use.

[0159] The system feedback path is completed by real-time acquisition and judgment logic inside the vehicle controller. In each control cycle, the shock absorber response will produce a set of observable physical quantity changes, including shock absorber compression stroke, vehicle body vertical acceleration rate of change, wheel dynamic load distribution change, pitch rate fluctuation, suspension rebound time, etc. These signals are recorded synchronously through the shock absorber body encoder, vehicle body IMU, wheel speed sensor and other devices, and are correspondingly bound with the control vector of the previous cycle. The system locally stores a short data window, maintains the control vector and vehicle response pairing information within the last few seconds, and calls it for real-time error analysis module.

[0160] The error analysis module calculates the adjustment offset based on the difference between the control target and the actual feedback, mainly including the following types of indicators: first, the delay difference between the target response rate and the actual shock response rate; second, the fitting degree between the vehicle body acceleration change trend guided by the control expected stiffness and the true measured value; third, whether the control output continuity is consistent with the actual vehicle vibration change rhythm; fourth, whether the system output appears flicker or command oscillation in the frequent vibration scene. The system sets dynamic threshold for each type of error, and automatically marks that the current model has adaptive deviation in this working condition when it exceeds the warning range.

[0161] When the error trigger threshold is exceeded, the system will enter the local model fine-tuning state. The fine-tuning process does not involve complete model structure update, but only weight correction within the key parameters or local sub-module range of the model. For example, when the shock risk identification model has a low score for a continuous vibration risk score, resulting in insufficient shock response, the system will locate the graph structure and prediction node corresponding to the score, update the node embedding parameters or edge weight estimation method, and retrain the score mapping relationship of the graph structure. For example, if the vehicle state vector output by the autoencoder is seriously inconsistent with the actual vehicle response trend, the system will start the local reconstruction error tuning mechanism to make subtle adjustments to the autoencoder intermediate layer channel weights, enhancing the model's ability to perceive the current type of vibration.

[0162] The individualized offset strategy of the shock adjustment model also accepts feedback correction. The system will analyze the degree of consistency between the driver's behavior response and the system adjustment result in continuous similar road conditions, and if the stiffness adjustment value is too large or too small for a long time, resulting in unnatural response, the model will reactivate the small sample fast adaptation process, load the previous behavior segment and regenerate the current strategy offset, replacing the old parameters to achieve dynamic evolution of individualized behavior strategy. This process is completed locally and does not depend on external connection, maintaining the system's self-adjustment ability under disconnected conditions.

[0163] All feedback corrections generate records in the local cache and are periodically uploaded to the backend data platform. The platform aggregates feedback from different vehicles, establishes a centralized sample set for common deviation scenarios, and updates the model version regularly. After each update, the model is distributed to vehicles through remote upgrades, and the system automatically completes switching and comparison. After ensuring that the new version has better adjustment effects, the system replaces the old version. If the new version does not adapt to specific user usage, the system supports quick rollback operations to ensure the continuity of control logic.

[0164] The feedback control module also has scene recognition capability. In some special working conditions, such as sharp lane changing, obstacle avoidance braking, and continuous pothole areas, the system can temporarily relax the error limit to prevent atypical states from triggering false adjustments. The recognition mechanism uses multiple features such as vehicle speed change curve, steering wheel angular velocity, brake intensity change, and terrain continuity parameters to make judgments. Once a temporary working condition anomaly is identified, the system will switch to a tolerant mode, stop model fine-tuning during the current period, and maintain stable control parameter output to prevent rhythm disorders caused by false corrections.

[0165] The construction of the closed-loop control mechanism is the core embodiment of the intelligentization of the vehicle damping control system. By systematically introducing feedback, error recognition, parameter fine-tuning, and behavior adaptation processes, the adjustment system evolves from rule-driven to data-driven and then to behavior-driven, allowing each vehicle to develop a unique response style on real roads. At the same time, this mechanism provides a data foundation for vehicle manufacturers to establish a model stability evaluation system, which can be used for model adaptation prediction, algorithm robustness evaluation, risk scenario simulation construction, and other engineering aspects before new vehicle models go online, and can comprehensively enhance the autonomous evolution and safety control capabilities of the vehicle chassis intelligent system.

[0166] It should be noted that the method of the embodiments of the present application can be executed by a single device, such as a computer or a server. The method of the embodiments of the present application can also be applied to a distributed scenario, and completed by multiple devices cooperating with each other. In this distributed scenario, one of the multiple devices can only execute one or more steps in the method of the embodiments of the present application, and the multiple devices can interact with each other to complete the method.

[0167] It should be noted that some embodiments of the present application have been described above. In some cases, the actions or steps described in the above embodiments can be executed in an order different from that in the above embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or can be advantageous.

[0168] Corresponding to the method of any of the above embodiments, the application also provides a vehicle control device based on the same inventive concept.

[0169] Reference Figure 2 The vehicle control device comprises:

[0170] The acquisition module 202 is configured to acquire terrain data and vehicle state data in a vehicle driving process;

[0171] The risk assessment module 204 is configured to generate a shock risk score by a pre-trained shock risk identification model based on the terrain data and the vehicle state data;

[0172] The state extraction module 206 is configured to generate a vehicle state vector by a pre-trained autoencoder based on the vehicle state data;

[0173] The adjustment generation module 208 is configured to generate a stiffness adjustment value by a pre-trained shock adjustment model based on the shock risk score, the vehicle state vector, the vehicle state data and a current driver identifier;

[0174] The fusion module 210 is configured to fuse the shock risk score, the vehicle state vector and the stiffness adjustment value to obtain a control vector;

[0175] The control module 212 is configured to send the control vector to a shock absorber control unit, so that the shock absorber control unit adjusts the stiffness of the shock absorber according to the control vector.

[0176] In some embodiments, the risk assessment module 204 is further configured to construct a vehicle state atlas according to the terrain data and the vehicle state data;

[0177] The vehicle state atlas is input into the shock risk identification model, and the shock risk score is output by the shock risk identification model.

[0178] In some embodiments, the risk assessment module 204 is further configured to generate a terrain node according to the terrain data and a state node according to the vehicle state data;

[0179] A directed edge from the terrain node to the state node is constructed according to a physical dependency relationship;

[0180] The vehicle state atlas is generated based on the terrain node, the state node and the directed edge.

[0181] In some embodiments, a training module is further included, which is configured to construct an initial autoencoder;

[0182] Collect vehicle historical state data of the vehicle in a plurality of driving scenes, a plurality of driving styles and a plurality of vehicle types respectively as first training samples;

[0183] Input the first training samples into the initial autoencoder, and output corresponding state vector samples through the initial autoencoder;

[0184] Decode the state vector samples through the decoder to obtain restoration samples;

[0185] Adjust the parameters of the initial autoencoder by minimizing the difference between the restoration samples and the first training samples, and obtain a trained autoencoder.

[0186] In some embodiments, the training module is further configured to construct an initial shock adjustment model;

[0187] Collect historical terrain data and historical vehicle state data when different drivers drive the vehicle;

[0188] According to the historical terrain data and the historical vehicle state data, a second training sample is constructed;

[0189] The second training sample is divided into a plurality of training tasks, and each training task corresponds to a driver;

[0190] Based on the plurality of training tasks, the initial shock adjustment model is trained using a meta-learning method to obtain a trained shock adjustment model.

[0191] In some embodiments, the fusion module 210 is further configured to calculate a weight coefficient corresponding to each of the shock risk score, the vehicle state vector and the stiffness adjustment value according to historical terrain data and historical vehicle state data within a time period before the current time;

[0192] Based on the weight coefficient, the shock risk score, the vehicle state vector and the stiffness adjustment value are weighted to obtain the control vector.

[0193] In some embodiments, the fusion module 210 is further configured to determine the vehicle type identifier of the current vehicle, and determine the corresponding preset weight proportion relationship according to the vehicle type identifier;

[0194] Based on the preset weight proportion relationship, the shock risk score, the vehicle state vector and the stiffness adjustment value are weighted to obtain the control vector.

[0195] In some embodiments, the state extraction module 206 is further configured to perform quality assessment on the vehicle state data to obtain a quality assessment result;

[0196] According to the quality evaluation result, weights of each channel data in the vehicle state data are determined, the vehicle state data is preprocessed according to the weights of the each channel data, and preprocessed vehicle state data is obtained.

[0197] Based on the preprocessed vehicle state data, a vehicle state vector is generated by using a self-encoder trained in advance.

[0198] For the convenience of description, the above apparatus is described in various modules according to functions. Of course, in the implementation of the present application, the functions of each module can be implemented in one or more software and / or hardware.

[0199] The apparatus of the above embodiment is used to implement the corresponding vehicle control method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which are not described here again.

[0200] Based on the same inventive concept, corresponding to the method of any of the above embodiments, the present application also provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the vehicle control method of any of the above embodiments when executing the program.

[0201] Figure 3 A more specific hardware structure of an electronic device is shown, which can include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are connected to each other through the bus 1050 for communication within the device.

[0202] The processor 1010 can be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, etc., for executing related programs to implement the technical solutions provided by the embodiments of the present application.

[0203] The memory 1020 can be implemented in the form of a ROM (Read Only Memory), a RAM (Random Access Memory), a static storage device, a dynamic storage device, etc. The memory 1020 can store an operating system and other application programs, and when the technical solutions provided in the embodiments of the present specification are implemented by software or firmware, the related program codes are stored in the memory 1020 and are called and executed by the processor 1010.

[0204] The input / output interface 1030 is configured to connect an input / output module to realize information input and output. The input / output module can be configured as a component in the device (not shown in the figure) or can be externally connected to the device to provide corresponding functions. The input device can include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device can include a display, a speaker, a vibrator, an indicator light, etc.

[0205] The communication interface 1040 is configured to connect a communication module (not shown in the figure) to realize the communication interaction between the device and other devices. The communication module can realize communication through a wired manner (such as a USB, a network cable, etc.) or through a wireless manner (such as a mobile network, WIFI, Bluetooth, etc.).

[0206] The bus 1050 includes a channel to transmit information between various components (such as the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040) of the device.

[0207] It should be noted that although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040, and the bus 1050, in the specific implementation process, the device can also include other components necessary for normal operation. In addition, those skilled in the art can understand that the above device can also only include the components necessary to implement the solutions of the embodiments of the present specification, and does not have to include all the components shown in the figure.

[0208] The electronic device of the above embodiments is used to implement the corresponding vehicle control method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which are not described here again.

[0209] Based on the same inventive concept, corresponding to the method of any of the above embodiments, the present application also provides a vehicle, which comprises the electronic device according to any of the above embodiments.

[0210] Based on the same inventive concept, the application also provides a non-transitory computer readable storage medium storing computer instructions for causing a computer to perform the vehicle control method according to any of the above embodiments.

[0211] The computer readable medium of the embodiments can include permanent and non-permanent, removable and non-removable media, which can be realized by any method or technology to store information. The information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device.

[0212] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to perform the vehicle control method according to any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which are not described here.

[0213] Based on the same concept, the application also provides a computer program product comprising computer program instructions for causing a computer to perform the method according to any of the above embodiments when the computer program instructions are run on the computer, having the beneficial effects of the corresponding method embodiments, which are not described here.

[0214] Those skilled in the art should understand that the above discussion of any of the embodiments is only exemplary and is not intended to imply that the scope of the application is limited to these examples; the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes of the different aspects of the embodiments of the application as described above. In order to be brief, they are not provided in detail.

[0215] Additionally, to simplify the description and discussion, and so as not to obscure the embodiments of the application being presented, the well-known functions or constructions of integrated circuit (IC) chips and other components can or can not be shown in the figures and will be omitted as not to unnecessarily obscure the embodiments of the application being presented. Moreover, the devices can be shown in block diagram form in order to avoid obscuring the embodiments of the application, and this also acknowledges the fact that the details in regards to the implementation of such block devices are highly dependent on the platform upon which the embodiments of the application are being implemented (i.e., these details should be well within the purview of one of ordinary skill in the art). Where specific details are set forth in order to describe an illustrative embodiment of the application, it will be apparent to one of ordinary skill in the art that the embodiments of the application can be practiced without, or with variation of, these specific details. Thus, the description is to be considered as illustrative and not restrictive, and the scope of the application should be determined not with reference to the above description, but should be given to the appended claims.

[0216] While the application has been described in connection with specific embodiments thereof, many alternatives, modifications and variations will be apparent to those skilled in the art in light of the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) can use the embodiments discussed.

[0217] Embodiments of the application are intended to cover all such alternatives, modifications and variations as falling within the scope of the application. Accordingly, any omission, modification, equivalent replacement, improvement, etc. made within the spirit and principle of the embodiments of the application should be included in the scope of protection of the application.

Claims

1. A vehicle control method characterized by, The method comprises: acquiring terrain data and vehicle state data during vehicle driving; generating a vibration risk score based on the terrain data and the vehicle state data by a pre-trained vibration risk identification model, comprising: constructing a vehicle state graph based on the terrain data and the vehicle state data, comprising: generating terrain nodes based on the terrain data and state nodes based on the vehicle state data; constructing directed edges from the terrain nodes to the state nodes based on physical dependencies, the directed edges defining causal relationships, temporal sequences, or physical interactions between nodes; each directed edge corresponds to an edge weight, which quantifies the relationship strength between nodes; generating the vehicle state graph based on the terrain nodes, the state nodes, and the directed edges; inputting the vehicle state graph into the vibration risk identification model, and outputting the vibration risk score by the vibration risk identification model; wherein the vibration risk identification model is a graph neural network; based on the vehicle state data, performing dimension compression and feature extraction on the vehicle state data by a pre-trained autoencoder to generate a vehicle state vector; based on the vibration risk score, the vehicle state vector, the vehicle state data, and the current driver identifier, generating a stiffness adjustment value by a pre-trained vibration adjustment model; wherein the vibration adjustment model is trained by historical terrain data and historical vehicle state data of different drivers driving vehicles, and is trained by a meta-learning method; performing weighted fusion on the vibration risk score, the vehicle state vector, and the stiffness adjustment value to obtain a control vector; sending the control vector to a shock absorber control unit to adjust the stiffness of the shock absorber according to the control vector by the shock absorber control unit.

2. The method of claim 1, wherein, The training method of the autoencoder comprises: constructing an initial autoencoder; collecting vehicle historical state data of vehicles in multiple driving scenarios, multiple driving styles, and multiple vehicle types as first training samples; inputting the first training samples into the initial autoencoder to output corresponding state vector samples by the initial autoencoder; decoding the state vector samples by a decoder to obtain restoration samples; adjusting the parameters of the initial autoencoder by minimizing the difference between the restoration samples and the first training samples to obtain a trained autoencoder.

3. The method of claim 1, wherein, The training method of the vibration adjustment model comprises: constructing an initial vibration adjustment model; collecting historical terrain data and historical vehicle state data of different drivers driving vehicles; constructing second training samples based on the historical terrain data and the historical vehicle state data; dividing the second training samples into multiple training tasks, one-to-one corresponding to drivers; based on the multiple training tasks, training the initial vibration adjustment model by a meta-learning method to obtain a trained vibration adjustment model.

4. The method of claim 1, wherein, The fusion of the vibration risk score, the vehicle state vector, and the stiffness adjustment value to obtain a control vector comprises: According to historical terrain data and historical vehicle state data within a time period before the current time, weight coefficients corresponding to the vibration risk score, the vehicle state vector and the stiffness adjustment value are respectively calculated; Based on the weight coefficients, the vibration risk score, the vehicle state vector and the stiffness adjustment value are weighted to obtain the control vector.

5. The method of claim 1, wherein, The fusion of the vibration risk score, the vehicle state vector and the stiffness adjustment value to obtain the control vector comprises: Determine the vehicle model identifier of the current vehicle, and determine the corresponding preset weight proportion relationship according to the vehicle model identifier; Based on the preset weight proportion relationship, the vibration risk score, the vehicle state vector and the stiffness adjustment value are weighted to obtain the control vector.

6. The method of claim 1, wherein, Based on the vehicle state data, the vehicle state vector is generated by a pre-trained autoencoder, comprising: Quality assessment is performed on the vehicle state data to obtain a quality assessment result; According to the quality assessment result, the weight of each channel data in the vehicle state data is determined, and the vehicle state data is preprocessed according to the weight of each channel data to obtain preprocessed vehicle state data; Based on the preprocessed vehicle state data, the vehicle state vector is generated by a pre-trained autoencoder.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, The processor executes the computer program to realize the method of any one of claims 1 to 6.

8. A vehicle characterized by comprising: The vehicle comprises the electronic device of claim 7. The vehicle comprises the electronic device of claim 7.

Images