Vehicle suspension control method and device, computer equipment and readable storage medium

Abstract

The invention relates to a vehicle suspension control method and device, computer equipment and a readable storage medium. The method comprises the following steps: combining obtained vehicle body dynamics time series data, vehicle control time series data and environment perception time series data in a vehicle driving process into time series data of multiple channels, determining a self-adaptive modal number of the time series data under each channel according to the time series data under each channel, and determining the self-adaptive modal number of the time series data under each channel according to the self-adaptive modal number of the time series data under each channel; and performing modal decomposition on the time sequence data based on the adaptive modal number, extracting frequency domain characteristics of the time sequence data, integrating the frequency domain characteristics of the time sequence data of the multiple channels, inputting the frequency domain characteristics into a trained suspension control parameter prediction model to obtain target time sequence characteristics for predicting suspension control parameters, and performing prediction on the suspension control parameters based on the target time sequence characteristics. Vehicle suspension control parameters are obtained through prediction, and a suspension system of the vehicle is controlled based on the vehicle suspension control parameters. By adopting the method, the risk of rollover of the vehicle can be obviously reduced.

Description

Technical Field

[0001] The present application relates to the field of vehicle control technology, and in particular to a vehicle suspension control method, device, vehicle, computer equipment, computer-readable storage medium, and computer program product. Background Art

[0002] With the rapid development of intelligent vehicles, autonomous driving systems, and active safety control technologies, ensuring vehicle stability in complex dynamic environments has become a research hotspot and a technical challenge. Extreme driving conditions, such as high-speed cornering, emergency obstacle avoidance, slippery roads, and asymmetric loads, can significantly increase the risk of rollover and place higher demands on the suspension system's dynamic response capabilities.

[0003] However, traditional suspension systems rely solely on mechanical structure and preset damping for passive response, resulting in problems such as response lag and limited adjustment capabilities, making it difficult to effectively suppress the vehicle's roll or flip tendency in transient dynamics.

[0004] Therefore, it is necessary to provide a solution that can effectively suppress the risk of vehicle rollover in transient dynamics. Summary of the Invention

[0005] Based on this, it is necessary to provide a vehicle suspension control method, device, vehicle, computer equipment, computer-readable storage medium and computer program product that can significantly reduce the risk of vehicle rollover in response to the above technical problems.

[0006] In a first aspect, the present application provides a vehicle suspension control method, comprising:

[0007] Acquire vehicle body dynamics time series data, vehicle control time series data, and environment perception time series data during vehicle driving;

[0008] combining the vehicle body dynamics time series data, the vehicle control time series data, and the environment perception time series data into multi-channel time series data;

[0009] For the time series data under each channel, determine the adaptive mode number of the time series data under the channel, and perform modal decomposition on the time series data based on the adaptive mode number to extract the frequency domain features of the time series data;

[0010] Integrating frequency domain features of time series data from multiple channels, using the integrated multimodal frequency domain features as input, calling a trained suspension control parameter prediction model, extracting time series features from the multimodal frequency domain features, performing weighted fusion on the extracted time series features to obtain target time series features for predicting suspension control parameters, and predicting vehicle suspension control parameters based on the target time series features, wherein the suspension control parameter prediction model is obtained by training a long short-term memory network including a self-attention mechanism based on the frequency domain features of historical multimodal time series data under various road conditions;

[0011] A suspension system of the vehicle is controlled based on the vehicle suspension control parameters.

[0012] In one embodiment, for each channel of time series data, the spectral entropy of the power spectrum of the channel's time series data is determined;

[0013] Based on the spectral entropy, an adaptive mode number of the time series data of the channel is determined.

[0014] In one embodiment, determining the adaptive modal number of the time series data of the channel based on the spectral entropy includes:

[0015] determining a signal change rate of the time series data of the channel;

[0016] Determining an adaptive penalty factor based on the signal change rate and a preset initialization difference factor;

[0017] Based on the adaptive penalty factor and the spectral entropy, an adaptive mode number of the time series data of the channel is determined.

[0018] In one embodiment, performing modal decomposition on the time series data based on the adaptive modal number to extract frequency domain features of the time series data includes:

[0019] Based on the adaptive mode number, decomposing the time series data into a plurality of intrinsic mode function components, wherein the intrinsic mode function components match the adaptive mode number;

[0020] Determining the variance of each of the intrinsic mode function components, and determining the weight of each of the intrinsic mode function components based on the variance;

[0021] Based on the weights of the respective intrinsic mode function components, weighted fusion is performed on the respective intrinsic mode function components to obtain frequency domain features of the time series data of the channel.

[0022] In one embodiment, the suspension control parameter prediction model includes a bidirectional LSTM layer, a fully connected layer, and a convolutional neural network layer, and the extraction of time series features from the frequency domain features of each mode includes:

[0023] Extracting the time series features of the vehicle body dynamics time series data through the bidirectional LSTM layer;

[0024] Mapping the vehicle manipulation time series data to a high-dimensional space through the fully connected layer to obtain time series features of the vehicle manipulation time series data;

[0025] The convolutional neural network is used to propose temporal features of environmental perception temporal data.

[0026] In one embodiment, performing weighted fusion on the extracted time series features to obtain target time series features for predicting suspension control parameters includes:

[0027] Determining the interaction weights between each time series feature and the suspension control task through the multi-head self-attention mechanism in the suspension control parameter prediction model;

[0028] Based on the interaction weights, the time series features are weightedly fused to obtain target time series features for predicting suspension control parameters.

[0029] In one embodiment, before acquiring the vehicle body dynamics time series data, vehicle control time series data, and environment perception time series data during vehicle driving, the method further includes:

[0030] Obtain vehicle body dynamics time series data, vehicle control time series data, and environmental perception time series data under different road conditions to obtain multimodal time series data;

[0031] Performing feature alignment on the multimodal time series data through a gradient reversal layer;

[0032] Based on the multimodal time series data after feature alignment, the constructed initial suspension control parameter prediction model is trained through an adversarial training strategy to obtain a trained suspension control parameter prediction model.

[0033] In one embodiment, after obtaining the trained suspension control parameter prediction model, the method further includes:

[0034] Removing redundant convolution kernels from the trained suspension control parameter prediction model through channel pruning to obtain a compressed suspension control parameter prediction model;

[0035] The compressed suspension control parameter prediction model is deployed on the vehicle.

[0036] In a second aspect, the present application further provides a vehicle suspension control device, comprising:

[0037] A data acquisition module is used to acquire vehicle body dynamics time series data, vehicle control time series data, and environment perception time series data during vehicle driving;

[0038] a data combination module, configured to combine the vehicle body dynamics time series data, the vehicle control time series data, and the environment perception time series data into multi-channel time series data;

[0039] A feature extraction module is used to determine the adaptive modal number of the time series data under each channel, and perform modal decomposition on the time series data based on the adaptive modal number to extract the frequency domain features of the time series data;

[0040] A suspension control parameter prediction module is configured to integrate frequency domain features of time series data from multiple channels, use the integrated multimodal frequency domain features as input, call a trained suspension control parameter prediction model, extract time series features from the multimodal frequency domain features, perform weighted fusion on the extracted time series features, obtain target time series features for predicting suspension control parameters, and predict vehicle suspension control parameters based on the target time series features. The suspension control parameter prediction model is trained using a long short-term memory network with a self-attention mechanism based on the frequency domain features of historical multimodal time series data under various road conditions.

[0041] A suspension control module is configured to control a suspension system of the vehicle based on the vehicle suspension control parameters.

[0042] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above-mentioned vehicle suspension control method embodiments when executing the computer program.

[0043] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which implements the steps in any one of the vehicle suspension control method embodiments when the computer program is executed by a processor.

[0044] In a fifth aspect, the present application also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps in any one of the vehicle suspension control method embodiments.

[0045] In a sixth aspect, the present application also provides a vehicle comprising a vehicle body, a memory and a controller, wherein the memory stores a computer program, and when the controller executes the computer program, it implements the steps in any one of the vehicle suspension control method embodiments.

[0046] The vehicle suspension control method, apparatus, computer device, computer-readable storage medium, computer program product, and vehicle described above can fully understand the current driving scenario by acquiring vehicle body dynamics time series data, vehicle control time series data, and environmental perception time series data during vehicle driving. Subsequently, by determining the adaptive modal number of the multimodal time series data and performing variational modal decomposition on the time series data based on the adaptive modal number, the frequency domain features of the time series data are extracted. The decomposition strength can be automatically adjusted for different road conditions, and more representative key frequency domain features can be extracted, providing more accurate input for subsequent suspension control parameter prediction. Subsequently, the extracted frequency domain feature data is input into a suspension control parameter prediction model. The model, through a long short-term memory network and an attention mechanism, can capture long-term dependencies in the input frequency domain feature data and select target time series features that are more suitable for predicting suspension control parameters. Subsequently, more accurate vehicle suspension control parameters are predicted based on the target time series features. The vehicle suspension is controlled based on the vehicle suspension control parameters, enabling early identification of risks and timely adjustment of the suspension system, significantly reducing the probability of vehicle rollover. The entire solution, through multimodal perception of driving scenarios, fine frequency domain feature extraction, deep learning modeling and active control strategies, achieves comprehensive perception, accurate prediction and rapid response to vehicle status. Especially when facing complex working conditions that may lead to rollover, it can identify risks in advance and actively adjust the suspension system, significantly reducing the probability of vehicle rollover. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments of the present application or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying any creative work.

[0048] Figure 1 1 is a flow chart of a vehicle suspension control method according to an embodiment;

[0049] Figure 2 1 is a flow chart of the adaptive modal number step in one embodiment;

[0050] Figure 3 1 is a flow chart of the adaptive modal number step in another embodiment;

[0051] Figure 4 Schematic diagram of a flow chart of a frequency domain feature extraction step in one embodiment;

[0052] Figure 5 1 is a flow chart of a suspension control parameter prediction model training process in one embodiment;

[0053] Figure 6 1 is a flow chart of a suspension control parameter prediction model training and deployment process in one embodiment;

[0054] Figure 7 is a structural block diagram of a vehicle suspension control device according to one embodiment;

[0055] Figure 8 is a structural block diagram of a vehicle suspension control device in another embodiment;

[0056] Figure 9 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0058] It should be noted that the terms "including" and "having" and any variations thereof used in this application are intended to cover non-exclusive inclusions. The term "plurality" used in this application refers to two or more. The term "and / or" used in this application refers to one or more solutions.

[0059] In one embodiment, Figure 1 As shown, a vehicle suspension control method is provided. This embodiment uses the method applied to a vehicle as an example. It is understood that the method can also be applied to a server, or to a system including a vehicle and a server, and implemented through interaction between the vehicle and the server. In this embodiment, the method includes the following steps (hereinafter referred to as S):

[0060] S100, obtaining vehicle body dynamics time series data, vehicle control time series data, and environment perception time series data during vehicle driving.

[0061] Vehicle body dynamics time series data refers to data on the vehicle's motion state that changes over time, including but not limited to lateral acceleration, displacement and velocity of each wheel suspension, and vehicle pitch and roll angles. Vehicle control time series data refers to vehicle control parameters that change over time, including but not limited to throttle data, brake data, and steering wheel angle data. Vehicle control time series data can be used to reflect the driver's operating intentions. Environmental perception time series data refers to data on the vehicle's external environmental state that changes over time, including but not limited to road adhesion coefficient, weather condition codes, and road height changes.

[0062] In a specific implementation, while the vehicle is in motion, sensors such as the vehicle's suspension sensors, body accelerometers, and inertial measurement units (IMUs) collect data in real time. A vehicle controller (hereinafter referred to as the controller) receives dynamic data from these sensors and performs a series of data processing to generate vehicle dynamic time series data. Accelerator pedal sensors, brake pedal sensors, and steering angle sensors collect vehicle operating parameters in real time. The controller receives vehicle operation data from these sensors to generate vehicle control time series data. Sensors such as wheel speed sensors, radars, temperature sensors, and cameras sense the vehicle's external environment in real time and transmit the collected data to the controller. The controller then organizes the received environmental perception data to generate environmental perception time series data. It is understood that after receiving the raw sensor data from these sensors, the controller may filter, denoise, and normalize the raw sensor data, align data from different sources using timestamps, and store the data as time series data by channel classification.

[0063] S200: combining vehicle body dynamics time series data, vehicle control time series data, and environment perception time series data into multi-channel time series data.

[0064] Multi-channel time series data refers to data collected simultaneously from multiple sensors or measurement points at different time points. This data comes from different sensors, and each channel records different physical quantities or data from different sources. The data between channels is independent.

[0065] For example, the vehicle body dynamics time series data includes suspension displacement and lateral acceleration, the vehicle control time series data includes throttle, brake, and steering wheel angle, and the environment perception time series data includes road adhesion coefficient and weather condition coding. The above vehicle body dynamics time series data, vehicle control time series data, and environment perception time series data are respectively referred to as , and . Among them, suspension displacement refers to the amount of movement in the vertical direction of the vehicle's suspension system during driving. Lateral acceleration refers to the acceleration perpendicular to the direction of vehicle movement, which manifests as a centrifugal effect when the vehicle turns. The road adhesion coefficient refers to the adhesion ability of the tire on different road surfaces. The value of the adhesion coefficient is mainly determined by factors such as the material of the road, the condition of the road surface, the tire structure, tread pattern, material, and the speed of the car. The weather condition code is a digital or letter code converted from various weather phenomena (such as sunny, rainy, snowy, windy, etc.). Specifically, the vehicle can obtain real-time weather update information through V2X communication technology, including weather condition codes.

[0066] By combining , and , get a multi-channel input data, that is, multi-channel time series data There are 7 channels, each of which corresponds to a type of sensor data or a data category. For example, suspension displacement is one channel and lateral acceleration is another channel.

[0067] S300 , for the time series data under each channel, determining the adaptive modal number of the time series data under the channel, performing modal decomposition on the time series data based on the adaptive modal number, and extracting the frequency domain features of the time series data.

[0068] The adaptive modal number characterizes the number of intrinsic mode functions (IMFs) a signal can be decomposed into. In this embodiment, the adaptive modal number refers to the optimal modal number automatically determined through an adaptive method. Typically, different types of signals have varying degrees of complexity, requiring different modal numbers. Frequency domain features describe the energy distribution characteristics of a signal at different frequencies, including but not limited to the dominant frequency, energy percentage, and power spectral density.

[0069] In specific implementation, the energy entropy, mutual information or K-means clustering and other adaptive methods of the time series data under each channel can be determined to estimate the optimal modal number K of the channel. Subsequently, the time series data of the channel is decomposed into K modal signals. Then, FFT (fast Fourier transform) or wavelet transform is performed on each modal signal to extract the frequency domain features. The above processing is performed on the time series under each channel in the above manner to obtain the frequency domain feature vector of each channel.

[0070] S400 integrates the frequency domain features of the time series data of multiple channels, uses the integrated multimodal frequency domain features as input, calls the trained suspension control parameter prediction model, extracts the time series features in the multimodal frequency domain features, constructs the target time series features for predicting the suspension control parameters from the extracted time series features, and predicts the vehicle suspension control parameters based on the target time series features.

[0071] Among them, the suspension control parameter prediction model is obtained by training a long short-term memory network with a self-attention mechanism based on the frequency domain characteristics of historical multimodal time series data under various road conditions.

[0072] Multimodal frequency-domain features integrate frequency-domain features from multiple channels (such as vehicle body dynamics, control input, and environmental perception) and process them into frequency-domain features that meet the input requirements of the suspension control parameter prediction model. Target time-series features are key features selected by the model for predicting optimal suspension control parameters. These can include energy, dominant frequency, and time-series rate of change in specific frequency bands.

[0073] Suspension systems include passive suspension systems, semi-active suspension systems, and active suspension systems. In this embodiment, the predicted suspension control parameters, also referred to as active suspension control parameters, refer to physical quantities used to control the active suspension system, including but not limited to shock absorber damping coefficient, active air spring stiffness, and active suspension actuator output force. It will be appreciated that in other embodiments, suspension control parameters of a semi-active suspension system can also be predicted to further control the semi-active suspension system.

[0074] In practical applications, the vehicle suspension control task can be modeled as a prediction problem for multimodal time series data. A prediction model can be designed by combining bidirectional LSTM (Long Short-Term Memory) and Transformer structures, integrating an attention mechanism, and training a cross-modal deep prediction model that can be used to predict suspension control parameters through multimodal time series data. Specifically, model training can be as follows: First, based on the temporal dependency and multimodal fusion requirements in the vehicle suspension control parameter prediction task, a backbone network of the suspension control parameter prediction model is constructed based on the LSTM-Transformer hybrid structure. The backbone network includes a temporal feature extraction module and a cross-modal attention fusion module. The temporal feature extraction module uses a bidirectional LSTM layer to capture the long-term dependency of vehicle dynamics parameters, and extracts the weights of the temporal features of different modalities for each cross-modal attention fusion module through a multi-head self-attention mechanism.

[0075] Subsequently, the vehicle body dynamics time series data, vehicle control time series data, environmental perception time series data and optimal suspension control parameters of the vehicle driving under road conditions of varying complexity during the historical time period are collected. Afterwards, the optimal suspension control parameter labels are added to the collected vehicle body dynamics time series data, vehicle control time series data and environmental perception time series data to obtain a training data set. Next, based on the training data set, a curriculum learning strategy is adopted, and the training process gradually transitions from simple urban road scenes to complex road conditions (such as rainy and snowy days, and icy roads) to ensure that the model masters different driving modes in stages. The optimizer of the model selects an adaptive algorithm that takes into account both convergence speed and stability. The loss function can adopt a composite loss function, combined with the mean square error term and the control parameter smoothness constraint to avoid sudden changes in the output instructions. The loss function combines the mean square error (MAE) with the control smoothness constraint to avoid sudden changes in the values of the suspension control parameters:

[0076]

[0077] in, represents the actual optimal suspension control parameters, Indicates the rate of change of the actual optimal suspension control parameters; are the suspension control parameters predicted by the model, To predict the rate of change of suspension control parameters, is the smoothing coefficient, Represents the total loss value.

[0078] Based on the above loss function, the difference between the actual optimal suspension control parameters and the predicted suspension control parameters is fitted to obtain a trained suspension control parameter prediction model. The trained suspension control parameter prediction model is then deployed in the vehicle control system.

[0079] In specific implementation, the controller integrates the frequency domain features from multiple channels to obtain multimodal frequency domain features ,in, Represents vehicle body dynamics time series data (suspension displacement, lateral acceleration), Represents vehicle control timing data (throttle, brake, steering wheel angle), Represents environmental perception time series data (road adhesion coefficient, weather condition encoding). Next, the multimodal frequency domain features are input into the trained suspension control parameter prediction model. The model extracts the time series features from the frequency domain features under each mode through the time series feature extraction module to obtain multiple time series features. Subsequently, the target time series features for predicting suspension control parameters are selected from the extracted multiple time series features. Finally, based on the target time series features, a multi-task learning architecture is used to predict the suspension control parameters. Specific control quantity The output layer predicts the best suspension control parameters through the fully connected layer and Tanh activation function:

[0080]

[0081] in, , Represents the weights and biases of the suspension control parameter prediction layer.

[0082] S500: Controlling a suspension system of the vehicle based on a vehicle suspension control parameter.

[0083] In this embodiment, taking the controlled suspension system as an active suspension system as an example, after predicting the optimal suspension control parameters, the controller can convert the suspension control parameters into control signals of specific actuators (such as PWM signals to control solenoid valves, voltage signals to drive motors) through control algorithms such as PID algorithms and fuzzy control methods, and then send the control signals to the active suspension system to adjust the action of the active suspension system. During this process, the vehicle status, such as actual damping value and body posture, can be monitored and fed back in real time for closed-loop adjustment to minimize the possibility of vehicle rollover.

[0084] There is no limit to the way to determine the number of adaptive modes. Figure 2As shown, in some embodiments, for the time series data in each channel, determining the adaptive modal number of the time series data in the channel includes:

[0085] S320 , for the time series data of each channel, determine the spectral entropy of the power spectrum of the time series data of the channel.

[0086] S340 : Determine the adaptive mode number of the time series data of the channel based on the spectral entropy.

[0087] The power spectrum describes the distribution of a signal's energy across different frequencies. Spectral entropy, a metric based on information theory, measures the "chaos" or uncertainty of a signal's frequency domain distribution. For example, if a signal's energy is concentrated in a few frequencies, the spectral entropy is low, representing fewer frequency domain components, and a smaller number of modes can be used to represent the signal. If the energy is distributed across multiple frequencies, the spectral entropy is high, representing a higher frequency domain component, and therefore, a larger number of modes is required to accurately represent the signal.

[0088] In specific implementation, the time series data (i.e. signal data) under each channel can be subjected to FFT or wavelet transform to obtain a complex spectrum. Then, based on the complex spectrum, the power spectrum of the time series data can be calculated. Next, the spectral entropy is calculated based on the power spectrum in the following way :

[0089]

[0090] in, Indicates the The power spectrum of the modal data, and .

[0091] Furthermore, after obtaining the spectral entropy, the adaptive modal number can be determined through a segmented mapping approach. For example, a correspondence between different spectral entropy ranges and different adaptive modal numbers K is pre-set, and then the adaptive modal number K is determined by finding the spectral entropy interval within which the calculated spectral entropy falls.

[0092] In other embodiments, a regression model for predicting the optimal number of modes can be pre-trained, and the optimal number of modes can be obtained by inputting the spectral entropy and time series data into the regression model. The regression model can be a random forest or a neural network, etc., and is not limited to this.

[0093] In this embodiment, by calculating the spectral entropy of each channel signal, the optimal number of modes required for modal decomposition can be accurately determined, thereby more effectively extracting frequency domain features and providing high-quality input features for subsequent suspension control parameter prediction.

[0094] like Figure 3As shown, in some exemplary embodiments, S340 includes:

[0095] S342, determining the signal change rate of the channel's time series data.

[0096] S344: Determine an adaptive penalty factor based on the signal change rate and a preset initialization difference factor.

[0097] S346 , determining the adaptive mode number of the time series data of the channel based on the adaptive penalty factor and the spectral entropy.

[0098] The signal change rate is used to describe the intensity of signal fluctuations in the time domain. The initialization difference factor is a preset constant used to control the intensity of the influence of the signal change rate on the penalty factor. The adaptive penalty factor is a coefficient used to control the decomposition accuracy and the number of modes. In this embodiment, it can be based on Adaptive penalty factor for signal change rate .

[0099] In specific implementation, the time domain volatility of time series data can be evaluated by the difference method to determine whether it is a non-stationary signal, and then the adaptive penalty factor can be determined by the signal change rate and the preset initialization difference factor. :

[0100]

[0101] in, Represents calculation The first-order difference of is the initialization differential factor. That is, the greater the signal fluctuation, the smaller the adaptive penalty factor The larger the value, the more it helps to limit the bandwidth of each mode and prevent aliasing.

[0102] Then, based on the penalty factor and spectral entropy , determine the adaptive mode number K:

[0103]

[0104] in, and are penalty factors, and is a constant.

[0105] In this embodiment, fixed parameters are not relied upon. Instead, the adaptive modal number is determined by the signal change rate and spectral entropy. The adaptive factor is dynamically adjusted according to the signal itself to determine a more reasonable modal number. This can reduce problems such as modal aliasing, over-decomposition or under-decomposition, and has better versatility for data from different road conditions, driving behaviors, and sensor types.

[0106] like Figure 4As shown, in some exemplary embodiments, modal decomposition is performed on time series data based on the adaptive modal number to extract frequency domain features of the time series data, including:

[0107] S420 , based on the adaptive mode number, decompose the time series data into a plurality of intrinsic mode function components, where the intrinsic mode function components match the adaptive mode number.

[0108] S440 , determining the variance of each eigenmode function component, and determining the weight of each eigenmode function component based on the variance.

[0109] S460 , performing weighted fusion on each eigenmode function component based on the weight of each eigenmode function component to obtain frequency domain features of the time series data of the channel.

[0110] The intrinsic mode function component, or IMF component, refers to each mode obtained by decomposing time series data using the variational mode decomposition (VMD) algorithm.

[0111] In this embodiment, an improved VMD algorithm based on spectral entropy to determine the number of adaptive modes is used to perform variational mode decomposition on the time series data under the channel and extract frequency domain features. In specific implementation, after determining the adaptive mode K of the time series data under each channel, the variational mode decomposition algorithm can be used to perform variational mode decomposition on the time series data under each channel, and the frequency domain features of each channel data can be extracted. After decomposition, the IMF component group is obtained:

[0112]

[0113] in, Channel data Decomposed IMF components.

[0114] Next, the variance of each IMF component can be determined to measure the volatility or energy of the IMF component. The larger the variance, the more "active" or "important" the IMF component is. Next, the weight of each IMF is determined based on the variance. :

[0115]

[0116] in, Indicates the IMF components, Indicates the The weighted values of the IMF components, and , represents the variance of the IMF component.

[0117] Finally, through the weights of each IMF , perform weighted fusion on all IMF components under the channel to obtain the frequency domain characteristics of the time series data under the channel, specifically:

[0118] The final fused frequency domain features are ,and Through the above method, variational mode decomposition is performed on the time series data under each channel to obtain the frequency domain characteristics of the time series data of each channel.

[0119] It is understandable that, in other embodiments, EMD (Empirical Mode Decomposition) or other modal decomposition methods may be used to perform modal decomposition on the time series data of each channel and extract frequency domain features.

[0120] In this embodiment, by performing variational modal decomposition on the time series data under the channel, the modal aliasing problem can be reduced and different frequency components can be separated more accurately.

[0121] In some exemplary embodiments, the suspension control parameter prediction model includes a bidirectional LSTM layer, a fully connected layer, and a convolutional neural network layer, which extracts time series features from the frequency domain features of each mode, including:

[0122] The time series features of the vehicle body dynamics time series data are extracted through the bidirectional LSTM layer.

[0123] The vehicle control time series data is mapped to a high-dimensional space through a fully connected layer to obtain the time series features of the vehicle control time series data.

[0124] The temporal features of environmental perception time series data are proposed through convolutional neural networks.

[0125] In this embodiment, in order to solve the heterogeneity problem of different sensor data, the modal embedding layer can be used to perform feature mapping on the input multi-modal time series data to extract the time series features. Specifically, the time series feature module of the suspension control parameter prediction model includes a bidirectional LSTM layer (forward LSTM and reverse LSTM), a fully connected layer, and a convolutional neural network layer. The controller can be used to transform the vehicle body dynamics time series data into Directly input the bidirectional LSTM layer, the processing process of the bidirectional LSTM layer is as follows:

[0126]

[0127] in, and Respectively expressed in The hidden states of the forward LSTM and the reverse LSTM at the time point, similarly, Indicates The hidden state of the forward LSTM at the time point, For The hidden state of the reverse LSTM at the time point, It is the multimodal frequency domain feature, that is, the splicing and integration result of the frequency domain features after multi-channel modal decomposition.

[0128] Then, the vehicle body dynamics time series characteristics are obtained through the output of forward LSTM and reverse LSTM :

[0129]

[0130] At the same time, the vehicle control data is transmitted through the fully connected layer (FC) Map to high-dimensional space, map the features of each time step, and obtain the time series features of vehicle control time series data. In addition, the environment perception time series data The data is input into a convolutional neural network, which uses a one-dimensional convolutional layer (Conv1D) and a maximum pooling layer (MaxPool1D) to extract the embedded representation of the environmental perception time series data. The processing process is as follows:

[0131]

[0132] in, Embedded representations representing context-aware features.

[0133] Then, the extracted temporal features can be uniformly sent to the cross-modal attention fusion module in the model, and the interaction weights of the temporal features of different modalities can be dynamically allocated through the multi-head self-attention mechanism. Finally, based on the interaction weights, the temporal features are weightedly fused to obtain the target temporal features for predicting suspension control parameters.

[0134] The interaction weights are used to represent the importance of each time-step temporal feature in the current task (e.g., suspension control) and are automatically learned by the attention mechanism. The target temporal features are the most representative temporal features extracted after weighted fusion and can be directly used for the final suspension control parameter prediction.

[0135] Specifically, the cross-modal attention fusion module uses the Multi-Head Self-Attention (MHSA) mechanism to calculate the interaction weights between temporal features and predictions of different modalities:

[0136]

[0137] in, represent the query matrix, key matrix and value matrix respectively, is the transpose of the key matrix, is the feature dimension.

[0138] Then, based on the interaction weight, the gating mechanism is used to retain important time series features, and the retained time series features are fused. Finally, the fused target time series features are Output through residual connection and layer normalization (LayerNorm):

[0139]

[0140] It includes time series features extracted from vehicle body dynamics time series data, vehicle control time series data, and environment perception time series data, among which, The core purpose of using 8 parallel attention heads is to enhance the model's ability to dynamically interact with complex multimodal time series data by modeling multiple independent feature subspaces. Each attention head can learn different attention patterns, which are ultimately combined to form a more comprehensive target time series feature.

[0141] Furthermore, a multi-task learning architecture is used to predict the suspension control parameters. The output layer obtains the optimal suspension control parameters through a fully connected layer and a Tanh activation function:

[0142]

[0143] in, , Represents the weights and biases of the active suspension control parameter prediction layer.

[0144] In this embodiment, by designing a modal embedding layer to perform feature mapping on multimodal time series data, different sensor data can be mapped to a vector space of the same dimension after passing through their respective embedding layers, which facilitates subsequent fusion and processing. In addition, heterogeneous data is more robust after embedding and has stronger adaptability to different road conditions and driving behaviors. By performing weighted fusion in a unified space, the fused target time series features have stronger expressive and generalization capabilities.

[0145] In order to make the suspension control parameter prediction model more generalizable, a multi-modal time series training model under multiple road conditions can be used. Figure 5 As shown, in some exemplary embodiments, before obtaining the vehicle body dynamics time series data, vehicle control time series data, and environment perception time series data during vehicle driving, the method further includes:

[0146] S102, acquiring vehicle body dynamics time series data, vehicle control time series data, and environment perception time series data under different road conditions to obtain multimodal time series data.

[0147] S104: Perform feature alignment on the multimodal time series data through a gradient reversal layer.

[0148] S106 , based on the multimodal time series data after feature alignment, the constructed initial suspension control parameter prediction model is trained through an adversarial training strategy to obtain a trained suspension control parameter prediction model.

[0149] In this embodiment, road conditions of varying complexity include, but are not limited to, urban roads, highways, mountain roads, rainy roads, and icy roads. Multimodal time series data can be generated by pre-collecting vehicle body dynamics time series data (suspension displacement, lateral acceleration), vehicle control time series data (throttle, brake, steering wheel angle), and environmental perception time series data (road adhesion coefficient, weather condition encoding) under different road conditions. Subsequently, this data is aligned in time series to form a multimodal time series dataset.

[0150] Next, an initial neural network model is constructed, including a gradient reversal layer. This model consists of a feature extractor, a road condition classifier, and a suspension control parameter predictor. During training, multimodal time series data is input into the feature extractor to obtain a feature vector, which is then passed to the road condition classifier and suspension control parameter predictor, respectively, through the gradient reversal layer. The gradient reversal layer transfers features normally during forward propagation, but during backward propagation, the gradient calculated by the road condition classifier is multiplied by a negative coefficient (inverting the gradient direction) and then passed back to the feature extractor. This causes the feature extractor to update its parameters in a direction that confuses the road condition classifier, thereby learning a feature representation that is robust to road condition variations. Specifically, the feature distributions under different road conditions are aligned.

[0151] Finally, adversarial training is performed on the feature-aligned multimodal time series data. In each iteration, the model simultaneously computes two losses: the suspension control parameter prediction loss (e.g., mean squared error) and the road condition classification loss (e.g., cross entropy). Through a gradient reversal mechanism, the feature extractor is optimized to simultaneously minimize the suspension control prediction loss and maximize the road condition classification loss (i.e., making it impossible for the road condition classifier to distinguish between different road conditions), while the road condition classifier is optimized to minimize the road condition classification loss. Through this adversarial training strategy, the model ultimately learns the ability to generalize to different road conditions, resulting in a trained suspension control parameter prediction model.

[0152] In this embodiment, the gradient reversal layer is used to suppress features related to road conditions, and adversarial training is then used to force the feature extractor to ignore features related to road conditions and focus on essential features directly related to suspension control, thereby improving the model's generalization ability under unknown road conditions.

[0153] In order to meet the real-time requirements of the vehicle electronic control unit, the model can be compressed before deployment. Figure 6 As shown, in some exemplary embodiments, after S106, the method further includes:

[0154] S108, removing redundant convolution kernels in the trained suspension control parameter prediction model through channel pruning to obtain a compressed suspension control parameter prediction model, and deploying the compressed suspension control parameter prediction model on the vehicle.

[0155] Channel pruning is a model compression technique that reduces model parameters and computational complexity by removing unimportant channels (i.e., certain output channels of the convolution kernel) from a convolutional neural network. In a suspension control parameter prediction model, if the output channels of certain convolution kernels contribute little to the final prediction results, they can be removed without significantly affecting model performance.

[0156] In specific implementation, the importance of the output channels of each convolutional layer in the suspension control parameter prediction model can be evaluated. Then, based on the target compression rate and model performance requirements, the proportion of channels to be pruned in each layer is determined (for example, removing 10% of the channels). Next, for each convolutional layer, unimportant output channels are removed, and the input channels corresponding to these channels in the next layer are also removed. After pruning, the model parameters W′ satisfy:

[0157]

[0158] in, is a binary mask, is the pruning threshold, are the complete model parameters before pruning; Represents the activation value of the jth neuron in the i-th layer network before pruning. When it is greater than the pruning threshold When , the corresponding mask is 1.

[0159] Since model performance may degrade, fine-tuning can be performed on the training data to restore performance. Furthermore, a quantized deployment approach can be used to convert the trained suspension control parameter prediction model into a format suitable for onboard hardware deployment and integrate it into the vehicle control system.

[0160] In other embodiments, in order to ensure that the model meets the real-time requirements of vehicle control, a closed-loop verification system can be built on a high-performance test platform. By injecting sensor data streams of typical working conditions such as urban roads, highways, rainy and snowy days, the end-to-end inference delay from multimodal time series data input to vehicle suspension control parameter output is tested, and the frame rate stability of the model under extreme road disturbances is evaluated. The verification process simultaneously monitors the resource usage of the on-board electronic control unit to ensure that the lightweight model still maintains millisecond-level active suspension control response capabilities under pruning and INT8 quantization conditions, providing fast and reliable guarantees for actual vehicle deployment.

[0161] In order to more clearly illustrate the vehicle suspension control method provided by the present application, a specific embodiment is described below. The embodiment includes the following contents:

[0162] S1, obtaining vehicle body dynamics time series data, vehicle control time series data and environment perception time series data during vehicle driving.

[0163] Specifically, the vehicle body dynamics time series data includes suspension displacement and lateral acceleration, the vehicle control time series data includes throttle, brake, and steering wheel angle, and the environmental perception time series data includes road adhesion coefficient and weather condition coding. For example, the above vehicle body dynamics time series data, vehicle control time series data, and environmental perception time series data are respectively referred to as , and .

[0164] S2, combining the vehicle body dynamics time series data, vehicle control time series data, and environment perception time series data into multi-channel time series data.

[0165] By combining , and , get a multi-channel input data, that is, multi-channel time series data .

[0166] S3, for the time series data under each channel, determine the spectral entropy of the power spectrum of the time series data of the channel, determine the adaptive modal number of the time series data under the channel based on the spectral entropy, and perform modal decomposition on the time series data based on the adaptive modal number to extract the frequency domain characteristics of the time series data.

[0167] Specifically, the signal change rate of the channel's time series data is first determined. Based on the signal change rate and a preset initialization differential factor, an adaptive penalty factor is determined. Based on the adaptive penalty factor and spectral entropy, the adaptive modal number K of the channel's time series data is determined. Based on the adaptive modal number K, the time series data under each channel is subjected to VMD decomposition to obtain K IMF components. Based on the variance of each IMF component, the weight of the IMF component is determined, and multiple IMF components are weightedly fused to obtain the frequency domain characteristics of the channel's time series data.

[0168] S4 integrates the frequency domain features of the time series data of multiple channels, takes the integrated multimodal frequency domain features as input, calls the trained suspension control parameter prediction model, extracts the time series features in the multimodal frequency domain features, performs weighted fusion on the extracted time series features, obtains the target time series features for predicting the suspension control parameters, and predicts the vehicle suspension control parameters based on the target time series features. The suspension control parameter prediction model is obtained by training a long short-term memory network containing a self-attention mechanism based on the frequency domain features of historical multimodal time series data under various road conditions.

[0169] Then, the controller integrates the frequency domain features from multiple channels to obtain multimodal frequency domain features. ,in, Represents vehicle body dynamics time series data (suspension displacement, lateral acceleration), Represents vehicle control timing data (throttle, brake, steering wheel angle), represents the environmental perception time series data (road adhesion coefficient, weather condition encoding). Next, the multimodal frequency domain features Input to the trained suspension control parameter prediction model, specifically, the controller converts the vehicle body dynamics time series data Directly input the bidirectional LSTM layer to obtain the vehicle body dynamics time series features, and then use the fully connected layer (FC) to convert the vehicle control data into Map to high-dimensional space, map the features of each time step, and obtain the time series features of vehicle control time series data. In addition, the environment perception time series data The input is fed into a convolutional neural network, specifically using a one-dimensional convolutional layer (Conv1D) and a maximum pooling layer (MaxPool1D) to extract an embedded representation of the environmental perception time series data. The extracted time series features are then fed into the model's cross-modal attention fusion module. A multi-head self-attention mechanism dynamically assigns interaction weights to the time series features of different modalities. Finally, based on the interaction weights, the time series features are weighted and fused to obtain the target time series features used to predict suspension control parameters. The suspension control parameters of the active suspension system are then predicted based on the target time series features.

[0170] S5, controlling the suspension system of the vehicle based on the vehicle suspension control parameters.

[0171] After predicting the optimal suspension control parameters, the controller can convert the suspension control parameters into control signals for specific actuators (such as PWM signals to control solenoid valves, voltage signals to drive motors) through control algorithms such as PID algorithms and fuzzy control methods. The control signals are then sent to the active suspension system to adjust the action of the active suspension system. During this process, the vehicle status, such as the actual damping value and body posture, can be monitored and fed back in real time for closed-loop adjustment to minimize the possibility of vehicle rollover.

[0172] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily performed in sequence in the order indicated by the arrows. Unless clearly stated herein, the execution of these steps is not strictly limited in order, and these steps can be performed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above may include multiple steps or multiple stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of the steps or stages in other steps or other steps. It is understandable that the various steps in different embodiments can be freely combined as needed, and the various non-contradictory schemes formed by the combination all fall within the scope of protection of this application.

[0173] Based on the same inventive concept, embodiments of the present application also provide a vehicle suspension control device for implementing the aforementioned vehicle suspension control method. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more vehicle suspension control device embodiments provided below can be found in the above-described limitations of the vehicle suspension control method and will not be further elaborated here.

[0174] In an exemplary embodiment, Figure 7 As shown, a vehicle suspension control device 700 is provided, comprising: a data acquisition module 710, a data combination module 720, a feature extraction module 730, a suspension control parameter prediction module 740 and a suspension control module 750, wherein:

[0175] The data acquisition module 710 is used to acquire vehicle body dynamics time series data, vehicle control time series data and environment perception time series data during vehicle driving.

[0176] The data combination module 720 is used to combine the vehicle body dynamics time series data, the vehicle control time series data and the environment perception time series data into multi-channel time series data.

[0177] The feature extraction module 730 is used to determine the adaptive modal number of the time series data under each channel, perform modal decomposition on the time series data based on the adaptive modal number, and extract frequency domain features of the time series data.

[0178] The suspension control parameter prediction module 740 is used to integrate the frequency domain features of the time series data of multiple channels, take the integrated multimodal frequency domain features as input, call the trained suspension control parameter prediction model, extract the time series features in the multimodal frequency domain features, perform weighted fusion on the extracted time series features, and obtain the target time series features for predicting the suspension control parameters. Based on the target time series features, the vehicle suspension control parameters are predicted, wherein the suspension control parameter prediction model is obtained by training a long short-term memory network containing a self-attention mechanism based on the frequency domain features of historical multimodal time series data under various road conditions.

[0179] The suspension control module 750 is configured to control the suspension system of the vehicle based on the vehicle suspension control parameters.

[0180] In some embodiments, the feature extraction module 730 is further used to determine the spectral entropy of the power spectrum of the time series data of each channel, and determine the adaptive mode number of the time series data of the channel based on the spectral entropy.

[0181] In some embodiments, the feature extraction module 730 is also used to determine the signal change rate of the channel's time series data, determine an adaptive penalty factor based on the signal change rate and a preset initialization differential factor, and determine the adaptive mode number of the channel's time series data based on the adaptive penalty factor and spectral entropy.

[0182] In some embodiments, the feature extraction module 730 is also used to decompose the time series data into multiple intrinsic mode function components based on the adaptive mode number, match the intrinsic mode function components with the adaptive mode number, determine the variance of each intrinsic mode function component, determine the weight of each intrinsic mode function component based on the variance, and perform weighted fusion on each intrinsic mode function component based on the weight of each intrinsic mode function component to obtain the frequency domain characteristics of the time series data of the channel.

[0183] In some embodiments, the suspension control parameter prediction model includes a bidirectional LSTM layer, a fully connected layer, and a convolutional neural network layer to extract the timing features in the frequency domain features of each mode. The suspension control parameter prediction module 740 is also used to extract the timing features of the vehicle body dynamics timing data through the bidirectional LSTM layer, map the vehicle control timing data to a high-dimensional space through the fully connected layer to obtain the timing features of the vehicle control timing data, and propose the timing features of the environmental perception timing data through the convolutional neural network.

[0184] In some embodiments, the suspension control parameter prediction module 750 is also used to determine the interaction weights between each timing feature and the suspension control task through the multi-head self-attention mechanism in the suspension control parameter prediction model, and perform weighted fusion on the timing features based on the interaction weights to obtain the target timing features for predicting the suspension control parameters.

[0185] like Figure 8 As shown, in some embodiments, the device also includes a model training module 702, which is used to obtain the vehicle body dynamics time series data, vehicle control time series data and environmental perception time series data under different road conditions to obtain multimodal time series data, and perform feature alignment on the multimodal time series data through a gradient reversal layer. Based on the multimodal time series data after feature alignment, the constructed initial suspension control parameter prediction model is trained through an adversarial training strategy to obtain a trained suspension control parameter prediction model.

[0186] In some embodiments, the device also includes a model deployment module 704, which is used to remove redundant convolution kernels in the trained suspension control parameter prediction model through channel pruning to obtain a compressed suspension control parameter prediction model, and deploy the compressed suspension control parameter prediction model on the vehicle.

[0187] Each module in the aforementioned vehicle suspension control device may be implemented in whole or in part through software, hardware, or a combination thereof. Each module may be embedded in or independent of a processor within a computer device in hardware form, or may be stored in a computer device memory in software form, allowing the processor to call and execute the corresponding operations of each module.

[0188] In an exemplary embodiment, the present application also provides a vehicle including a vehicle body, a memory and a controller, wherein the memory stores a computer program, and when the controller executes the computer program, the steps of any one of the above-mentioned vehicle suspension control method embodiments are implemented.

[0189] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Figure 9As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store data such as vehicle body dynamics timing data, vehicle control timing data, environmental perception timing data, and suspension control parameters. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a vehicle suspension control method is implemented.

[0190] Those skilled in the art will understand that Figure 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0191] In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps of any one of the above-mentioned vehicle suspension control method embodiments when executing the computer program.

[0192] In an exemplary embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of any one of the above-mentioned vehicle suspension control method embodiments are implemented.

[0193] In an exemplary embodiment, a computer program product is provided, including a computer program, which implements the steps of any one of the above vehicle suspension control method embodiments when executed by a processor.

[0194] It should be noted that the user information (including but not limited to user device information, vehicle information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0195] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), quantum computing-based data processing logic devices, artificial intelligence (AI) processors, and the like.

[0196] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0197] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A vehicle suspension control method, characterized in that: The method comprises: Acquire vehicle body dynamics time series data, vehicle control time series data, and environment perception time series data during vehicle driving; combining the vehicle body dynamics time series data, the vehicle control time series data, and the environment perception time series data into multi-channel time series data; For the time series data under each channel, determine the adaptive modal number of the time series data under the channel, and perform modal decomposition on the time series data based on the adaptive modal number to extract the frequency domain features of the time series data; Integrating frequency domain features of time series data from multiple channels, using the integrated multimodal frequency domain features as input, calling a trained suspension control parameter prediction model, extracting time series features from the multimodal frequency domain features, performing weighted fusion on the extracted time series features to obtain target time series features for predicting suspension control parameters, and predicting vehicle suspension control parameters based on the target time series features, wherein the suspension control parameter prediction model is obtained by training a long short-term memory network including a self-attention mechanism based on the frequency domain features of historical multimodal time series data under various road conditions; A suspension system of the vehicle is controlled based on the vehicle suspension control parameters.

2. The method according to claim 1, characterized in that The step of determining the adaptive modal number of the time series data in each channel includes: For each channel of time series data, determine the spectral entropy of the power spectrum of the time series data of the channel; Based on the spectral entropy, an adaptive mode number of the time series data of the channel is determined.

3. The method according to claim 2, characterized in that The determining, based on the spectral entropy, the adaptive modal number of the time series data of the channel includes: determining a signal change rate of the time series data of the channel; Determining an adaptive penalty factor based on the signal change rate and a preset initialization difference factor; Based on the adaptive penalty factor and the spectral entropy, an adaptive mode number of the time series data of the channel is determined.

4. The method according to claim 1, wherein The performing modal decomposition on the time series data based on the adaptive modal number to extract frequency domain features of the time series data includes: Based on the adaptive mode number, decomposing the time series data into a plurality of intrinsic mode function components, wherein the intrinsic mode function components match the adaptive mode number; Determining the variance of each of the intrinsic mode function components, and determining the weight of each of the intrinsic mode function components based on the variance; Based on the weights of the respective intrinsic mode function components, weighted fusion is performed on the respective intrinsic mode function components to obtain frequency domain features of the time series data of the channel.

5. The method according to any one of claims 1 to 4, characterized in that The suspension control parameter prediction model includes a bidirectional LSTM layer, a fully connected layer, and a convolutional neural network layer. The extraction of time series features from the frequency domain features of each mode includes: Extracting the time series features of the vehicle body dynamics time series data through the bidirectional LSTM layer; Mapping the vehicle manipulation time series data to a high-dimensional space through the fully connected layer to obtain time series features of the vehicle manipulation time series data; The convolutional neural network is used to propose temporal features of environmental perception temporal data.

6. The method according to any one of claims 1 to 4, characterized in that The weighted fusion of the extracted time series features to obtain target time series features for predicting suspension control parameters includes: Determining the interaction weights between each time series feature and the suspension control task through the multi-head self-attention mechanism in the suspension control parameter prediction model; Based on the interaction weights, the time series features are weightedly fused to obtain target time series features for predicting suspension control parameters.

7. The method according to any one of claims 1 to 4, characterized in that Before acquiring the vehicle body dynamics time series data, vehicle control time series data, and environment perception time series data during vehicle driving, the method further includes: Obtain vehicle body dynamics time series data, vehicle control time series data, and environmental perception time series data under different road conditions to obtain multimodal time series data; Performing feature alignment on the multimodal time series data through a gradient reversal layer; Based on the multimodal time series data after feature alignment, the constructed initial suspension control parameter prediction model is trained through an adversarial training strategy to obtain a trained suspension control parameter prediction model.

8. The method according to claim 7, characterized in that After obtaining the trained suspension control parameter prediction model, the method further includes: Removing redundant convolution kernels from the trained suspension control parameter prediction model through channel pruning to obtain a compressed suspension control parameter prediction model; The compressed suspension control parameter prediction model is deployed on the vehicle.

9. A vehicle suspension control device, characterized in that: The device comprises: A data acquisition module is used to acquire vehicle body dynamics time series data, vehicle control time series data, and environment perception time series data during vehicle driving; a data combination module, configured to combine the vehicle body dynamics time series data, the vehicle control time series data, and the environment perception time series data into multi-channel time series data; A feature extraction module is used to determine the adaptive modal number of the time series data under each channel, and perform modal decomposition on the time series data based on the adaptive modal number to extract the frequency domain features of the time series data; A suspension control parameter prediction module is configured to integrate frequency domain features of time series data from multiple channels, use the integrated multimodal frequency domain features as input, call a trained suspension control parameter prediction model, extract time series features from the multimodal frequency domain features, perform weighted fusion on the extracted time series features, obtain target time series features for predicting suspension control parameters, and predict vehicle suspension control parameters based on the target time series features. The suspension control parameter prediction model is trained using a long short-term memory network with a self-attention mechanism based on the frequency domain features of historical multimodal time series data under various road conditions. A suspension control module is configured to control a suspension system of the vehicle based on the vehicle suspension control parameters.

10. A vehicle comprising a vehicle body, a memory and a controller, characterized in that: The memory stores a computer program, and is characterized in that when the controller executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

11. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

13. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images