An integrated electronically controlled automobile chassis suspension control system

Abstract

The application discloses an integrated electric control type automobile chassis suspension control system and relates to the technical field of image recognition. The integrated electric control type automobile chassis suspension control system comprises a data acquisition module, an AI prediction module, a control execution module and a suspension control parameter. The data acquisition module is used for collecting vehicle state data, road surface data and driving behavior data in real time. The AI prediction module is used for extracting spatial features based on the road surface data. The control execution module is used for constructing a comfort evaluation index based on the suspension control parameter, referencing deep reinforcement learning based on the comfort evaluation index to optimize the suspension control parameter, and strengthening key spatial feature extraction of the road surface to accurately focus on abnormal areas such as pits and protrusions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, and in particular to an integrated electric control type automobile chassis suspension control system. BACKGROUND

[0002] Early automobile suspensions are mainly simple mechanical structures, such as steel plate spring suspensions and coil spring suspensions, which can only provide basic support and cushioning functions and have poor adaptability to complex road conditions and driving conditions; with the development of electronic technology, people began to explore electric control suspension technology, and through the intervention of a simple electronic control unit, limited damping adjustment was realized.

[0003] With the development of intelligent and networked vehicles, vehicles need to cope with more complex and variable driving scenarios, and traditional suspensions have been difficult to meet the high requirements of users for ride experience and driving performance. The existing system has limitations in data processing and feature extraction: in data acquisition, there is a lack of multi-dimensional and high-precision synchronous perception of vehicle state, road information and driving behavior, making it difficult to build a comprehensive driving scenario model; in feature processing, the extraction of road space features lacks pertinence and cannot effectively focus on key road abnormalities such as bumps and potholes, and the processing capability of vehicle state and driving behavior time series is insufficient, traditional time series models are complex to calculate and difficult to highlight the features of key scenes such as sudden acceleration and turning, resulting in limited prediction accuracy of suspension control parameters; and the control strategy of the existing system is mostly based on single target optimization, without combining human comfort evaluation and multi-task learning mechanism, making it difficult to achieve dynamic balance between comfort, handling and energy efficiency, and unable to meet the needs of intelligent vehicles for fine and forward-looking suspension control.

[0004] In view of the above technical defects, a solution is proposed. SUMMARY

[0005] The purpose of the present application is to provide an integrated electric control type automobile chassis suspension control system to solve the problems in the background art, which comprises data acquisition, AI prediction and control execution modules; vehicle state, road and driving behavior data are collected, features are extracted using an improved CNN model and GRU-Attention, suspension parameters are predicted through a fully connected network, and deep reinforcement learning optimization is combined to improve comfort and handling.

[0006] The purpose of the present application can be achieved by the following technical solution: an integrated electric control type automobile chassis suspension control system, comprising:

[0007] A data acquisition module acquires vehicle state data, road data and driving behavior data in real time;

[0008] An AI prediction module extracts spatial features based on the road surface data; inputs vehicle state and driving behavior into a GRU-Attention architecture to extract time sequence features; inputs the spatial features and the time sequence features into a fully connected network to predict suspension control parameters;

[0009] The suspension control parameters include optimal damping force, target suspension height, and vibration increment prediction.

[0010] A control execution module constructs a comfort evaluation index based on the suspension control parameters, and optimizes the suspension control parameters based on the comfort evaluation index using deep reinforcement learning.

[0011] Preferably, the vehicle state data is collected by a high-precision inertial measurement unit and a wheel speed sensor, and is represented as V.

[0012] The road surface data is obtained by constructing a road surface digital twin model, and is represented as R.

[0013] The driving behavior data is generated in real time by combining three dimensions of driving scene, driving operation category, and specific scene and operation, and is represented as D.

[0014] Preferably, the process of extracting the spatial features is as follows:

[0015] An improved CNN model is constructed based on a convolutional neural network architecture and by introducing a CBAM convolution block attention module.

[0016] The input road surface data R is operated by a convolutional neural network to extract spatial unevenness features, outputting preliminary spatial unevenness features; a CBAM convolution block attention module is connected to the deep network after preliminary spatial unevenness feature extraction; by combining channel and spatial dual attention mechanisms, the output spatial features .

[0017] Preferably, the process of extracting the time sequence features is as follows:

[0018] Vehicle state data V and driving behavior data D are obtained, and a GRU-Attention architecture is used to dynamically adjust the vehicle state data V and the driving behavior data D using an attention mechanism, including feature screening, memory updating, and parameter optimization; an attention mechanism is introduced to calculate time sequence step weights, and dynamic time sequence features are obtained by fusing hidden states at each time .

[0019] Preferably, the process of predicting the suspension control parameters is as follows:

[0020] Obtain spatial features and time sequence features , concatenate the spatial features with timing characteristics Concatenate by dimension, form fusion feature vector, train through fully connected network, predict suspension control parameters.

[0021] Preferably, the process of training the fully connected network is:

[0022] Input the fusion feature vector into the fully connected network, and extract and abstract features layer by layer through the weighted calculation and nonlinear activation of multiple neurons inside the network; in the training process, the best damping force , target suspension height and vibration increment prediction are the goals, and multi-task learning loss is used for loss function optimization; the loss function considers the errors of the three prediction tasks comprehensively, and the formula is:

[0023]

[0024] Among them, is the loss function; , and are bias parameters for measuring the mean square error of the predicted damping force and the true value; , and are weight coefficients, which are adjusted according to the gradient size and direction using the gradient descent method.

[0025] Preferably, the process of adjusting the weight coefficient and bias parameter is:

[0026] Calculate the gradient of each layer weight and bias in back propagation, the gradient direction is the direction in which the loss function grows fastest, along the opposite direction of the gradient, that is, the loss reduction direction, multiply the gradient by the set learning rate to get the parameter adjustment amount; update the weight and bias according to the parameter adjustment amount, so that the loss function gradually decreases, and complete a parameter optimization iteration; through continuous iteration, the loss function is minimized.

[0027] Preferably, the process of constructing the comfort evaluation index is:

[0028] Through frequency domain analysis of the vibration acceleration suffered by different parts of the human body in the vertical and horizontal directions, combined with the weighting function W(f), where f is the vibration frequency, the weighted acceleration root mean square value is calculated, and then the comfort index C is mapped.

[0029] Preferably, the process of optimizing the suspension control parameters is:

[0030] When using deep reinforcement learning to optimize the suspension control parameters, first define the state space :

[0031] , wherein, is the energy recovery amount, according to the energy recovery amount in the driving mode, a dynamic adjustment reward function r is designed, and the formula is:

[0032]

[0033] wherein C is the human comfort level; is the optimal damping force, is the target suspension height and is the vibration increment prediction.

[0034] The beneficial effects of the present application are as follows:

[0035] (1) The present application realizes comprehensive capture of vehicle state, road surface and driving behavior data by precise collection of multi-dimensional data through high-precision sensors and multi-modal perception system, provides a high-quality data basis for suspension control, and through the improvement of the CNN model combined with the CBAM convolution block attention module, the key spatial feature extraction of the road surface is strengthened, the abnormal areas such as pits and protrusions are accurately focused, and the perception accuracy of the road surface environment is improved.

[0036] (2) The present application highlights key scene data such as sudden acceleration and turning through GRU-Attention architecture optimization of time sequence feature processing, captures vehicle dynamic changes within milliseconds, improves the real-time performance and pertinence of suspension parameter prediction, and through the combination of spatiotemporal feature decoupling and multi-task learning, the full connection network optimizes parameters such as the optimal damping force and the target suspension height, dynamically adjusts the weight through the loss function, and comprehensively improves the prediction accuracy.

[0037] (3) The present application dynamically optimizes the suspension parameters by combining the ISO standard comfort evaluation with deep reinforcement learning; and defines a state space containing the energy recovery amount, designs a dynamic adjustment reward function suitable for different driving modes, and can balance comfort, energy efficiency and handling, and accurately adapt to various energy recovery requirements and complex driving conditions. BRIEF DESCRIPTION OF DRAWINGS

[0038] The present application will be further described below in conjunction with the accompanying drawings;

[0039] Figure 1 is the system flowchart of the present application;

[0040] Figure 2 is the AI prediction flowchart of the system of the present application. DETAILED DESCRIPTION

[0041] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be clearly and completely described below, obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work belong to the protection scope of the present application.

[0042] Embodiment: please refer to Figure 1 , Figure 2 The embodiment provides an integrated electric control automobile chassis suspension control system, which comprises the following modules.

[0043] A data acquisition module acquires vehicle state data, road surface data and driving behavior data in real time.

[0044] The vehicle state data :

[0045] As a core parameter set for real-time feedback of vehicle dynamic characteristics, the vehicle state data vector is acquired by a high-precision inertial measurement unit and a wheel speed sensor in cooperation; wherein the linear speed vx is calculated by using a Kalman filtering algorithm on the wheel speed pulse signal, and the accuracy can reach 0.1 m / s, so that the speed change of the vehicle under complex working conditions can be captured; the linear acceleration aj is output by a three-axis accelerometer, and after temperature compensation and zero offset calibration, the resolution reaches , and the acceleration fluctuation caused by slight bumps of the road surface can be identified. The angular velocity includes the rotation rate of the vehicle body in three directions of yaw, pitch and roll, is acquired by a gyroscope array and is subjected to attitude calculation, and the unit is rad / s; the vehicle body pitch angle and the roll angle use a multi-sensor data fusion technology, combine IMU and visual odometry information, and control the measurement accuracy to , so as to provide an attitude reference for dynamic adjustment of the suspension system.

[0046] The road surface data :

[0047] A multi-modal perception system based on laser radar point cloud and visual image constructs a road surface digital twin model by using a three-dimensional reconstruction algorithm. The mj rows in the matrix represent the sampling point sequence along the driving direction of the vehicle, and the nl columns contain multi-dimensional information such as coordinates and height; wherein pc represents the row index of the matrix , corresponding to the sampling points of the road surface at different positions in the longitudinal direction; qj represents the column index of the matrix , corresponding to the sampling points of the road surface at different positions in the transverse direction, and the matrix element The characteristic parameters of the corresponding road surface position of the pc-th row and the qj-th column are recorded, such as unevenness, friction coefficient, etc.; the road unevenness is mapped to a quantized value of 0-50 (mm / m) according to the international road unevenness index (IRI) standard, and is used to evaluate the overall flatness of the road; the local slope is calculated by fitting the sampling points by the least square method, and the slope change within ±15° can be identified.

[0048] For the size of the sudden obstacle, the system uses a deep learning algorithm to perform semantic segmentation on the laser radar point cloud, and combines the depth information of the visual image to accurately measure the depth of the pothole and the height of the protrusion, providing key data for suspension preview control.

[0049] Driving behavior data :

[0050] wherein, is an element of the tensor D, ib generally represents different driving scene dimensions, such as high-speed driving, urban congestion, and curve driving; jc represents driving operation categories, such as acceleration, braking, and steering; and kp represents specific parameter indicators under specific scenes and operations, such as acceleration size, braking force, and steering angle; through the combination of the three dimensions (ib, jc, kp), complex driving behavior can be comprehensively described; the three-dimensional tensor is generated in real time by a driver operation intention recognition system, and fuses a steering wheel angle sensor, an acceleration / braking pedal stroke sensor, and historical driving data. The longitudinal acceleration / braking intensity is quantized in the interval [-1, 1], 1 represents full throttle acceleration, and -1 represents emergency braking, and the accurate mapping of the pedal stroke and the power output is realized through a PID control algorithm; the lateral steering angular velocity is collected by a steering system angular velocity sensor, and is converted into rad / s units in combination with the steering ratio parameter, and is used to predict the vehicle steering trend.

[0051] The AI prediction module obtains road surface data and inputs them into an improved CNN model to extract spatial features; obtains vehicle state and driving behavior and inputs them into an improved LSTM time series feature to extract time series features, and splices the spatial features and the time series features to input them into a fully connected network to predict suspension control parameters;

[0052] The process of extracting spatial features based on road surface data is as follows:

[0053] An improved CNN model is constructed based on a convolutional neural network (CNN).

[0054] After the conventional convolutional neural network, a CBAM convolution block attention module is introduced. The CBAM convolution block attention module can dynamically allocate attention weights according to the importance of the channel and spatial dimensions of the input feature map, enhance the attention ability of the model to key features, and emphasize the extraction of key static information such as deceleration zones and pothole profiles.

[0055] Specifically, for the input road surface data , spatial unevenness features are extracted by convolutional neural network operations such as dilated convolution, and preliminary spatial unevenness features are output.

[0056] A CBAM convolutional block attention module is connected to the deep network after preliminary spatial unevenness feature extraction; the CBAM convolutional block attention module is composed of two sub-modules: a channel attention module and a spatial attention module.

[0057] Channel attention module: through global average pooling and fully connected layers, the dependency relationship between feature channels is learned, and different weights are assigned to different channels, thereby strengthening the feature channels related to road obstacle shape, height, etc.

[0058] Spatial attention module: by performing global maximum pooling and / or average pooling in the channel dimension, a two-dimensional spatial attention map is generated. This spatial attention map is used to focus on key spatial regions in the feature map, such as pits, protrusions, and other significant abnormal regions, and to enhance the features of these regions while suppressing the responses of background or irrelevant regions. This mechanism enables the model to more effectively capture local details and spatial structures of the road surface.

[0059] By combining channel and spatial dual attention mechanisms, the improved CNN model can adaptively highlight key information in road point cloud data. The final output spatial features , are representations enriched with key static features after attention mechanism enhancement.

[0060] Specifically, attention-enhanced extraction:

[0061] S201: Connect a CBAM convolutional block attention module to the deep network for preliminary spatial unevenness feature extraction.

[0062] S202: In the channel attention module, the dependency relationship between feature channels is learned through global average pooling and fully connected layers, and higher weights are assigned to feature channels related to road obstacle shape and height to strengthen key channel information.

[0063] S203: In the spatial attention module, global maximum pooling or average pooling is performed in the channel dimension to generate a two-dimensional spatial attention map, focusing on pits, protrusions, and other significant abnormal regions, enhancing key spatial region features, and suppressing background interference. After CBAM convolutional block attention module enhancement, the output spatial feature vector is a structured representation of road key static features.

[0064] Time series feature extraction:

[0065] In real-time data stream processing generated by vehicle-mounted sensors, LSTM network is usually used for processing. The traditional LSTM network has problems such as high computational complexity and insufficient data processing capability in key scenarios. Although the traditional LSTM network has the ability to capture long and short term dependencies, it has obvious shortcomings. On the one hand, the LSTM network contains multiple gating structures such as forget gate, input gate and output gate. The complex calculation logic leads to high consumption of computing resources and low efficiency in model training and inference process, which is difficult to meet the strict requirements of real-time performance of vehicle-mounted systems. On the other hand, LSTM treats all time sequence data equally and cannot focus on the time sequence data of key driving scenarios such as sudden acceleration, emergency braking and corner driving, which makes it difficult for the model to accurately extract key information and affects the accurate analysis and prediction of vehicle dynamic behavior.

[0066] For the time sequence data in vehicle state data V and driving behavior data D, the LSTM network is improved to a gated recurrent unit-attention (GRU-Attention) architecture, and the key time steps are strengthened. Specifically, the forget gate and input gate are combined in the gated recurrent unit GRU, which is simplified to the update gate and reset gate, which reduces the computational complexity while retaining the long and short term memory capability, and is more suitable for processing real-time data streams generated by vehicle-mounted sensors. After introducing the attention mechanism, the model can automatically learn the weight distribution of different time steps and give higher weights to the time sequence data corresponding to key driving scenarios such as sudden acceleration, emergency braking and corner driving.

[0067] When the vehicle is turning, the vehicle-mounted sensors will collect multi-dimensional time sequence data such as steering angle, lateral acceleration, wheel speed difference, vehicle speed and suspension displacement in real time. In the traditional LSTM network, these data are processed in order and the weights of each time step are relatively average, but the GRU-Attention architecture dynamically adjusts through the attention mechanism, including feature screening, memory updating and parameter optimization. GRU updates the hidden state through the gating mechanism, so that the model can better capture long-term dependencies when processing long sequence data, overcoming the defects of traditional RNN. GRU update formula.

[0068] Memory update, based on vehicle state data V, by determining the sudden change of steering angle, when the sudden change of steering angle is detected, the reset gate is activated, and the irrelevant historical information is cleared to focus on the current cornering state; at the same time, the update gate dynamically adjusts the memory retention ratio according to the change of lateral acceleration to avoid noise accumulation.

[0069] Specifically, the reset gate :

[0070]

[0071] wherein, is the reset gate output at time t, which is used to control the hidden state at the previous time How much information can be passed to the candidate hidden state at the current time step; is a sigmoid activation function, which maps the calculation result to the interval [0, 1]; is the weight matrix for linear transformation of the previous time step hidden state in the reset gate; is the hidden state at time step t-1, which contains the time series information at the previous time step, and is the key carrier for storing historical information in the GRU network; is the weight matrix for linear transformation of the current time step input in the reset gate; is the vehicle state V and driving behavior data D input to the GRU network at time step t.

[0072] Update gate :

[0073]

[0074] wherein, is the update gate output at time step t, which is used to control the contribution of the previous time step hidden state and the current time step candidate hidden state to the final hidden state at the current time step; is the weight matrix for linear transformation of the previous time step hidden state in the update gate; is the weight matrix for linear transformation of the current time step input in the update gate, which has a similar effect as and cooperates to complete the fusion calculation of the input and historical state. Feature selection: By calculating the correlation between each time step and the "turning" scene, higher weights are assigned to core features such as steering angle (representing driver intent), lateral acceleration (reflecting centrifugal force influence), and wheel speed difference (reflecting the motion difference between inside and outside wheels), and the influence of non-key data such as engine speed is weakened.

[0075] Specifically, the candidate hidden state

[0076] :

[0077]

[0078] wherein, is the candidate hidden state calculated at time step t, which, under the action of the reset gate, combines the previous time step hidden state and the current input to attempt to generate a new hidden state representation; tanh is the hyperbolic tangent activation function, and the formula is ​Mapping the calculation results to the interval [-1, 1] can increase the nonlinear expression capability of the model; is the hidden state of the previous moment after being processed by the reset gate in the candidate hidden state calculation, , is the element-wise multiplication operator. is the weight matrix for linear transformation of the input at the current moment in the candidate hidden state calculation, together construct a linear transformation relationship that fuses input and historical information.

[0079] Hidden state update :

[0080]

[0081] where, is the final hidden state at moment t, which fuses the information of the hidden state of the previous moment and the candidate hidden state at the current moment, and the fusion ratio of the two is controlled by the update gate , and the formula "·" is also an element-wise multiplication operator. By introducing the attention mechanism, the time step weight

[0082] is calculated, and the formula is:

[0083]

[0084] where, is the scoring function (such as dot product , q is the query vector), and exp is the exponential function, which is used here to calculate the weight proportion of the scoring function result; is the scoring function used to calculate the relevance score between the hidden state at moment t and the query vector q; T is the total length of the time series data (total steps), indicating that the weights of the hidden states at T moments are calculated and normalized; is the hidden state at moment k, and similarly, k is traversed from 1 to T to calculate the weight of each moment; q is the query vector, which is a reference vector used to calculate the relevance of each moment hidden state; the final time series feature , the formula is:

[0085]

[0086] Output dynamic time series feature vector where L is the time series feature dimension. ​

[0087] Parameter optimization: Based on the weighted timing features, the model predicts the optimal suspension parameter combination under the current working condition. For example, in continuous curves, the system will predict in advance and increase the damping force of the outer suspension to suppress body roll; in the acceleration stage after the curve, the damping is quickly reduced to reduce power loss.

[0088] Through this mechanism, the system can accurately capture the change pattern of driving behavior and vehicle state within milliseconds, and then optimize the dynamic adjustment strategy of suspension damping force, spring stiffness and other control parameters. Compared with traditional control algorithms, the response speed is improved, and the ride comfort is improved in complex working conditions.

[0089] Traditional automobile suspension control is often limited by the problems of prediction accuracy and real-time performance. Through in-depth research, the innovation is realized from the decoupling of space-time features and multi-task learning.

[0090] In traditional road-vehicle state prediction, a single feature prediction method is usually used. For example, only the roughness of the road is considered to predict the vibration of the vehicle. This method cannot fully consider the complex dynamic relationship between the road and the vehicle, resulting in large prediction errors. Moreover, traditional methods do not effectively distinguish between time and space features when processing data, making it difficult to meet the real-time and accuracy requirements of high-precision suspension control in different speeds and road conditions.

[0091] Therefore, the innovative method of space-time feature decoupling and multi-task learning is proposed. Space-time feature decoupling is to separate the time and space information in road and vehicle state data, i.e. extracting spatial features and timing features as mentioned above. For example, when a vehicle is driving on a curve with different curvatures, the curvature change of the curve affects the vehicle roll, and the change of the vehicle speed also affects the roll. Through decoupling, the influence of different dimensional features on the vehicle state can be more accurately analyzed.

[0092] Multi-task learning is to handle multiple related tasks at the same time, such as predicting the road bump degree while predicting the tire contact force, body pitch angle, etc. Through the shared bottom feature extraction network, each task is mutually complementary and mutually promoting, for example, the prediction result of road bump degree can help to more accurately judge the tire contact force, and vice versa, thereby significantly improving the overall prediction accuracy.

[0093] The process of predicting suspension control parameters based on a fully connected network is as follows:

[0094] The spatial features and the timing features After splicing, the complex relationships in the data are mined through feature learning and mapping of the fully connected network, and the precise prediction of the suspension control parameters, including the optimal damping force, target suspension height and vibration increment prediction, is realized.

[0095] From the data preprocessing, network training and prediction output stages, the optimal damping force, target suspension height and vibration increment prediction are obtained:

[0096] Data preprocessing and feature fusion:

[0097] Spatial features Generally contain static information of the current driving environment of the vehicle, such as road slope, bend curvature, etc., reflecting the geometric and physical characteristics of the space where the vehicle is located; time series features Record the dynamic changes in the vehicle driving process, such as suspension displacement, vehicle speed fluctuations over time. Splice these two types of features by dimension to form a fusion feature vector, which integrates the information of spatial and temporal dimensions, providing a data basis for subsequent prediction.

[0098] Fully connected network training:

[0099] The fusion feature vector is input into the fully connected network, and the network extracts and abstracts features layer by layer through weighted calculation and nonlinear activation of multiple neurons inside the network. In the training process, the optimal damping force , target suspension height and vibration increment prediction are the targets, and multi-task learning loss is used for optimization. The loss function considers the errors of the three prediction tasks, and the formula is:

[0100]

[0101] Among them, is the loss function; , and are weight coefficients, and . used to measure the mean square error between the predicted value of the damping force and the true value , and are the same. Through the back propagation algorithm, the network parameters are constantly adjusted to minimize the loss function and improve the prediction accuracy of the network.

[0102] Specifically, through the back propagation algorithm, the loss function of each layer parameter is propagated back layer by layer from the output layer. In the process of neural network training, the spatial features and time series features The input network is used to obtain a predicted suspension control output through forward propagation. At this time, the predicted output is compared with the actual ideal suspension control state, and the error between the two is calculated, which is quantified by a loss function.

[0103] The connection weights and bias parameters between neurons in the network are adjusted according to the size and direction of the gradient using the gradient descent method. The direction of the gradient represents the direction in which the loss function grows fastest. Along the opposite direction of the gradient, the parameter adjustment amount is obtained by multiplying the gradient by the learning rate, and the weights and biases are updated to gradually reduce the loss. In each backpropagation process, after calculating the gradient of each layer parameter, the weights and biases are updated according to a certain learning rate. The learning rate determines the step size of parameter updating. If it is set too large, it may cause the parameter to miss the optimal solution or even diverge. If it is set too small, the training process will become extremely slow.

[0104] Predicted output:

[0105] By continuously repeating the process of data input, loss calculation, backpropagation, and parameter adjustment, the prediction accuracy of the neural network for the suspension control state of the vehicle chassis will gradually improve. As the training progresses, the network can better learn the mapping relationship between input data such as road conditions, vehicle speed, and ideal suspension control strategies, thereby achieving more accurate suspension control strategy optimization. This allows the vehicle to automatically adjust the damping, stiffness, and other parameters of the suspension according to different road conditions during driving, improving ride comfort and vehicle handling stability. This process is like self-learning and self-improvement for the neural network, which continuously adapts to complex and variable driving conditions of the vehicle to achieve the best suspension control effect and minimize the loss function.

[0106] The trained fully connected network can quickly output the optimal damping force, target suspension height, and vibration increment prediction when receiving new fused feature vectors.

[0107] The control execution module constructs a comfort evaluation index based on the suspension control parameters and optimizes the suspension control parameters using deep reinforcement learning.

[0108] The process of constructing the comfort evaluation index is as follows:

[0109] In the comfort evaluation and parameter optimization process of the vehicle chassis suspension control system, when constructing the comfort evaluation index C, the human body vibration perception model is used, and the ISO2631-1 vibration evaluation standard is strictly followed. This standard quantifies human comfort as C by weighting the human body's response to vibration at different frequencies, and where 1 represents the most uncomfortable state.

[0110] Specifically, by performing frequency domain analysis on the vibration acceleration a(t) of different parts of the human body in the vertical, horizontal and other directions, combined with the weighting function W(f), where f is the vibration frequency, the weighted acceleration root mean square value is calculated , and then the comfort index C is obtained by mapping, and the formula is:

[0111] ;

[0112] wherein, , is the lower limit and upper limit of the vibration frequency analysis, which defines the frequency range of the frequency domain analysis of the vibration signal, ensuring that the vibration frequency interval affecting the human body is covered; is the frequency domain representation of the vibration acceleration, which converts the time domain vibration acceleration into a function of frequency f by Fourier transform or similar frequency domain analysis method.

[0113] When using deep reinforcement learning (DRL) to optimize suspension control parameters, first define the state space; wherein, represents the energy recovery amount, which reflects the energy recovered by the suspension system through dampers and other devices during the movement process.

[0114] According to the energy recovery amount in the driving mode, the reward function is dynamically adjusted to r, and the formula is:

[0115]

[0116] wherein C is the human comfort based on the ISO2631-1 standard described above; , and are weight parameters, for example, set by adjusting the proportion of the three to balance comfort, energy efficiency and control smoothness. is the vibration increment prediction.

[0117] Task weights , and will be dynamically adjusted according to the driving mode. In the comfort mode, in order to improve the ride comfort, increase the weight of , so that the network pays more attention to the accurate prediction of the optimal damping force, so as to effectively suppress the suspension vibration; in the sports mode, increase the weight of , let the network focus on optimizing the prediction of the target suspension height, and improve the vehicle handling performance.

[0118] The above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented by software, the above-described embodiments can be implemented in whole or in part in the form of a computer program product. A person of ordinary skill in the art can be aware that units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether the functions are performed by hardware or software depends on the specific application and design constraints of the technical solutions.

[0119] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, and can be located in one place or distributed on multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiments according to actual needs.

[0120] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. An integrated electronically controlled automotive chassis suspension control system, characterized by, The application relates to a vehicle suspension control method based on deep reinforcement learning, comprising the following steps: a data acquisition module acquires vehicle state data, road surface data and driving behavior data in real time; the vehicle state data is acquired by a high-precision inertial measurement unit and a wheel speed sensor, and is represented as V; the road surface data is obtained by constructing a road surface digital twin model and mapping original point cloud data, and is represented as R; the driving behavior data is generated in real time by combining three dimensions of driving scene, driving operation category, specific scene and operation, and is represented as D; an AI prediction module extracts spatial features based on the road surface data; vehicle state and driving behavior are input into a GRU-Attention architecture to extract time sequence features, the spatial features and the time sequence features are spliced into a full connection network, and a suspension control parameter is predicted; the process of extracting the spatial features is as follows: a convolutional neural network is used as a basic architecture, a CBAM convolution block attention module is introduced, and an improved CNN model is constructed; The input road surface data R is operated through a convolutional neural network to extract spatial unevenness features and output preliminary spatial unevenness features; a CBAM convolution block attention module is connected to a deep network after preliminary spatial unevenness feature extraction; and spatial features are output through the combination of channel and spatial double attention mechanisms ; the process of extracting the time sequence features is as follows: The vehicle state data V and the driving behavior data D are acquired, the vehicle state data V and the driving behavior data D are dynamically adjusted by using a GRU-Attention architecture, feature screening, memory updating and parameter optimization are included, a timing step weight is calculated by introducing an attention mechanism, and dynamic timing features are obtained by fusing hidden states at each moment ; the suspension control parameter comprises an optimal damping force, a target suspension height and a vibration increment prediction; a control execution module constructs a comfort evaluation index based on the suspension control parameter, and optimizes the suspension control parameter based on the comfort evaluation index by using deep reinforcement learning.

2. The integrated electronically controlled chassis suspension control system of claim 1, wherein, the process of predicting the suspension control parameter is as follows: Obtaining spatial features with timing features spatial features with timing features Concatenate by dimension, form fusion feature vector, train through fully connected network, predict suspension control parameters.

3. The integrated electronically controlled chassis suspension control system of claim 2, wherein, the process of training the full connection network is as follows: a fusion feature vector is input into a full connection network, the network is internally calculated by weighting and nonlinear activation of multiple neurons, and features are extracted and abstracted layer by layer; During training, the optimal damping force , target suspension height and vibration increment prediction are targeted, and a multi-task learning loss is used for loss function optimization; the loss function comprehensively considers the errors of the three prediction tasks, and the formula is: ， wherein, is a loss function; , and are bias parameters for measuring the mean square error of the predicted value of the damping force and the true value; , and are weight coefficients, and the weight coefficients and the bias parameters are adjusted according to the size and direction of the gradient using the gradient descent method.

4. The integrated electronically controlled chassis control system for an automotive vehicle according to claim 3, wherein the process of adjusting the weight coefficient and the bias parameter is as follows: in the back propagation, the gradient of each layer weight and bias is calculated, the gradient direction is the direction in which the loss function grows fastest, the parameter adjustment amount is obtained by multiplying the gradient by a set learning rate in the opposite direction of the gradient, that is, the loss reduction direction; the weight and the bias are updated according to the parameter adjustment amount, so that the loss function is gradually reduced, and one parameter optimization iteration is completed; the loss function is minimized through continuous iteration.

5. An integrated electronically controlled vehicle chassis control system according to claim 4, wherein, the process of constructing the comfort evaluation index is as follows: By analyzing the vibration acceleration experienced by different parts of the human body in the vertical and horizontal directions... Frequency domain analysis is performed, and the weighted root mean square value of the acceleration is calculated using the weighting function W(f), where f is the vibration frequency. This is then mapped to the comfort index C.

6. An integrated electronically controlled automotive chassis control system according to claim 5, wherein, the process of optimizing the suspension control parameter is as follows: When optimizing suspension control parameters using deep reinforcement learning, first define the state space , , wherein, is the amount of energy recovered, the dynamic adjustment reward function r is designed according to the amount of energy recovered in the driving mode, and the formula is: ， where C is the human comfort level; is the optimal damping force, is the target suspension height, and is the vibration increment prediction.

Images