Neural network driven vehicle adaptive cruise control method

Abstract

The invention relates to the technical field of vehicle cruise control, and discloses a neural network driven vehicle adaptive cruise control method. Real-time driving data are collected through a multi-modal sensor, spatial-temporal feature fusion is carried out through a multi-layer convolutional neural network to generate dynamic environment sensing data, and a pre-trained adaptive decision model is input to obtain driving strategy parameters. And constructing a mixed integer programming model on the basis, globally planning a cruise path by adopting an incremental branch and bound algorithm integrating a dynamic relaxation threshold and a heuristic pruning strategy, and outputting optimal cruise trajectory data. A hierarchical control framework comprising a planning layer, a coordination layer and an execution layer is constructed, the planning layer generates a global trajectory sequence, the coordination layer dynamically corrects a local trajectory, and the execution layer achieves vehicle longitudinal acceleration and transverse steering angle tracking based on a robust sliding mode control algorithm and outputs a vehicle control instruction to complete self-adaptive cruise control. And the cruise control performance, safety and stability of the vehicle in a complex environment are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle cruise control, and in particular to a vehicle adaptive cruise control method driven by a neural network. Background Art

[0002] With the rapid development of the automotive industry and the increasing demand for driving experience and safety, vehicle adaptive cruise control (ACC) technology has gradually become a research hotspot. Traditional adaptive cruise control methods are mainly based on fixed rules and simple sensor data processing, which have many limitations when facing complex and changing driving environments.

[0003] In terms of sensor data processing, traditional methods often rely on only a single or a few types of sensors, such as using only millimeter-wave radar to monitor vehicle distance and relative speed. This method obtains limited environmental information and makes it difficult to fully perceive the complex situation around the vehicle. Moreover, traditional methods process sensor data in a simple way and do not fully exploit the spatiotemporal features in the data. For example, the point cloud data collected by the lidar and the image data obtained by the visual camera contain rich environmental geometry and texture information, but traditional methods fail to effectively fuse these multimodal data and cannot provide the vehicle with accurate environmental perception, limiting the performance of adaptive cruise control.

[0004] At the decision-making and planning level, traditional cruise control decision models are usually based on preset fixed rules and lack the ability to flexibly respond to complex environments and multi-objective optimization. Its path planning algorithm also often uses a relatively simple search algorithm, which has high computational complexity and low efficiency, and cannot quickly plan a reasonable cruise trajectory in vehicle driving scenarios with extremely high real-time requirements. Especially in congested urban roads, traditional path planning algorithms are prone to fall into local optimal solutions, resulting in unreasonable vehicle driving routes, increasing driving time and energy consumption.

[0005] In terms of vehicle control execution, traditional control algorithms are highly dependent on vehicle dynamics models. However, actual vehicles are affected by various uncertain factors during driving, such as changes in road friction and wear of the vehicle's own components. This makes it difficult for traditional control algorithms to ensure stable driving of the vehicle under different operating conditions. Summary of the invention

[0006] The object of the present invention is to provide a vehicle adaptive cruise control method driven by a neural network to solve the problems raised in the above-mentioned background technology.

[0007] To achieve the above object, the present invention provides the following technical solution: a vehicle adaptive cruise control method driven by a neural network, the method comprising: Collect real-time driving data of vehicles through multimodal sensors; Based on a multi-layer convolutional neural network, the real-time driving data is fused with temporal and spatial features to generate dynamic environment perception data; Inputting the dynamic environment perception data into a pre-trained adaptive decision model to generate driving strategy parameters; A mixed integer programming model is constructed according to the driving strategy parameters, wherein the mixed integer programming model takes maximizing the safe driving distance and balancing the energy consumption as optimization goals, and uses an incremental branch-and-bound algorithm to globally plan the cruise path, wherein the incremental branch-and-bound algorithm integrates a dynamic relaxation threshold and a heuristic pruning strategy; and optimal cruise trajectory data is output based on the mixed integer programming model; A hierarchical control framework is constructed according to the optimal cruise trajectory data, and the hierarchical control framework includes a planning layer, a coordination layer and an execution layer, wherein the planning layer generates a global trajectory sequence based on the driving strategy parameters, the coordination layer uses a rolling time domain optimization algorithm to dynamically correct the local trajectory, and the execution layer realizes the longitudinal acceleration and lateral steering angle tracking of the vehicle based on a robust sliding mode control algorithm; the vehicle control command is output through the hierarchical control framework to complete the adaptive cruise control.

[0008] Preferably, the performing spatiotemporal feature fusion on the real-time driving data based on a multi-layer convolutional neural network to generate dynamic environment perception data includes: Multimodal sensors include LiDAR, millimeter wave radar, inertial measurement unit, visual camera, and ultrasonic sensor; Perform multi-scale voxel processing on the lidar point cloud data and visual camera images to generate a three-dimensional grid feature map; perform time-series alignment on the millimeter-wave radar data and ultrasonic sensor data to construct a time-series distance-velocity matrix; Constructing a dual-channel convolutional neural network, wherein the first channel uses a three-dimensional sparse convolution kernel to extract the geometric features of the three-dimensional grid feature map, and the second channel uses a temporal convolution kernel to extract the dynamic features of the temporal distance-velocity matrix; The geometric features and dynamic features are fused through a cross-channel feature splicing layer to generate a joint feature tensor; the time dependency of the joint feature tensor is modeled based on a gated recurrent unit to output dynamic environment perception data including the position, speed and motion trend of environmental obstacles.

[0009] Preferably, the adaptive decision model adopts a staged graph attention network structure to perform multi-objective optimization on the driving strategy based on a dynamic priority allocation mechanism; the staged graph attention network structure includes: Construct a vehicle-environment interaction graph, where nodes include the vehicle node, surrounding vehicle nodes, lane line nodes, and obstacle nodes. Node attributes include position, velocity, and acceleration vectors. A two-stage attention mechanism is adopted. In the first stage, the association weights between the vehicle node and the surrounding nodes are calculated through the spatial attention layer. In the second stage, the importance of the historical state sequence is weighted through the temporal attention layer. Based on the multi-head graph attention module, the node features are iteratively updated, and each attention head uses a learnable position encoding to enhance the spatial relative relationship; the training process is stabilized by the residual connection and layer normalization mechanism, and finally the driving strategy parameters containing multi-objective constraints are output.

[0010] Preferably, the incremental branch and bound algorithm integrates a dynamic relaxation threshold and a heuristic pruning strategy including: The path planning problem is modeled as a mixed integer linear programming problem, where the decision variables include discrete lane selection variables and continuous acceleration variables. Initialize the relaxation problem and calculate the initial lower bound, and use a dynamic relaxation threshold adjustment mechanism to adaptively reduce the relaxation range of integer variables according to the number of iterations; In the branching process, the fractional variables that have the greatest impact on the objective function are prioritized for branching; in the pruning stage, a probabilistic pruning model is constructed based on the historical optimal solution, and Bayesian optimization is used to predict invalid branches and prune them in advance; The relaxed problem is iteratively optimized using an incremental solver, where only a subset of local variables is updated in each iteration until the integer feasibility condition is met.

[0011] Preferably, the rolling time domain optimization algorithm dynamically corrects the local trajectory including: A time-varying prediction model is constructed to discretize the vehicle dynamics equation into a state transfer matrix, wherein the state transfer matrix includes longitudinal velocity, lateral deviation, and yaw angle differential terms; Design a sliding window optimization objective function, including trajectory tracking error term, control input smoothing term and obstacle repelling potential energy term; Robust feasibility constraints are introduced, and the uncertain parameters are described by ellipsoid sets through semidefinite programming methods to construct robust linear matrix inequality constraints. The alternating direction multiplier method is used to solve the optimization problem in a distributed manner, and parallel computing is used to accelerate real-time response.

[0012] Preferably, the execution layer implements the vehicle's longitudinal acceleration and lateral steering angle tracking based on the robust sliding mode control algorithm, including: A super-helical sliding surface is designed to couple the longitudinal acceleration error and the lateral steering angle error into a composite sliding mode variable. Construct an adaptive convergence law to dynamically adjust the convergence speed and switching gain according to the error amplitude to suppress high-frequency chattering; A disturbance observer is used to estimate the unmodeled dynamic disturbance and the estimated value is fed forward to the control law. The finite-time convergence of the closed-loop system is verified by Lyapunov stability analysis.

[0013] Preferably, the three-dimensional sparse convolution kernel uses an octree structure to accelerate feature extraction, including: Divide the LiDAR point cloud into multi-level octree nodes, each of which stores the point cloud density and normal vector statistical features; Only non-empty nodes are activated in convolution operations, and multi-resolution geometric details are preserved through skip connections; A dynamic pooling layer is used to perform maximum aggregation on the features of adjacent nodes to generate a compact three-dimensional feature representation.

[0014] Preferably, the spatial attention layer adopts a relative position encoding mechanism, including: Define the relative position vector between the ego vehicle node and the target node, including the Euclidean distance, heading angle difference and relative velocity projection; Map the relative position vector to an attention bias term through a learnable nonlinear transformation; The bias term is added to the standard attention weight calculation process to enhance the geometric interpretability of spatial correlation.

[0015] Preferably, the probability pruning model is constructed based on Gaussian process regression, including: Collect the objective function values ​​and variable relaxations of historical branch nodes as training samples; Construct the covariance kernel function, which includes a combination of linear kernel and Marten kernel; By maximizing the marginal likelihood, we optimize the hyperparameters and predict the feasibility probability of unexplored branches. When the feasibility probability is lower than the dynamic threshold, early pruning is triggered.

[0016] Preferably, the ellipsoid set description is implemented by singular value decomposition, including: The uncertainty parameters are modeled as a set of bounded ellipsoids, with the center vectors as nominal parameters and the shape matrix determined by the parameter covariance; Perform singular value decomposition on the shape matrix to extract the direction and length of the main semi-axis; In the optimization constraints, all major semi-axis directions are required to satisfy linear inequality conditions.

[0017] Compared with the prior art, the present invention has the following beneficial effects: At the environmental perception level, multimodal sensors are used to collect data, and spatiotemporal features are fused through multi-layer convolutional neural networks. Multimodal sensors such as LiDAR, millimeter-wave radar, inertial measurement unit, visual camera and ultrasonic sensor work together to obtain vehicle driving environment information from different dimensions. LiDAR provides high-precision distance information, visual camera captures rich image texture and semantics, millimeter-wave radar monitors relative speed in real time, ultrasonic sensor assists in close-range detection, and inertial measurement unit perceives vehicle posture. Multi-layer convolutional neural network effectively fuses these multi-source heterogeneous data to generate accurate dynamic environment perception data. By multi-scale voxel processing of LiDAR point cloud data and visual camera images, and time-series alignment of millimeter-wave radar data and ultrasonic sensor data, a dual-channel convolutional neural network is constructed to extract geometric features and dynamic features respectively, and then through cross-channel feature splicing and gated recurrent unit modeling, the vehicle can accurately obtain the position, speed and movement trend of environmental obstacles. This greatly improves the vehicle's perception of complex environments. Compared with the traditional method of relying on a single sensor, it can more comprehensively and accurately grasp the surrounding environment information, providing a solid foundation for subsequent decision-making.

[0018] In terms of decision-making and planning, the adaptive decision model adopts a phased graph attention network structure and a dynamic priority allocation mechanism. The constructed vehicle-environment interaction graph covers nodes such as the vehicle, surrounding vehicles, lane lines and obstacles. The node attributes are rich and can accurately reflect the status of each object. In the two-stage attention mechanism, the spatial attention layer calculates the association weights through relative position encoding, enhances the geometric interpretation of spatial association, and enables the vehicle to pay more attention to surrounding objects; the temporal attention layer weights the historical state sequence and makes full use of historical information. The multi-head graph attention module iteratively updates the node features, combines residual connection and layer normalization to stabilize training, and outputs driving strategy parameters containing multi-objective constraints. This model can comprehensively consider multiple objectives such as driving safety, energy consumption, and traffic smoothness in complex traffic scenarios to optimize driving strategies. On highways, the vehicle speed and following distance can be dynamically adjusted according to traffic flow and road conditions, which not only ensures safety but also reduces energy consumption; on urban roads, it can quickly respond to changes in pedestrians and traffic lights, make reasonable decisions, and improve traffic efficiency.

[0019] When planning the path, a mixed integer programming model is constructed based on the driving strategy parameters, and an incremental branch-and-bound algorithm integrating dynamic relaxation threshold and heuristic pruning strategy is adopted. The path planning problem is modeled as a mixed integer linear programming problem, and the decision variables reasonably cover lane selection and acceleration control. The dynamic relaxation threshold adaptively adjusts the relaxation range of integer variables according to the number of iterations to improve the search efficiency; the heuristic pruning strategy uses the historical optimal solution to construct a probabilistic pruning model, prunes invalid branches in advance, and reduces the amount of calculation. The incremental solver iteratively optimizes and only updates a subset of local variables each time. While ensuring the planning accuracy, it quickly outputs the optimal cruise trajectory data, making the vehicle driving path more reasonable and efficient.

[0020] The hierarchical control framework further ensures the accuracy and stability of vehicle control. The planning layer generates a global trajectory sequence based on the driving strategy parameters to provide macro-guidance for vehicle driving; the coordination layer uses a rolling time domain optimization algorithm to build a time-varying prediction model, design an optimization objective function containing multiple constraints, and introduce robust feasibility constraints. Through semi-definite programming and alternating direction multiplier method distributed solution, the local trajectory is dynamically corrected to respond to changes in road conditions in real time; the execution layer is based on a robust sliding mode control algorithm, designs a super-helical sliding mode surface to couple the longitudinal acceleration and lateral steering angle error, constructs an adaptive reaching law to suppress vibration, and uses a disturbance observer to compensate for unmodeled disturbances. The Lyapunov stability analysis ensures that the closed-loop system converges in finite time. This allows the vehicle to accurately track the expected longitudinal acceleration and lateral steering angle under different driving conditions, improve driving safety and comfort, effectively respond to uncertain factors such as changes in road friction and external interference, and ensure stable vehicle driving. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A working principle diagram of the vehicle adaptive cruise control method of the present invention; Figure 2 Flowchart for multimodal data processing and dynamic environment perception data generation; Figure 3 It is the workflow diagram of the adaptive decision model; Figure 4 The following is a workflow diagram of the probability pruning model based on Gaussian process regression. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] See also Figure 1-Figure 4 The present invention provides a technical solution: a vehicle adaptive cruise control method driven by a neural network, and the specific implementation steps are as follows: Multimodal sensors are used to collect real-time driving data of vehicles. These multimodal sensors include lidar, millimeter-wave radar, inertial measurement unit, visual camera and ultrasonic sensor, etc. They obtain various information of the vehicle's driving process from different angles and methods, such as the distance, speed, image information of the surrounding environment and the vehicle's own posture data, providing a basis for subsequent analysis and decision-making.

[0024] Based on a multi-layer convolutional neural network, the collected real-time driving data is fused with spatiotemporal features. The data collected by multimodal sensors have different characteristics and need to be fused through specific processing methods. The laser radar point cloud data and the visual camera image are processed by multi-scale voxelization to generate a three-dimensional grid feature map; the millimeter wave radar data and the ultrasonic sensor data are aligned in time series to construct a time series distance-speed matrix. Then a two-channel convolutional neural network is constructed. The first channel uses a three-dimensional sparse convolution kernel to extract the geometric features of the three-dimensional grid feature map, and the second channel uses a time series convolution kernel to extract the dynamic features of the time series distance-speed matrix. The geometric features and dynamic features are fused through the cross-channel feature splicing layer to generate a joint feature tensor, and then the joint feature tensor is modeled for time dependence based on the gated recurrent unit. Finally, the dynamic environment perception data containing the location, speed and movement trend of environmental obstacles is output, so that the vehicle can have a more accurate perception of the surrounding dynamic environment.

[0025] The generated dynamic environment perception data is input into the pre-trained adaptive decision model, which adopts a staged graph attention network structure and performs multi-objective optimization of driving strategies based on a dynamic priority allocation mechanism. By constructing a vehicle-environment interaction graph, the nodes in the graph include the vehicle node, surrounding vehicle nodes, lane line nodes and obstacle nodes, and the node attributes include position, velocity and acceleration vectors. A two-stage attention mechanism is adopted. In the first stage, the association weights of the vehicle node and the surrounding nodes are calculated through the spatial attention layer, and in the second stage, the importance of the historical state sequence is weighted through the temporal attention layer. The node features are iteratively updated based on the multi-head graph attention module. Each attention head uses a learnable position encoding to enhance the spatial relative relationship. Then, the training process is stabilized through the residual connection and layer normalization mechanism. Finally, the driving strategy parameters containing multi-objective constraints are output to provide a basis for the vehicle's driving decision.

[0026] A mixed integer programming model is constructed based on the driving strategy parameters. The optimization objectives of this model are to maximize the safe driving distance and balance the energy consumption. An incremental branch-and-bound algorithm is used to globally plan the cruise path. The incremental branch-and-bound algorithm integrates dynamic relaxation threshold and heuristic pruning strategy. In the algorithm implementation, the path planning problem is modeled as a mixed integer linear programming problem. The decision variables include discrete lane selection variables and continuous acceleration variables. The relaxation problem is initialized and the initial lower bound is calculated. The dynamic relaxation threshold adjustment mechanism is used to adaptively narrow the relaxation range of integer variables according to the number of iterations. In the branching process, the fractional variables that have the greatest impact on the objective function are preferred for branching. In the pruning stage, a probabilistic pruning model is constructed based on the historical optimal solution, and the invalid branches are predicted and pruned in advance using Bayesian optimization. The relaxation problem is iteratively optimized using an incremental solver. Only a subset of local variables is updated in each iteration until the integer feasibility condition is met. Finally, the optimal cruise trajectory data is output to ensure that the vehicle driving path is both safe and energy-saving.

[0027] A hierarchical control framework is constructed based on the optimal cruise trajectory data. The hierarchical control framework includes a planning layer, a coordination layer, and an execution layer. The planning layer generates a global trajectory sequence based on the driving strategy parameters. The coordination layer uses a rolling horizon optimization algorithm to dynamically correct the local trajectory. The execution layer implements the vehicle's longitudinal acceleration and lateral steering angle tracking based on a robust sliding mode control algorithm. The rolling horizon optimization algorithm constructs a time-varying prediction model, discretizes the vehicle dynamics equation into a state transfer matrix, and designs a sliding window optimization objective function, which includes a trajectory tracking error term, a control input smoothing term, and an obstacle repelling potential energy term. Robust feasibility constraints are introduced, and the uncertainty parameters are described by an ellipsoid set through a semidefinite programming method. Robust linear matrix inequality constraints are constructed, and the alternating direction multiplier method is used to perform distributed solutions to the optimization problem. The execution layer designs a super-helical sliding surface, couples the longitudinal acceleration error and the lateral steering angle error into a composite sliding mode variable, constructs an adaptive approach law, dynamically adjusts the approach speed and switching gain according to the error amplitude, suppresses high-frequency chattering, uses a disturbance observer to estimate the unmodeled dynamic disturbance, and feeds the estimated value forward to the control law, and verifies the finite time convergence of the closed-loop system through Lyapunov stability analysis. The vehicle control command is output through the hierarchical control framework to complete the adaptive cruise control, so that the vehicle can drive safely and stably according to the planned trajectory.

[0028] The present invention will be further described below in conjunction with Examples 1 to 5:

[0029] Example 1: In this example, the specific process of generating dynamic environment perception data by fusing spatiotemporal features based on a multi-layer convolutional neural network and accelerating feature extraction using an octree structure using a three-dimensional sparse convolution kernel is described in detail.

[0030] After the multimodal sensor collects data, the processing of the LiDAR point cloud data and the visual camera image is a key step. When performing multi-scale voxel processing, first determine the size and resolution of the voxel, and select the appropriate scale parameters according to the complexity of the vehicle driving environment and the limitation of computing resources. For example, in the urban road environment, due to the large number of obstacles and frequent changes, a smaller voxel size is selected to capture details more accurately; while in the highway environment, the voxel size can be appropriately increased to reduce the amount of calculation. The LiDAR point cloud data is divided according to the set voxel size, and the point cloud information in each voxel is counted and encoded to generate a three-dimensional grid feature map, which contains the geometric structure information of the environment. At the same time, the visual camera image is also voxelized, the image pixels are matched with the voxels, and the color, texture and other information in the image are integrated into the three-dimensional grid feature map.

[0031] For millimeter wave radar data and ultrasonic sensor data, because they collect information about distance and speed, and there are differences in the collection time, time alignment is required. Through the time synchronization algorithm, the data collected at different times are adjusted to the same time scale to construct a time series distance-speed matrix. This matrix records the changes in the distance and speed of objects around the vehicle at different time points, providing data support for the subsequent extraction of dynamic features.

[0032] When constructing a dual-channel convolutional neural network, the first channel uses a three-dimensional sparse convolution kernel. The three-dimensional sparse convolution kernel uses an octree structure to accelerate feature extraction. The specific operation is as follows: the lidar point cloud is divided into multi-level octree nodes, and each node stores the point cloud density and normal vector statistical features. In the convolution operation, only non-empty nodes are activated, which greatly reduces the amount of calculation. At the same time, multi-resolution geometric details are retained through jump connections, so that geometric features can be effectively extracted at different scales. The dynamic pooling layer is used to maximize the features of adjacent nodes to generate a compact three-dimensional feature representation. For example, in a certain layer of the octree, several adjacent nodes may contain different geometric features. The most representative features are selected through the dynamic pooling layer to form a more concise and effective feature representation. The second channel uses a temporal convolution kernel to perform a convolution operation on the temporal distance-velocity matrix to extract its dynamic features, such as the movement trend of the object, the rate of change of velocity, etc.

[0033] Through the cross-channel feature concatenation layer, the geometric features extracted from the first channel are fused with the dynamic features extracted from the second channel to generate a joint feature tensor. The joint feature tensor contains the static geometric information and dynamic change information of the environment, and its time dependency is modeled based on the gated recurrent unit. The gated recurrent unit can adaptively adjust the transmission and update of information according to the current input and the state of the previous moment, thereby outputting dynamic environmental perception data including the location, speed and movement trend of environmental obstacles. Through such processing, the vehicle can understand the dynamic changes of the surrounding environment more comprehensively and accurately, providing a reliable basis for subsequent decision-making and control.

[0034] Example 2: This example focuses on the staged graph attention network structure adopted by the adaptive decision model and the relative position encoding mechanism adopted by the spatial attention layer.

[0035] When constructing a vehicle-environment interaction graph, it is crucial to accurately determine the attributes and connection relationships of the nodes in the graph. The ego vehicle node represents the vehicle, and its attributes include the vehicle's position, velocity, and acceleration vector, which can be obtained through the vehicle's own sensors and control system. The surrounding vehicle nodes represent other vehicles around the vehicle, and also contain position, velocity, and acceleration vectors. This information is collected by multimodal sensors and obtained through data fusion processing. The lane line node records the position, curvature, and other information of the lane line, and the obstacle node represents the obstacles on the road and their related attributes. The connection relationship between the nodes is determined based on the spatial position and interaction between the vehicle and the surrounding objects. For example, there is a relative position and speed relationship between the ego vehicle and the surrounding vehicles, and these relationships are reflected by connecting edges.

[0036] A two-stage attention mechanism is adopted. In the first stage of the spatial attention layer, a relative position encoding mechanism is adopted. The relative position vector of the ego vehicle node and the target node is defined, including the Euclidean distance, the heading angle difference and the relative speed projection. The Euclidean distance reflects the spatial distance between the ego vehicle and the target node, the heading angle difference reflects the difference in the driving directions of the two, and the relative speed projection represents the speed component of the two in the relative direction. The relative position vector is mapped to the attention bias term through a learnable nonlinear transformation. This nonlinear transformation can be implemented using a neural network structure such as a multi-layer perceptron. The bias term is superimposed on the standard attention weight calculation process to enhance the geometric interpretation of spatial correlation. For example, when calculating the attention weights of the ego vehicle node and the surrounding vehicle nodes, the bias term obtained by the nonlinear transformation of the relative position vector will make the attention weight more inclined to vehicles with a closer distance, similar driving direction and smaller relative speed, thereby more accurately reflecting the spatial relationship and mutual influence between vehicles.

[0037] In the second stage, the temporal attention layer weights the importance of the historical state sequence. The historical state sequence is formed by recording the state information of the vehicle and the surrounding environment at multiple time points. The temporal attention layer weights the information at different moments in the historical state sequence according to the current environmental changes and decision-making needs, so that the recent information related to the current decision is given a greater weight, thereby better utilizing historical information for decision-making.

[0038] Based on the multi-head graph attention module, the node features are iteratively updated, and each attention head uses a learnable position encoding to enhance the spatial relative relationship. The multi-head graph attention module can simultaneously pay attention to and update node features from different angles to improve the expressiveness of the model. The learnable position encoding of each attention head is implemented through a specific encoding method, such as the combined encoding of sine and cosine functions. The training process is stabilized by residual connections and layer normalization mechanisms. The residual connection makes it easier for the network to converge during training and avoids the gradient vanishing problem; the layer normalization mechanism normalizes the input of each layer to make the network training more stable. Finally, the adaptive decision model outputs driving strategy parameters containing multi-objective constraints. These parameters comprehensively consider multiple objectives such as vehicle driving safety, comfort, efficiency, etc., and provide a reasonable basis for vehicle driving decisions.

[0039] Embodiment 3: This embodiment introduces in detail the process of integrating the incremental branch and bound algorithm with the dynamic relaxation threshold and the heuristic pruning strategy and constructing the probabilistic pruning model based on Gaussian process regression.

[0040] When modeling the path planning problem as a mixed integer linear programming problem, the selection and definition of decision variables are critical. The discrete lane selection variable is used to represent the lane selected by the vehicle during driving. For example, integers such as 0, 1, and 2 are used to represent different lane numbers. The lane selection of the vehicle at a certain moment can be described by this variable. The continuous acceleration variable is used to control the acceleration and deceleration of the vehicle. It is a continuous value that can be adjusted according to the vehicle's dynamic model and driving requirements.

[0041] Initialize the relaxation problem and calculate the initial lower bound. First, determine the constraints and objective function of the relaxation problem. Constraints include vehicle dynamics constraints, road rules constraints, such as the maximum acceleration of the vehicle, the minimum safe distance, etc. The objective function takes the maximization of the safe driving distance and the balance of energy consumption as the optimization goal, and converts it into a linear form through certain mathematical transformations for subsequent calculations. A dynamic relaxation threshold adjustment mechanism is adopted to adaptively reduce the relaxation range of integer variables according to the number of iterations. At the beginning of the algorithm, a larger relaxation threshold is set. As the number of iterations increases, the threshold is gradually reduced so that the value range of the integer variable gradually approaches the optimal solution. For example, at the beginning of the iteration, the relaxation threshold may be set to 0.5, allowing the integer variable to take non-integer values ​​within a certain range; as the iteration proceeds, the threshold is gradually reduced to 0.1, making the value of the integer variable closer to the integer.

[0042] In the branching process, the fractional variable with the greatest impact on the objective function is preferentially selected for branching. By calculating the degree of influence of each fractional variable on the objective function, the variable with the greatest impact is selected for branching, so that the optimal solution can be found faster. In the pruning stage, a probabilistic pruning model is constructed based on the historical optimal solution, and the probabilistic pruning model is constructed based on Gaussian process regression. The objective function value and variable relaxation of the historical branch node are collected as training samples. The objective function value reflects the path planning effect under the branch node, and the variable relaxation indicates the degree of relaxation of the integer variable under the node. Construct a covariance kernel function, which includes a combination of linear kernel and Marten kernel. The linear kernel is used to capture the linear relationship between variables, and the Marten kernel can better handle nonlinear relationships. The hyperparameters are optimized by maximizing the marginal likelihood to predict the feasibility probability of the unexplored branch. When the feasibility probability is lower than the dynamic threshold, the early pruning operation is triggered. For example, the dynamic threshold can be adaptively adjusted according to historical data and the current search situation, and is generally set between 0.1-0.3, which can effectively reduce the amount of calculation and improve the efficiency of path planning.

[0043] The incremental solver is used to iteratively optimize the relaxation problem, and only a subset of local variables is updated in each iteration. The incremental solver uses the results of the previous iteration to update only some variables, avoiding repeated calculations and greatly improving the calculation efficiency. Through continuous iterations, the optimal cruise trajectory data is output until the integer feasibility condition is met, providing the optimal path planning for the vehicle.

[0044] Embodiment 4: This embodiment specifically describes the process of dynamically correcting the local trajectory using the rolling time domain optimization algorithm and implementing the ellipsoid set description through singular value decomposition.

[0045] When constructing a time-varying prediction model, the vehicle dynamics equation is discretized into a state transfer matrix. The vehicle dynamics equation describes the relationship between the vehicle's motion state and time, including longitudinal velocity, lateral deviation, and yaw angle differential terms. According to the discretized time step, the continuous dynamics equation is converted into a discrete state transfer matrix. For example, assuming the time step is Δt, the state transfer equation for the longitudinal velocity can be expressed as ,in Indicates The longitudinal velocity at time, Indicates The longitudinal acceleration at time, Indicates The longitudinal velocity at the moment . The lateral deviation and yaw angle differential terms also have similar discretized equations.

[0046] Design a sliding window optimization objective function, which includes a trajectory tracking error term, a control input smoothing term, and an obstacle repulsion potential energy term. The trajectory tracking error term is used to measure the difference between the current trajectory and the desired trajectory. For example, the Euclidean distance can be used to calculate the deviation between the current position of the vehicle and the corresponding position on the desired trajectory. The control input smoothing term ensures that the vehicle's control input (such as acceleration, steering angle, etc.) changes smoothly to avoid the impact of drastic changes on the vehicle's driving comfort and safety. The obstacle repulsion potential energy term generates a repulsive force based on the distance between the vehicle and the obstacle to prevent the vehicle from colliding with the obstacle. For example, the obstacle repulsion potential energy term can be expressed as ,in is a constant, It is the distance between the vehicle and the obstacle. The closer the distance, the greater the repulsive potential energy.

[0047] Robust feasibility constraints are introduced, and the uncertainty parameters are described by ellipsoid sets through semidefinite programming methods. The ellipsoid set description is achieved through singular value decomposition. The uncertainty parameters are modeled as bounded ellipsoid sets, the center vector is the nominal parameter, and the shape matrix is ​​determined by the parameter covariance. The shape matrix is ​​subjected to singular value decomposition to extract the direction and length of the main semi-axis. In the optimization constraints, it is mandatory that all the main semi-axis directions satisfy the linear inequality conditions. For example, assuming that the uncertainty parameters are the dynamic parameters of the vehicle (such as mass, moment of inertia, etc.), the main semi-axis direction and length of the ellipsoid set are obtained through singular value decomposition, and during the optimization process, it is ensured that the vehicle can still drive stably within the range of these uncertain parameters.

[0048] The alternating direction multiplier method is used to perform distributed solution of optimization problems, and real-time response is accelerated through parallel computing. The alternating direction multiplier method decomposes complex optimization problems into multiple sub-problems and solves them in parallel on different processors or computing units, greatly improving the computing efficiency. For example, the trajectory tracking error term, control input smoothing term, and obstacle repulsion potential energy term can be calculated on different computing units respectively, and then through information interaction and coordination, the optimal local trajectory correction result is finally obtained, so that the vehicle can dynamically adjust the driving trajectory according to real-time environmental changes.

[0049] Embodiment 5: At the execution layer of the vehicle adaptive cruise control, the longitudinal acceleration and lateral steering angle tracking of the vehicle are realized based on the robust sliding mode control algorithm. This process is crucial to ensure the stability and accuracy of the vehicle driving.

[0050] When designing the super-helical sliding surface, it is necessary to fully consider the coupling relationship between the longitudinal acceleration error and the lateral steering angle error. is the actual longitudinal acceleration of the vehicle and expected longitudinal acceleration The difference, that is , which reflects the degree of deviation of the vehicle from the expected speed change during acceleration or deceleration. is the actual lateral steering angle of the vehicle and the desired lateral steering angle The difference between , which reflects the deviation between the actual driving direction of the vehicle and the planned path direction. These two errors are coupled into a composite sliding mode variable , for example, can be expressed as ,in and It is a weight coefficient carefully set according to the vehicle dynamics. Dynamic factors such as vehicle mass, tire characteristics, and driving speed will affect the value of the weight coefficient. When driving at high speed, in order to ensure the stability of the vehicle, the The weight of the longitudinal acceleration error is more concerned about the impact of the longitudinal acceleration error on the overall state of the vehicle; in complex cornering scenarios, the The weight of the lateral steering angle error is adjusted to make the vehicle better along the curve.

[0051] An adaptive reaching law is constructed to suppress high-frequency chattering. The formula of the adaptive reaching law is: , where each parameter has a clear physical meaning and function. is the approach speed parameter, which determines the composite sliding mode variable The speed of approaching to zero. When the actual state of the vehicle deviates greatly from the expected state, the larger The value can make the system respond quickly and prompt the vehicle to adjust to the ideal driving state as soon as possible; as the deviation gradually decreases, the value can be appropriately reduced. value to avoid over-adjustment. is a value in The constant between and affects the convergence characteristics of the reaching law. When the value is small, the approach process is relatively gentle, but the convergence speed may be slow; Values ​​close to 1 will result in faster convergence but may cause the system to respond too drastically. is the switching gain, is a sign function. When the error amplitude is large, by increasing To speed up the approach; when the error amplitude is small, reduce High-frequency jitter can be effectively suppressed, because excessive switching gain will cause unnecessary high-frequency oscillations when the system is close to the ideal state, affecting the comfort and stability of the vehicle.

[0052] The use of disturbance observers to estimate unmodeled dynamic disturbances is a key step in improving the robustness of control systems. In actual driving, vehicles are affected by many factors that are not fully considered in the model, such as changes in road friction and wind interference. The disturbance observer uses a specific algorithm to estimate the unmodeled dynamic disturbances by real-time monitoring and analysis of vehicle inputs (such as control signals such as accelerator pedal position and steering wheel angle) and outputs (actual longitudinal acceleration, lateral steering angle and other measurement data). For example, the method based on the extended state observer regards the unknown disturbance and uncertainty in the system as an "extended state" and realizes real-time tracking of disturbances by observing and estimating the system state. Feedforward compensation is added to the control law. It can be expressed as ,in It is the control quantity calculated according to the ideal model. Through this feedforward compensation mechanism, the influence of unmodeled disturbance on vehicle driving can be effectively offset, so that the vehicle can still accurately track the expected longitudinal acceleration and lateral steering angle in a complex and changeable driving environment.

[0053] The finite-time convergence of the closed-loop system is verified by Lyapunov stability analysis. Constructing the Lyapunov function , which is about the composite sliding mode variable The quadratic function of , which directly reflects the energy state of the system. Taking the derivation, we can get . The adaptive reaching law Substitution In the expression of .because , , ,so , which shows that the Lyapunov function It decreases monotonically over time. According to Lyapunov's stability theory, when When it approaches zero, the composite sliding mode variable It also approaches zero, that is, the longitudinal acceleration error and the lateral steering angle error gradually decrease and eventually approach zero, which proves that the closed-loop system can converge to a stable state within a finite time, ensuring the driving stability and control accuracy of the vehicle.

[0054] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0055] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A neural network driven vehicle adaptive cruise control method, characterized in that: include: Collect real-time driving data of vehicles through multimodal sensors; Based on a multi-layer convolutional neural network, the real-time driving data is fused with temporal and spatial features to generate dynamic environment perception data; Inputting the dynamic environment perception data into a pre-trained adaptive decision model to generate driving strategy parameters; A mixed integer programming model is constructed according to the driving strategy parameters, wherein the mixed integer programming model takes maximizing the safe driving distance and balancing the energy consumption as optimization goals, and uses an incremental branch and bound algorithm to perform global planning on the cruising path, wherein the incremental branch and bound algorithm integrates a dynamic relaxation threshold and a heuristic pruning strategy; Outputting optimal cruise trajectory data based on the mixed integer programming model; A hierarchical control framework is constructed according to the optimal cruise trajectory data, and the hierarchical control framework includes a planning layer, a coordination layer and an execution layer, wherein the planning layer generates a global trajectory sequence based on the driving strategy parameters, the coordination layer uses a rolling time domain optimization algorithm to dynamically correct the local trajectory, and the execution layer realizes the longitudinal acceleration and lateral steering angle tracking of the vehicle based on a robust sliding mode control algorithm; the vehicle control command is output through the hierarchical control framework to complete the adaptive cruise control.

2. The vehicle adaptive cruise control method according to claim 1, characterized in that: The step of fusing the spatiotemporal features of the real-time driving data based on a multi-layer convolutional neural network to generate dynamic environment perception data includes: Multimodal sensors include LiDAR, millimeter wave radar, inertial measurement unit, visual camera, and ultrasonic sensor; Perform multi-scale voxel processing on the lidar point cloud data and visual camera images to generate a three-dimensional grid feature map; perform time-series alignment on the millimeter-wave radar data and ultrasonic sensor data to construct a time-series distance-velocity matrix; Constructing a dual-channel convolutional neural network, wherein the first channel uses a three-dimensional sparse convolution kernel to extract the geometric features of the three-dimensional grid feature map, and the second channel uses a temporal convolution kernel to extract the dynamic features of the temporal distance-velocity matrix; The geometric features and dynamic features are fused through a cross-channel feature splicing layer to generate a joint feature tensor; the time dependency of the joint feature tensor is modeled based on a gated recurrent unit to output dynamic environment perception data including the position, speed and motion trend of environmental obstacles.

3. The vehicle adaptive cruise control method according to claim 1, characterized in that: The adaptive decision model adopts a staged graph attention network structure to perform multi-objective optimization of driving strategies based on a dynamic priority allocation mechanism; the staged graph attention network structure includes: Construct a vehicle-environment interaction graph, where nodes include the vehicle node, surrounding vehicle nodes, lane line nodes, and obstacle nodes. Node attributes include position, velocity, and acceleration vectors. A two-stage attention mechanism is adopted. In the first stage, the association weights between the vehicle node and the surrounding nodes are calculated through the spatial attention layer. In the second stage, the importance of the historical state sequence is weighted through the temporal attention layer. Based on the multi-head graph attention module, the node features are iteratively updated, and each attention head uses a learnable position encoding to enhance the spatial relative relationship; the training process is stabilized by the residual connection and layer normalization mechanism, and finally the driving strategy parameters containing multi-objective constraints are output.

4. The vehicle adaptive cruise control method according to claim 1, characterized in that: The incremental branch-and-bound algorithm integrates a dynamic relaxation threshold and a heuristic pruning strategy including: The path planning problem is modeled as a mixed integer linear programming problem, where the decision variables include discrete lane selection variables and continuous acceleration variables. Initialize the relaxation problem and calculate the initial lower bound, and use a dynamic relaxation threshold adjustment mechanism to adaptively reduce the relaxation range of integer variables according to the number of iterations; In the branching process, the fractional variables that have the greatest impact on the objective function are prioritized for branching; in the pruning stage, a probabilistic pruning model is constructed based on the historical optimal solution, and Bayesian optimization is used to predict invalid branches and prune them in advance; The relaxed problem is iteratively optimized using an incremental solver, where only a subset of local variables is updated in each iteration until the integer feasibility condition is met.

5. The vehicle adaptive cruise control method according to claim 1, characterized in that: The rolling time domain optimization algorithm dynamically corrects the local trajectory including: A time-varying prediction model is constructed to discretize the vehicle dynamics equation into a state transfer matrix, wherein the state transfer matrix includes longitudinal velocity, lateral deviation, and yaw angle differential terms; Design a sliding window optimization objective function, including trajectory tracking error term, control input smoothing term and obstacle repulsion potential energy term; introduce robust feasibility constraints, describe the uncertainty parameters with an ellipsoid set through semidefinite programming method, and construct robust linear matrix inequality constraints; The alternating direction multiplier method is used to solve the optimization objective function in a distributed manner.

6. The vehicle adaptive cruise control method according to claim 1, characterized in that: The execution layer realizes the longitudinal acceleration and lateral steering angle tracking of the vehicle based on the robust sliding mode control algorithm, including: A super-helical sliding surface is designed to couple the longitudinal acceleration error and the lateral steering angle error into a composite sliding mode variable. An adaptive reaching law is constructed to dynamically adjust the approaching speed and switching gain according to the error amplitude to suppress high-frequency chattering.

7. The vehicle adaptive cruise control method according to claim 2, characterized in that: The three-dimensional sparse convolution kernel uses an octree structure to accelerate feature extraction, including: Divide the LiDAR point cloud into multi-level octree nodes, each of which stores the point cloud density and normal vector statistical features; Only non-empty nodes are activated in convolution operations, and multi-resolution geometric details are preserved through skip connections; A dynamic pooling layer is used to perform maximum aggregation on the features of adjacent nodes to generate a compact three-dimensional feature representation.

8. The vehicle adaptive cruise control method according to claim 3, characterized in that: The spatial attention layer adopts a relative position encoding mechanism, including: Define the relative position vector between the ego vehicle node and the target node, including the Euclidean distance, heading angle difference and relative velocity projection; Map the relative position vector to an attention bias term through a learnable nonlinear transformation; The bias term is added to the standard attention weight calculation process to enhance the geometric interpretability of spatial correlation.

9. The vehicle adaptive cruise control method according to claim 4, characterized in that: The probability pruning model is constructed based on Gaussian process regression and includes: Collect the objective function values ​​and variable relaxations of historical branch nodes as training samples; Construct the covariance kernel function, which includes a combination of linear kernel and Marten kernel; By maximizing the marginal likelihood, we optimize the hyperparameters and predict the feasibility probability of unexplored branches. When the feasibility probability is lower than the dynamic threshold, early pruning is triggered.

10. The vehicle adaptive cruise control method according to claim 5, characterized in that: The ellipsoid set description is implemented by singular value decomposition, including: The uncertainty parameters are modeled as a set of bounded ellipsoids, with the center vectors as nominal parameters and the shape matrix determined by the parameter covariance; Perform singular value decomposition on the shape matrix to extract the direction and length of the main semi-axis; In the optimization constraints, all major semi-axis directions are required to satisfy linear inequality conditions.

Images