Neural Network-Driven Vehicle Adaptive Cruise Control Method

Abstract

The present 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 is collected through multi-modal sensors, and a multi-layer convolutional neural network is used for spatio-temporal feature fusion to generate dynamic environment perception data, which is input into a pre-trained adaptive decision-making model to obtain driving strategy parameters. Based on this, a mixed integer programming model is constructed, and an incremental branch and bound algorithm integrating a dynamic relaxation threshold and a heuristic pruning strategy is used to globally plan the cruise path, and the optimal cruise trajectory data is output. Then, a hierarchical control framework including 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 the local trajectory, and the execution layer realizes the tracking of the vehicle longitudinal acceleration and the lateral steering angle based on the robust sliding mode control algorithm, and outputs vehicle control instructions to complete the adaptive cruise control, improving the performance, safety and stability of the vehicle cruise control in complex environments.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle cruise control, and particularly to a neural network-driven vehicle adaptive cruise control method. Background Technique

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

[0003] In terms of sensor data processing, traditional methods often rely on only one or a few types of sensors, such as only using millimeter-wave radar to monitor the vehicle distance and relative speed. The environmental information obtained in this way is limited, and it is difficult to comprehensively perceive the complex situations around the vehicle. Moreover, the traditional method of processing sensor data is relatively simple, and the spatio-temporal features in the data are not fully exploited. For example, the point cloud data collected by lidar and the image data obtained by visual cameras contain rich environmental geometry and texture information, but traditional methods fail to effectively fuse these multi-modal data, unable to provide accurate environmental perception for the vehicle, and restricting 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, lacking the ability to flexibly respond to complex environments and multi-objective optimization. Their path planning algorithms also often use relatively simple search algorithms, with high computational complexity and low efficiency, and cannot quickly plan a reasonable cruise trajectory in the vehicle driving scenario with extremely high real-time requirements. Especially in the congested urban roads, traditional path planning algorithms are prone to falling into local optimal solutions, resulting in unreasonable vehicle driving routes, increasing the driving time and energy consumption.

[0005] In terms of vehicle control execution, traditional control algorithms rely strongly on the vehicle dynamics model, while actual vehicles are affected by various uncertain factors during driving, such as changes in road surface friction and wear of vehicle components, which makes it difficult for traditional control algorithms to ensure the stable driving of vehicles under different working conditions. Summary of the Invention

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

[0007] To achieve the above purpose, the present invention provides the following technical solution: A neural network-driven vehicle adaptive cruise control method, the method includes:

[0008] Collect real-time driving data of the vehicle through multi-modal sensors;

[0009] Perform spatio-temporal feature fusion on the real-time driving data based on a multi-layer convolutional neural network to generate dynamic environment perception data;

[0010] Input the dynamic environment perception data into a pre-trained adaptive decision-making model to generate driving strategy parameters;

[0011] Construct a mixed-integer programming model according to the driving strategy parameters. The mixed-integer programming model takes maximizing the driving safety distance and balancing energy consumption as optimization objectives, and uses an incremental branch and bound algorithm to globally plan the cruise path, where the incremental branch and bound algorithm integrates a dynamic relaxation threshold and a heuristic pruning strategy; output the optimal cruise trajectory data based on the mixed-integer programming model;

[0012] Construct a hierarchical control framework according to 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 dynamically corrects the local trajectory using a receding horizon optimization algorithm. The execution layer realizes the tracking of the vehicle's longitudinal acceleration and lateral steering angle based on a robust sliding mode control algorithm; output vehicle control instructions through the hierarchical control framework to complete adaptive cruise control.

[0013] Preferably, the performing spatio-temporal feature fusion on the real-time driving data based on a multi-layer convolutional neural network to generate dynamic environment perception data includes:

[0014] The multi-modal sensors include lidar, millimeter wave radar, inertial measurement unit, vision camera, and ultrasonic sensor;

[0015] Perform multi-scale voxelization processing on the lidar point cloud data and the vision camera image to generate a three-dimensional grid feature map; perform temporal alignment on the millimeter wave radar data and the ultrasonic sensor data to construct a temporal distance-velocity matrix;

[0016] Construct a two-channel convolutional neural network. The first channel uses a three-dimensional sparse convolutional kernel to extract the geometric features of the three-dimensional grid feature map, and the second channel uses a temporal convolutional kernel to extract the dynamic features of the temporal distance-velocity matrix;

[0017] Fuse the geometric features and the dynamic features through a cross-channel feature splicing layer to generate a joint feature tensor; perform time-dependence modeling on the joint feature tensor based on a gated recurrent unit, and output dynamic environment perception data including the positions, velocities, and motion trends of environmental obstacles.

[0018] Preferably, the adaptive decision-making model adopts a phased graph attention network structure to perform multi-objective optimization on driving strategies based on a dynamic priority allocation mechanism; the phased graph attention network structure includes:

[0019] Construct a vehicle-environment interaction graph, where the nodes in the graph include the ego vehicle node, surrounding vehicle nodes, lane line nodes, and obstacle nodes, and the node attributes include position, speed, and acceleration vectors;

[0020] Adopt a two-stage attention mechanism. In the first stage, calculate the association weights between the ego vehicle node and surrounding nodes through a spatial attention layer. In the second stage, perform importance weighting on the historical state sequence through a temporal attention layer;

[0021] Based on the multi-head graph attention module, iteratively update the node features. Each attention head uses learnable position encoding to enhance the spatial relative relationship; stabilize the training process through residual connection and layer normalization mechanisms, and finally output the driving strategy parameters containing multi-objective constraints.

[0022] Preferably, the incremental branch and bound algorithm integrating a dynamic relaxation threshold and a heuristic pruning strategy includes:

[0023] Model the path planning problem as a mixed-integer linear programming problem, and the decision variables include discrete lane selection variables and continuous acceleration variables;

[0024] Initialize the relaxation problem and calculate the initial lower bound. Adopt a dynamic relaxation threshold adjustment mechanism to adaptively narrow the relaxation range of integer variables according to the number of iterations;

[0025] During the branching process, preferentially select the fractional variable that has the greatest impact on the objective function for branching; in the pruning stage, construct a probability pruning model based on the historical optimal solution, and use Bayesian optimization to predict and prune invalid branches in advance;

[0026] Use an incremental solver to iteratively optimize the relaxation problem, and only update a subset of local variables in each iteration until the integer feasibility condition is met.

[0027] Preferably, the rolling horizon optimization algorithm dynamically corrects the local trajectory including:

[0028] Construct a time-varying prediction model, discretize the vehicle dynamics equation into a state transition matrix, and the state transition matrix includes longitudinal speed, lateral deviation, and yaw rate differential terms;

[0029] Design a sliding window optimization objective function, including a trajectory tracking error term, a control input smoothing term, and an obstacle repulsion potential energy term;

[0030] Introduce robust feasibility constraints, describe the uncertain parameters by the ellipsoid set through the semi - definite programming method, and construct robust linear matrix inequality constraints;

[0031] Use the alternating direction multiplier method to solve the optimization problem distributively and accelerate the real - time response through parallel computing.

[0032] Preferably, the execution layer realizes the tracking of the longitudinal acceleration and the lateral steering angle of the vehicle based on the robust sliding - mode control algorithm, including:

[0033] Design a super - twisting sliding - mode surface, and couple the longitudinal acceleration error and the lateral steering angle error into a composite sliding - mode variable;

[0034] Construct an adaptive reaching law, dynamically adjust the reaching speed and the switching gain according to the error amplitude, and suppress the high - frequency chattering phenomenon;

[0035] Use an interference observer to estimate the unmodeled dynamic disturbances and feed - forward compensate the estimated value to the control law;

[0036] Verify the finite - time convergence of the closed - loop system through Lyapunov stability analysis.

[0037] Preferably, the three - dimensional sparse convolution kernel uses an octree structure to accelerate feature extraction, including:

[0038] Divide the lidar point cloud into multi - level octree nodes, and each node stores the point cloud density and the statistical features of the normal vector;

[0039] Only activate the non - empty nodes in the convolution operation, and retain the multi - resolution geometric details through skip connections;

[0040] Use a dynamic pooling layer to perform maximum aggregation on the features of adjacent nodes to generate a compact three - dimensional feature representation.

[0041] Preferably, the spatial attention layer uses a relative - position encoding mechanism, including:

[0042] Define the relative - position vector between the ego - vehicle node and the target node, which includes the Euclidean distance, the heading - angle difference, and the relative - velocity projection;

[0043] Map the relative - position vector to an attention bias term through a learnable non - linear transformation;

[0044] Superimpose the bias term on the standard attention - weight calculation process to enhance the geometric interpretability of the spatial correlation.

[0045] Preferably, the probability pruning model is constructed based on Gaussian process regression, including:

[0046] Collect the objective - function values and variable relaxation degrees of historical branch nodes as training samples;

[0047] Construct a covariance kernel function, including a combined form of a linear kernel and a Matérn kernel;

[0048] Optimize the hyperparameters by maximizing the marginal likelihood to predict the feasibility probability of unexplored branches;

[0049] When the feasibility probability is lower than the dynamic threshold, trigger the early pruning operation.

[0050] Preferably, the description of the ellipsoidal set is achieved through singular value decomposition, including:

[0051] Model the uncertainty parameters as a bounded ellipsoidal set, with the central vector being the nominal parameter and the shape matrix determined by the parameter covariance;

[0052] Perform singular value decomposition on the shape matrix to extract the principal axis directions and lengths;

[0053] Enforce that all principal axis directions satisfy the linear inequality conditions in the optimization constraints.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] At the environmental perception level, multi-modal sensors are used to collect data, and spatio-temporal feature fusion is performed through a multi-layer convolutional neural network. Multi-modal sensors such as lidar, millimeter-wave radar, inertial measurement unit, vision camera, and ultrasonic sensor work together to obtain vehicle driving environment information from different dimensions. The lidar provides high-precision distance information, the vision camera captures rich image textures and semantics, the millimeter-wave radar monitors the relative speed in real time, the ultrasonic sensor assists in close-range detection, and the inertial measurement unit senses the vehicle attitude. The multi-layer convolutional neural network effectively fuses these multi-source heterogeneous data to generate accurate dynamic environment perception data. By performing multi-scale voxelization on lidar point cloud data and vision camera images, and performing temporal alignment on 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 motion trend of environmental obstacles. This greatly improves the vehicle's perception ability of complex environments. Compared with the traditional method that relies on a single sensor, it can comprehensively and accurately grasp the surrounding environment information, providing a solid foundation for subsequent decision-making.

[0056] In terms of decision-making and planning, the adaptive decision-making 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 ego vehicle, surrounding vehicles, lane lines, and obstacles. The node attributes are rich and can accurately reflect the states of various objects. The two-stage attention mechanism includes a spatial attention layer that calculates the association weights through relative position encoding to enhance the geometric interpretability of spatial correlation, enabling the vehicle to pay more reasonable attention to surrounding objects; and a temporal attention layer that weights the historical state sequence to make full use of historical information. The multi-head graph attention module iteratively updates the node features, combines residual connections and layer normalization to stabilize the training, and outputs the driving strategy parameters containing multi-objective constraints. This model can comprehensively consider multiple objectives such as driving safety, energy consumption, and traffic flow in complex traffic scenarios and optimize the driving strategy. On highways, it can dynamically adjust the vehicle speed and following distance according to traffic flow and road conditions, ensuring safety while reducing energy consumption; in urban roads, it can quickly respond to changes in pedestrians and traffic lights, make reasonable decisions, and improve traffic efficiency.

[0057] When performing path planning, a mixed-integer programming model is constructed based on the driving strategy parameters, and an incremental branch-and-bound algorithm integrating a dynamic relaxation threshold and a 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 to prune invalid branches in advance and reduce the computational amount. The incremental solver iteratively optimizes, only updating a subset of local variables each time. While ensuring the planning accuracy, it quickly outputs the optimal cruise trajectory data, making the vehicle's driving path more reasonable and efficient.

[0058] 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 macroscopic guidance for vehicle driving; the coordination layer adopts a receding horizon optimization algorithm, constructs a time-varying prediction model, designs an optimization objective function containing multiple constraint terms, and introduces a robust feasibility constraint. It is solved distributively through semidefinite programming and the alternating direction method of multipliers to dynamically correct the local trajectory and respond to road condition changes in real time; the execution layer is based on the robust sliding mode control algorithm, designs a super-twisting sliding mode surface to couple the longitudinal acceleration and the lateral steering angle error, constructs an adaptive reaching law to suppress chattering, and uses a disturbance observer to compensate for unmodeled disturbances. Through Lyapunov stability analysis, it ensures the finite-time convergence of the closed-loop system. This enables the vehicle to accurately track the desired longitudinal acceleration and lateral steering angle under different driving conditions, improving driving safety and comfort, and effectively coping with uncertain factors such as road surface friction changes and external disturbances to ensure stable vehicle driving. Description of the Drawings

[0059] Figure 1It is the working principle diagram of the vehicle adaptive cruise control method described in the present invention;

[0060] Figure 2 It is the flow chart of multi-modal data processing and dynamic environment perception data generation;

[0061] Figure 3 It is the working flow chart of the adaptive decision-making model;

[0062] Figure 4 It is the working flow chart of the probability pruning model based on Gaussian process regression. Specific embodiments

[0063] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0064] Please refer to Figures 1 - 4 , the present invention provides a technical solution: a neural network-driven vehicle adaptive cruise control method, and the specific implementation steps are as follows:

[0065] Use multi-modal sensors to collect real-time driving data of the vehicle. These multi-modal sensors include lidar, millimeter-wave radar, inertial measurement unit, vision camera, and ultrasonic sensor, etc. They obtain various information during the vehicle driving process from different angles and ways, such as the distance, speed, image information of the surrounding environment, and the attitude data of the vehicle itself, etc., providing a basis for subsequent analysis and decision-making.

[0066] Based on a multi-layer convolutional neural network, perform spatio-temporal feature fusion on the collected real-time driving data. The data collected by multi-modal sensors have different characteristics and need to be fused through specific processing methods. Perform multi-scale voxelization processing on lidar point cloud data and vision camera images to generate a three-dimensional grid feature map; perform time series alignment on millimeter-wave radar data and ultrasonic sensor data to construct a time series distance-velocity matrix. Then construct a two-channel convolutional neural network. The first channel uses a three-dimensional sparse convolutional kernel to extract the geometric features of the three-dimensional grid feature map, and the second channel uses a time series convolutional kernel to extract the dynamic features of the time series distance-velocity matrix. Fusion the geometric features and dynamic features through a cross-channel feature splicing layer to generate a joint feature tensor, and then perform time-dependent modeling on the joint feature tensor based on a gated recurrent unit, and finally output dynamic environment perception data including the position, speed, and motion trend of environmental obstacles, enabling the vehicle to have a more accurate perception of the surrounding dynamic environment.

[0067] The generated dynamic environment perception data is input into a pre-trained adaptive decision-making model, which adopts a phased graph attention network structure and conducts 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 ego vehicle node, surrounding vehicle nodes, lane line nodes, and obstacle nodes, and the node attributes include position, speed, and acceleration vectors. A two-stage attention mechanism is adopted. In the first stage, the association weights between the ego vehicle node and surrounding nodes are calculated through a spatial attention layer. In the second stage, the importance of the historical state sequence is weighted through a temporal attention layer. The node features are iteratively updated based on a multi-head graph attention module. Each attention head uses learnable position encoding to enhance the spatial relative relationship, and then the training process is stabilized through a residual connection and layer normalization mechanism. Finally, the driving strategy parameters containing multi-objective constraints are output, providing a basis for the driving decision-making of the vehicle.

[0068] A mixed-integer programming model is constructed according to the driving strategy parameters, and the model aims to maximize the driving safety distance and balance the energy consumption. An incremental branch and bound algorithm is used to globally plan the cruise path, and the incremental branch and bound algorithm integrates a dynamic relaxation threshold and a heuristic pruning strategy. In the algorithm implementation, the path planning problem is modeled as a mixed-integer linear programming problem, and the decision variables include discrete lane selection variables and continuous acceleration variables. The relaxed problem is initialized and the initial lower bound is calculated. A dynamic relaxation threshold adjustment mechanism is adopted to adaptively narrow the relaxation range of integer variables according to the number of iterations. During the branching process, the fractional variable that has the greatest impact on the objective function is preferentially selected 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 and prune invalid branches in advance. An incremental solver is used to iteratively optimize the relaxed problem, and 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-efficient.

[0069] Construct a hierarchical control framework 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 driving strategy parameters. The coordination layer dynamically corrects the local trajectory using a rolling horizon optimization algorithm. The execution layer realizes the tracking of the vehicle's longitudinal acceleration and lateral steering angle based on a robust sliding mode control algorithm. When constructing the rolling horizon optimization algorithm, a time-varying prediction model is built, the vehicle dynamics equation is discretized into a state transition matrix, a sliding window optimization objective function is designed, including a trajectory tracking error term, a control input smoothing term, and an obstacle repulsion potential energy term. A robust feasibility constraint is introduced, the uncertain parameters are described by an ellipsoidal set through a semidefinite programming method, a robust linear matrix inequality constraint is constructed, and the alternating direction multiplier method is used to solve the optimization problem distributively. The execution layer designs a super-twisting sliding surface, couples the longitudinal acceleration error and the lateral steering angle error into a composite sliding mode variable, constructs an adaptive reaching law, dynamically adjusts the reaching speed and switching gain according to the error amplitude, suppresses the high-frequency chattering phenomenon, uses a disturbance observer to estimate the unmodeled dynamic disturbance, and feeds the estimated value forward to compensate the control law. The finite-time convergence of the closed-loop system is verified through Lyapunov stability analysis. Vehicle control commands are output through the hierarchical control framework to complete adaptive cruise control, enabling the vehicle to drive safely and stably along the planned trajectory.

[0070] The present invention will be further described below in conjunction with Embodiments 1 to 5:

[0071] Embodiment 1: In this embodiment, the specific process of generating dynamic environment perception data based on multi-layer convolutional neural network for spatio-temporal feature fusion and accelerating feature extraction using an octree structure for a three-dimensional sparse convolutional kernel is elaborated in detail.

[0072] After the multi-modal sensors collect data, the processing of lidar point cloud data and visual camera images is a key step. When performing multi-scale voxelization processing, first determine the size and resolution of the voxels, and select appropriate scale parameters according to the complexity of the vehicle driving environment and the limitation of computing resources. For example, in an urban road environment, due to more obstacles and frequent changes, a smaller voxel size is selected to capture details more precisely; while in a highway environment, the voxel size can be appropriately increased to reduce the computational amount. Divide the lidar point cloud data according to the set voxel size, statistically analyze and encode the point cloud information in each voxel to generate a three-dimensional grid feature map, which contains the geometric structure information of the environment. At the same time, perform voxelization processing on the visual camera image, correspond the image pixels to the voxels, and integrate the color, texture and other information in the image into the three-dimensional grid feature map.

[0073] For millimeter-wave radar data and ultrasonic sensor data, since they collect information about distance and speed and there are differences in the acquisition time, it is necessary to perform time series alignment. Through the time synchronization algorithm, the data collected at different times is adjusted to the same time scale to construct a time series distance-velocity matrix. This matrix records the distance and speed changes of the objects around the vehicle at different time points, providing data support for the subsequent extraction of dynamic features.

[0074] When constructing a two-channel convolutional neural network, the first channel uses a three-dimensional sparse convolutional kernel. The three-dimensional sparse convolutional kernel adopts an octree structure to accelerate feature extraction. The specific operations are as follows: The lidar point cloud is divided into multi-level octree nodes, and each node stores the point cloud density and the statistical features of the normal vector. In the convolution operation, only the non-empty nodes are activated, which greatly reduces the computational amount. At the same time, the multi-resolution geometric details are retained through skip connections, enabling effective extraction of geometric features at different scales. A dynamic pooling layer is used to perform maximum aggregation on 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 dynamic pooling layer selects the most representative feature among them to form a more concise and effective feature representation. The second channel uses a time series convolutional kernel to perform convolution operations on the time series distance-velocity matrix to extract its dynamic features, such as the motion trend of the object, the rate of change of speed, etc.

[0075] Through the cross-channel feature splicing 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. Based on the gated recurrent unit, time dependence modeling is performed on it. The gated recurrent unit can adaptively adjust the transmission and update of information according to the current input and the state at the previous moment, so as to output dynamic environment perception data including the position, speed, and motion trend of environmental obstacles. Through such processing, the vehicle can more comprehensively and accurately understand the dynamic changes in the surrounding environment, providing a reliable basis for subsequent decision-making and control.

[0076] Embodiment 2: This embodiment focuses on explaining the stage-based graph attention network structure adopted by the adaptive decision-making model and the relative position encoding mechanism adopted by the spatial attention layer.

[0077] 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 itself, and its attributes include the vehicle's position, speed, and acceleration vector, which can be obtained through the vehicle's own sensors and control systems. The surrounding vehicle nodes represent other vehicles around the vehicle and also include position, speed, and acceleration vectors, which are collected by multi-modal sensors and processed through data fusion. The lane line node records information such as the position and curvature of the lane line, and the obstacle node represents the obstacles on the road and their related attributes. The connection relationships between the nodes are determined based on the spatial positions and interactions of the vehicle and the surrounding objects. For example, there are relative position and speed relationships between the ego vehicle and the surrounding vehicles, which are reflected by the connecting edges.

[0078] Adopt a two-stage attention mechanism. In the spatial attention layer of the first stage, a relative position encoding mechanism is used. Define the relative position vector between the ego vehicle node and the target node, which includes 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 their driving directions, and the relative speed projection represents the speed component in the relative direction of the two. The relative position vector is mapped to an attention bias term through a learnable non-linear transformation, which can be implemented using neural network structures such as multi-layer perceptrons. The bias term is superimposed on the standard attention weight calculation process to enhance the geometric interpretability of the spatial correlation. For example, when calculating the attention weights between the ego vehicle node and the surrounding vehicle nodes, the bias term obtained by the non-linear transformation of the relative position vector will make the attention weights more inclined to the vehicles with shorter distances, similar driving directions, and smaller relative speeds, thus more accurately reflecting the spatial relationships and mutual influences between vehicles.

[0079] In the temporal attention layer of the second stage, importance weighting is performed on the historical state sequence. By recording the state information of the vehicle and the surrounding environment at multiple time points, a historical state sequence is formed. The temporal attention layer weights the information at different moments in the historical state sequence according to the current environmental changes and decision-making requirements, so that the recent information relevant to the current decision-making gets a greater weight, thus better utilizing historical information for decision-making.

[0080] Iteratively update the node features based on the multi-head graph attention module. Each attention head uses learnable positional encoding to enhance the spatial relative relationship. The multi-head graph attention module can simultaneously focus on and update the node features from different perspectives, improving the model's expressive ability. The learnable positional encoding for each attention head is achieved through a specific encoding method, such as a combination encoding of sine and cosine functions. The training process is stabilized through residual connections and layer normalization mechanisms. Residual connections make it easier for the network to converge during training, avoiding the problem of gradient vanishing. The layer normalization mechanism normalizes the input of each layer, making the training of the network more stable. Finally, the adaptive decision-making model outputs driving strategy parameters containing multi-objective constraints. These parameters comprehensively consider multiple objectives such as vehicle driving safety, comfort, and efficiency, providing a reasonable basis for vehicle driving decisions.

[0081] Embodiment 3: This embodiment details the process of integrating the incremental branch and bound algorithm with a dynamic relaxation threshold, a heuristic pruning strategy, and the construction of a probabilistic pruning model based on Gaussian process regression.

[0082] When modeling the path planning problem as a mixed-integer linear programming problem, the selection and definition of decision variables are crucial. Discrete lane selection variables are used to represent the lanes selected by the vehicle during driving. For example, integer values such as 0, 1, 2, etc. are used to represent different lane numbers, and the lane selection of the vehicle at a certain moment can be described by this variable. Continuous acceleration variables are 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.

[0083] Initialize the relaxation problem and calculate the initial lower bound. First, determine the constraint conditions and objective function of the relaxation problem. The constraint conditions include vehicle dynamic limitations, road rule limitations, etc., such as the maximum acceleration of the vehicle, the minimum safety distance, etc. The objective function aims to maximize the driving safety distance and balance the energy consumption, and it is transformed into a linear form through certain mathematical transformations for subsequent calculations. Adopt a dynamic relaxation threshold adjustment mechanism to adaptively narrow the relaxation range of integer variables according to the number of iterations. At the beginning of the algorithm, set a relatively large relaxation threshold. As the number of iterations increases, gradually reduce the threshold, so that the value range of integer variables gradually approaches the optimal solution. For example, at the initial stage of 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 progresses, the threshold gradually decreases to 0.1, making the value of the integer variable closer to an integer.

[0084] During the branching process, the fractional variable that has 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 and choosing the variable with the greatest influence for branching, the optimal solution can be found more quickly. 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 values and variable relaxations of historical branching nodes are collected as training samples. The objective function values reflect the path planning effect under this branching node, and the variable relaxation represents the relaxation degree of integer variables under this node. A covariance kernel function is constructed, which includes a combined form of a linear kernel and a Matérn kernel. The linear kernel is used to capture the linear relationship between variables, and the Matérn kernel can better handle non-linear relationships. The hyperparameters are optimized by maximizing the marginal likelihood to predict the feasibility probability of unexplored branches. When the feasibility probability is lower than the dynamic threshold, an 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 and 0.3. This can effectively reduce the computational amount and improve the efficiency of path planning.

[0085] An 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 utilizes the results of the previous iteration and only updates some variables, avoiding repeated calculations and greatly improving the computational efficiency. Through continuous iteration until the integer feasibility condition is met, the optimal cruise trajectory data is output to provide the optimal path planning for the vehicle's travel.

[0086] Example 4: This example specifically illustrates the process of dynamically correcting the local trajectory by the rolling horizon optimization algorithm and the realization of the ellipsoidal set description through singular value decomposition.

[0087] When constructing the time-varying prediction model, the vehicle dynamics equation is discretized into a state transition matrix. The vehicle dynamics equation describes the relationship between the vehicle's motion state and time, including the longitudinal velocity, lateral deviation, and yaw rate differential terms, etc. According to the discretized time step, the continuous dynamics equation is transformed into a discrete state transition matrix. For example, assuming the time step is Δt, the state transition equation of the longitudinal velocity can be expressed as , where represents the longitudinal velocity at the th moment, represents the longitudinal acceleration at the th moment, represents the longitudinal velocity at the th moment. Similar discretization equations exist for the lateral deviation and yaw rate differential terms.

[0088] Design a sliding window to optimize the 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 control inputs of the vehicle (such as acceleration, steering angle, etc.) change smoothly, avoiding the impact of drastic changes on the driving comfort and safety of the vehicle. 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 , where is a constant,[[]] is the distance between the vehicle and the obstacle, and the closer the distance, the greater the repulsive potential energy.

[0089] Introduce robust feasibility constraints. The uncertain parameters are described by an ellipsoidal set through semi-definite programming methods. The ellipsoidal set description is achieved through singular value decomposition. The uncertain parameters are modeled as a bounded ellipsoidal set, with the center vector being the nominal parameter and the shape matrix determined by the parameter covariance. Perform singular value decomposition on the shape matrix to extract the principal axis directions and lengths. In the optimization constraints, it is required that all principal axis directions satisfy linear inequality conditions. For example, assume that the uncertain parameters are the dynamic parameters of the vehicle (such as mass, moment of inertia, etc.). By singular value decomposition, the principal axis directions and lengths of the ellipsoidal set are obtained, and during the optimization process, it is ensured that the vehicle can still drive stably within the range of changes of these uncertain parameters.

[0090] Use the alternating direction method of multipliers to solve the optimization problem distributively, and accelerate the real-time response through parallel computing. The alternating direction method of multipliers decomposes the complex optimization problem into multiple sub-problems, which are solved in parallel on different processors or computing units, greatly improving the computational efficiency. For example, the trajectory tracking error term, the control input smoothing term, and the 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, enabling the vehicle to dynamically adjust the driving trajectory according to the real-time environmental changes.

[0091] Example 5: In the execution layer of 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, which is crucial for ensuring the stability and accuracy of vehicle driving.

[0092] When designing the super-twisting sliding mode surface, the coupling relationship between the longitudinal acceleration error and the lateral steering angle error needs to be fully considered. The longitudinal acceleration error is the difference between the actual longitudinal acceleration of the vehicle and the desired longitudinal acceleration , that is , which reflects the degree of deviation of the vehicle from the expected speed change during acceleration or deceleration. Lateral steering angle error is the actual lateral steering angle of the vehicle and the desired lateral steering angle The difference, that is , reflecting the deviation of the actual driving direction of the vehicle from the planned path direction. Coupling these two errors into a composite sliding mode variable , for example, can be expressed as , where and are weight coefficients carefully set according to the vehicle dynamics characteristics. Vehicle dynamics factors such as vehicle mass, tire characteristics, and driving speed will all affect the value of the weight coefficient. When driving at high speed, to ensure the driving stability of the vehicle, the weight of may be increased, and more attention is paid to the impact of the longitudinal acceleration error on the overall state of the vehicle; while in the complex curve driving scenario, the weight of will be appropriately increased, focusing on adjusting the lateral steering angle error to enable the vehicle to better follow the curve.

[0093] Construct an adaptive reaching law to suppress the high-frequency chattering phenomenon. The formula of the adaptive reaching law is , where each parameter has clear physical meanings and functions. is the reaching speed parameter, which determines the speed at which the composite sliding mode variable approaches zero. When the deviation between the actual state and the desired state of the vehicle is large, a larger value can make the system respond quickly, prompting the vehicle to adjust to a state close to the ideal driving state as soon as possible; as the deviation gradually decreases, the value of can be appropriately reduced to avoid over-adjustment. is a constant with a value between , and its magnitude affects the convergence characteristics of the reaching law. When the value is small, the reaching process is relatively gentle, but the convergence speed may be slow; When the value approaches 1, the convergence speed increases, but it may cause the system response to be too violent. is the switching gain, is the sign function. When the error amplitude is large, increasing to accelerate the reaching speed; when the error amplitude is small, reducing can effectively suppress the high-frequency chattering phenomenon, because too large a switching gain will cause unnecessary high-frequency oscillations when the system approaches the ideal state, affecting the comfort and stability of the vehicle driving.

[0094] An interference observer is used to estimate the unmodeled dynamic disturbances, which is a key link to improve the robustness of the control system. During actual vehicle driving, the vehicle is affected by various factors not fully considered in the model, such as changes in road surface friction and wind interference. The interference observer monitors and analyzes the vehicle's inputs (such as control signals like throttle pedal position and steering wheel angle) and outputs (measured data such as actual longitudinal acceleration and lateral steering angle) in real time, and uses specific algorithms to estimate the unmodeled dynamic disturbances. . For example, the method based on the extended state observer regards the unknown disturbances and uncertainties in the system as an "extended state", and realizes the real-time tracking of the disturbances through the observation and estimation of the system state. The estimated disturbance value is fed forward and compensated into the control law, and the control law can be expressed as , where is the control quantity calculated according to the ideal model. Through this feed-forward compensation mechanism, the influence of unmodeled disturbances on vehicle driving can be effectively offset, enabling the vehicle to accurately track the desired longitudinal acceleration and lateral steering angle in a complex and changing driving environment.

[0095] The finite-time convergence of the closed-loop system is verified through Lyapunov stability analysis. A Lyapunov function is constructed, which is a quadratic function of the composite sliding mode variable , and intuitively reflects the energy state of the system. Taking the derivative of , we can obtain . Substituting the adaptive reaching law into the expression of , we get . Since , , , so , which indicates that the Lyapunov function decreases monotonically with time. According to the Lyapunov stability theory, when approaches zero, the composite sliding mode variable also approaches zero, that is, the longitudinal acceleration error and the lateral steering angle error gradually decrease and finally approach zero, thus proving 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.

[0096] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, article or device.

[0097] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and 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 Including: Collecting real-time driving data of the vehicle through multi-modal sensors; Performing spatio-temporal feature fusion on the real-time driving data based on a multi-layer convolutional neural network to generate dynamic environment perception data; Inputting the dynamic environment perception data into a pre-trained adaptive decision-making model to generate driving strategy parameters; Constructing a mixed-integer programming model according to the driving strategy parameters, where the mixed-integer programming model takes maximizing the driving safety distance and balancing energy consumption as optimization objectives, and uses an incremental branch and bound algorithm to globally plan the cruise path, and 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; Constructing a hierarchical control framework according to the optimal cruise trajectory data, where 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 dynamically corrects the local trajectory by using a receding horizon optimization algorithm, and the execution layer realizes the tracking of the longitudinal acceleration and lateral steering angle of the vehicle based on a robust sliding mode control algorithm; Outputting vehicle control instructions through the hierarchical control framework to complete adaptive cruise control.

2. The vehicle adaptive cruise control method according to claim 1, wherein The performing spatio-temporal feature fusion on the real-time driving data based on a multi-layer convolutional neural network to generate dynamic environment perception data includes: The multi-modal sensors include lidar, millimeter-wave radar, inertial measurement unit, vision camera and ultrasonic sensor; Performing multi-scale voxelization processing on lidar point cloud data and vision camera images to generate a three-dimensional grid feature map; Aligning the millimeter-wave radar data and ultrasonic sensor data in time series to construct a time series distance-velocity matrix; Constructing a two-channel convolutional neural network, where 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-velocity matrix; Fusing the geometric features and dynamic features through a cross-channel feature splicing layer to generate a joint feature tensor; Modeling the time dependence of the joint feature tensor based on a gated recurrent unit, and outputting dynamic environment perception data including the positions, speeds and motion trends of environmental obstacles.

3. The vehicle adaptive cruise control method according to claim 1, wherein, The adaptive decision-making model adopts a multi-stage graph attention network structure to perform multi-objective optimization on the driving strategy based on a dynamic priority allocation mechanism; The multi-stage graph attention network structure includes: Constructing a vehicle-environment interaction graph, where the nodes in the graph include a self-vehicle node, surrounding vehicle nodes, lane line nodes and obstacle nodes, and the node attributes include position, speed and acceleration vectors; Adopting a two-stage attention mechanism, where the first stage calculates the association weights between the self-vehicle node and surrounding nodes through a spatial attention layer, and the second stage performs importance weighting on the historical state sequence through a time attention layer; Iteratively updating the node features based on a multi-head graph attention module, and each attention head uses a learnable position encoding to enhance the spatial relative relationship; Stabilizing the training process through a residual connection and layer normalization mechanism, and finally outputting driving strategy parameters including multi-objective constraints.

4. The vehicle adaptive cruise control method according to claim 1, wherein The incremental branch and bound algorithm integrating dynamic relaxation threshold and heuristic pruning strategy includes: Model the path planning problem as a mixed-integer linear programming problem, where the decision variables include discrete lane selection variables and continuous acceleration variables; Initialize the relaxed problem and calculate the initial lower bound. Adopt a dynamic relaxation threshold adjustment mechanism to adaptively narrow the relaxation range of integer variables according to the number of iterations; During the branching process, preferentially select the fractional variable that has the greatest impact on the objective function for branching; in the pruning stage, construct a probabilistic pruning model based on the historical optimal solution, and use Bayesian optimization to predict invalid branches and prune them in advance; Use an incremental solver to iteratively optimize the relaxed problem, and only update a subset of local variables in each iteration until the integer feasibility condition is met.

5. The vehicle adaptive cruise control method according to claim 1, wherein The rolling horizon optimization algorithm dynamically corrects the local trajectory, including: Construct a time-varying prediction model, discretize the vehicle dynamics equation into a state transition matrix, and the state transition matrix includes longitudinal velocity, lateral deviation, and yaw rate differential terms; Design a sliding window optimization objective function, including a trajectory tracking error term, a control input smoothing term, and an obstacle repulsion potential energy term; introduce a robust feasibility constraint, describe the uncertain parameters by an ellipsoidal set through semi-definite programming method, and construct a robust linear matrix inequality constraint; Use the alternating direction method of multipliers to solve the optimization objective function distributively.

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

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, and each node stores the point cloud density and the normal vector statistical features; Only activate non-empty nodes during the convolution operation, and retain the multi-resolution geometric details through skip connections; Use a dynamic pooling layer 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, wherein 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, the heading angle difference, and the relative velocity projection; Map the relative position vector to an attention bias term through a learnable non-linear transformation; Superimpose the bias term on the standard attention weight calculation process to enhance the geometric interpretability of the spatial correlation.

9. The vehicle adaptive cruise control method according to claim 4, wherein The probabilistic pruning model is constructed based on Gaussian process regression, including: Collect the objective function values and variable relaxation degrees of historical branch nodes as training samples; Construct a covariance kernel function, including a combined form of a linear kernel and a Matérn kernel; Optimize the hyperparameters by maximizing the marginal likelihood, and predict the feasibility probability of unexplored branches; When the feasibility probability is lower than the dynamic threshold, trigger the early pruning operation.

10. The vehicle adaptive cruise control method according to claim 5, characterized in that, The ellipsoidal set description is realized through singular value decomposition, including: Model the uncertain parameters as a bounded ellipsoidal set, the center vector is the nominal parameter, and the shape matrix is determined by the parameter covariance; Perform singular value decomposition on the shape matrix to extract the principal axis direction and length; All the half-axle directions are forced to satisfy the linear inequality conditions in the optimization constraints.

Images